Welding Journal - March 2023

SECTION District Activities NEWS District 2 — New Jersey January 12 Location: The Neighborhood Pub & Grill at Ellery’s, Middlesex, N.J. Presenter: Albert J. Moore Jr. Summary: Moore presented to members on short circuitry, gas metal arc welding, surface tension, and several other topics. Moore is known for his contributions to welding and cutting. There were 28 people in attendance. District 4 — Southwest Virginia December 9 Location: Wohlfahrt Haus Dinner Theatre, Wytheville, Va. Summary: Members participated in a planning meeting before attending a musical performance with their spouses. December 12 Location: Botetourt Technical Education Center, Fincastle ,Va. Summary: The Section hosted a successful Weld-Off event. Participating schools included Salem High School, Roanoke Technical Education Center, Botetourt Technical Education Center (BTEC), Giles County Technology Center, Floyd County High School, and Burton Center for Arts & Technology. Troy Linkenhoker, welding instructor at BTEC, was the host and coordinator. The Section is appreciative to judges Cameron Hatter, Brian Price, Richard Barner, Gary Young, Nathan Minnix, Jason Van Ness, Mike Bryant, Josh Graham, Micah Denson, and Mathew Fowler. Corporate donations also made SOUTHWEST VIRGINIA — Seen at the Wohlfahrt Haus Dinner Theatre are (from left) Section Chair David Shinault and wife Maureen and Gladys Johnson with Section Program Chair Wayne Johnson. SOUTHWEST VIRGINIA — Seen are (from left) Section Chair David Shinault, third-place winner Jayden Smith, and District Director Anver Classens. SOUTHWEST VIRGINIA — Section Chair David Shinault (left) is seen with second-place winner Ethan Carr (center) and District Director Anver Classens. MARCH 2023 | 49

SECTION NEWS the event possible. Donations included steel and sawing services from Steel Services of Roanoke; welding gloves and cutting glasses from Protective Industrial Products; skull caps and stickers from Revco Industries; welding helmets and drill bit sets from Walter Abrasives; ball caps and stickers from United Abrasives; a helmet, jacket, and school pack from Lincoln Electric; welding helmets from Miller Electric; and labor and logistics from Arc3 Gases. The Section sends a special thank you to District Director Anver Classens who presented $5000 in scholarships. All student participants put forth a good effort and worked professionally. The instructors should be very proud of each one. SOUTHWEST VIRGINIA — First-place winner Camden Italiano (center) is seen with Section Chair David Shinault (left) and District Director Anver Classens. SOUTHWEST VIRGINIA — Weld-Off participants are seen in a work area. District 4 — Triangle November 3 Location: Altec Industries, Creedmoor, N.C. Summary: The Section gathered at Altec Industries for a shop tour. The Creedmoor plant primarily manufactures truck equipment servicing the power and forestry industries. May 6 Location: ShopSpace, Raleigh, N.C. Summary: ShopSpace Cofounder Lucas House gave Section members a shop tour then a blacksmithing demonstration where he used the shop’s recently upgraded power hammer. ShopSpace is a nonprofit organization offering welding, fabrication, and blacksmithing classes. It is a community space for students and members to build their creations. 50 | WELDING JOURNAL

SECTION NEWS District 6 — Northern New York November 29 Location: Capital Region BOCES, Albany, N.Y. Presenter: Donald Mattoon, instructor, Capital Region BOCES Summary: In coordination with the Section, the Capital Region BOCES Career and Technical Education Center welding program held its annual industry review. Members from industry as well as current and former students toured the brand-new training facility, discussed industry trends, and spoke about the needs of the workforce. NORTHERN NEW YORK — Participants of the November meeting are seen during an informative discussion. District 8 — Nashville December 2 Location: Tennessee College of Applied Technology, Livingston, Tenn. Presenter: Curtis Duncan, welding engineer, Y-12, and AWS Nashville Section chair Summary: Duncan stopped by the Livingston campus of the Tennessee College of Applied Technology (TCAT) to visit with students. Joined by Welding Instructor Jason Wilborn, they discussed welding-related jobs available now and how technological advances are changing the future of welding while at the same time adding to the number of welding career possibilities. The discussion was aided by the presence of program graduate Colby Norrod. Norrod completed the program a year ago and is currently welding heavy wall nickel piping. He was able to share firsthand the experiences of finding a job as well as add to the career possibilities discussion. Thanks to TCAT Livingston VP Jeff Slagle for allowing the visit to take place. December 17 Location: Longhorn Steakhouse, Mt. Juliet, Tenn. Summary: Section members and their families met at Longhorn Steakhouse to enjoy fellowship and dine on the Saturday before Christmas. All attendees seemed to have a good time and their attendance is appreciated. NASHVILLE — Section Chair Curtis Duncan addressed TCAT-Livingston students. NASHVILLE — Heath Jackson, National Standard Welding Products district sales manager, and retired welding educator David Porter spoke while Jackson’s daughter, Brynn, looked at her phone to find the Jacksons a good route home. MARCH 2023 | 51

SECTION NEWS District 9 — Mobile November 17 Location: AIDT Maritime Training Center, Mobile, Ala. Presenter: Tony Hufford, category manager, Weiler Abrasives Summary: Hufford gave a presentation on the safe and effective use of portable cutting wheels followed by a questionand-answer session on proper usage. He passed out “swag” to those answering the questions correctly, thus engaging students and attendees. Afterward, there were hands-on abrasive demonstrations. Gorman Sales sponsored the meeting and provided all students with a pair of John Tillman welding gloves. Students Josue Calderon and Bowen McIlwain were the winners of the door prizes and both took home a Hyundai welding hood. MOBILE — Section Chair Tomasz Andraka (left) presented guest speaker Tony Hufford with a plaque of appreciation. MOBILE — Students Josue Calderon and Bowen McIlwain won door prizes donated by Hyundai Welding. MOBILE — Guest speaker Tony Hufford had some fun taking selfies with students. MOBILE — Meeting attendees looked on as Tony Hufford of Weiler Abrasives demonstrated proper use of cutting wheels. 52 | WELDING JOURNAL

SECTION NEWS District 10 — Drake Well Student Chapter December 2 Location: Pittsburgh, Pa. Summary: Students from the Venango Technology Center competed in the Pittsburgh District Annual AWS Weld-Off. The competition consisted of a written test of the student’s basic knowledge of welding and welding tests in the vertical and overhead positions. The welding tests were on 3/8-in. plate using a backup bar. The joint welding procedure was to be either a stringer or weave bead technique based on acceptable industry practice. Electrodes for the test were 1/8- and 5/32-in. E7018. All weldments were given visual inspection; those that passed were sent away to be radiographed in accordance with AWS D1.1, Structural Welding Code — Steel. The results will be announced this month. December 14 Summary: Members met with District 7 Director Roger Hilty and District 10 Director Tom Kostreba for the presentation of a $1000 scholarship to Matthew Woolcock for winning the Annual Pittsburgh Section AWS Weld-Off in March of 2022. Woolcock is currently pursuing a welding engineering degree at Pennsylvania College of Technology. DRAKE WELL STUDENT CHAPTER — Pictured (from left) are competitors Emily Murray, Jack Mumford, and Michael Crocker. DRAKE WELL STUDENT CHAPTER — District 10 Director Thomas Kostreba (left) and District 7 Director Roger Hilty (right) are seen with scholarship recipient Matthew Woolcock. MAHONING VALLEY — Past Section Chair Dave Hughes spoke to students about becoming good welders and responsible employees. District 10 — Mahoning Valley October 27 Location: Mahoning County Career and Technical Center (MCCTC), Canfield, Ohio Presenter: Tim Echan, technical sales representative, Lincoln Electric Summary: The Section held a morning meeting at MCCTC that included Section members, educators, and high school welding students throughout Mahoning Valley. The purpose of the meeting was to introduce students to welding industry MARCH 2023 | 53

SECTION NEWS opportunities as well as post–high school educational opportunities. December 22 Location: Davidson’s Restaurant, Canfield, Ohio Summary: An executive committee meeting was held as a planning session for 2023 Section events and activities. MAHONING VALLEY — Executive committee members gathered for a December planning meeting. Seen are (sitting, from left) Section Chair Joe Tufaro, Denny Naples, James Cox, and Sam Murray as well as (standing, from left) Pat Prokop, Chuck Moore, and Dave Hughes. District 12 — Racine-Kenosha November 22 Location: Gateway Technical College SC Johnson iMET Center, Sturtevant, Wis. Summary: The Section and Local 601 Steamfitters supported Gateway Technical College’s Women in Manufacturing event. Local 601 representatives discussed apprentice opportunities and Gateway Technical College had welding booths set up for participants to try welding. December 1 Location: American Bin and Conveyor, Burlington, Wis. Summary: Section members and students from Gateway Technical College toured American Bin & Conveyor. RACINE-KENOSHA — Women in Manufacturing event participants gathered for a group photo. RACINE-KENOSHA — Seen are the American Bin & Conveyor facility tour attendees. 54 | WELDING JOURNAL

SECTION NEWS District 13 — Chicago December 14 Location: Mama Luigi’s Restaurant, Bridgeview, Ill. Summary: Members gathered to recap 2022 and plan for their 2023 goal to recruit new members. Attendees also said goodbye to long-time Section member Marty Vondra who is moving to Texas and will be missed dearly. January 5 Location: Gibsons Bar & Steakhouse, Oak Brook, Ill. Summary: The Section held an appreciation dinner for District Director Ronald Ashelford. CHICAGO — December meeting attendees included Cliff Iftimie, Anghelina Iftimie, Craig Tichelar, Amanda Young, Marty Vondra, Jeff Stanczak, James Greer, John Hesseltine, and Section Chair Dave Viar. CHICAGO — Seen at an appreciation dinner for District Director Ronald Ashelford are Cliff Iftimie (sitting) and (front standing, from left) Cynthia Bern, Rita Vondra, Amanda Young, Craig and Kim Tichelar, Anghelina Iftimie, Cathy and John Hesseltine, District Director Ron Ashelford, and Crystal Ashelford. In the back row (from left) are Jeff Stanczak, Marty Vondra, and Section Chair David Viar. MARCH 2023 | 55

SECTION NEWS District 14 — Indiana January 11 Location: Maddox Industrial Group, Indianapolis, Ind. Presenter: Adam Bolyard Summary: Members toured Maddox Industrial Group for a first-hand look at the company’s work. Maddox specializes in cyrogenic work, air separation plants, ASME power and pressure piping, and sanitary piping. The company has its own certified oxygen cleaning room used to clean pipes, tubes, and fittings for oxygen service. INDIANA — Maddox LOX Box (liquid oxygen) in production. District 16 — Kansas December 8 Location: ICM Inc., Colwich, Kan. Summary: Members toured ICM Inc. to see how the company manufactures products that support the ethanol industry. KANSAS — Members are seen at a facility tour of ICM Inc. District 16 — Nebraska November 17 Location: Valmont Industries, Valley, Neb. Presenter: Chris TenEyck, Valmont Industries Summary: Members had the opportunity to tour Valmont’s small pole manufacturing facility and see how the company implements submerged arc and resistance welding processes as well as burnishes light poles to achieve desired strength properties. 56 | WELDING JOURNAL

SECTION NEWS NEBRASKA — Pictured are presenter Chris TenEyck (left) and Section Secretary Jeffrey Tyler. NEBRASKA — Valmont tour participants Adam Wietzki, Brandon Kielisch, Michael Gerhardt, Mark Roland, Chris TenEyck, and Chris TenEyck Sr. gathered for a group photo. District 20 — Colorado November 10 Location: Denver, Colo. Presenter: Micah Fonoroff, Shoot Indoors Denver Summary: Fonoroff hosted the Section for its first November Turkey Shoot at Shoot Indoors Denver. After a safety presentation, shooters headed to the range to show off their skills on old-fashioned turkey targets to see who would take home the three top prizes and the consolation prize package for last place. First-place winner Damon Nab from Northeastern Junior College received a $75 gift card; second-place winner Jason Hill, director of welding technology at Northeastern Junior College, received a $50 card; and third-place winner Dustin Nelson, Section certification officer, received a $25 card toward Thanksgiving dinner. The consolation prize package included gag gifts like a rubber chicken and targets to use for practice to place better for the next event. COLORADO — Section members, students, and host Micah Fonoroff gathered to celebrate the winners at Shoot Indoors Denver. (Photo credit: Gabe Costales of Unleashed Creative.) MARCH 2023 | 57

SECTION NEWS District 20 — New Mexico December 15 Location: Albuquerque, N.Mex Presenter: Robert Scheehl, cryogenic engineer and subject matter expert, Airgas Summary: Scheehl discussed additive manufacturing and 3D printing at the Section’s December meeting. NEW MEXICO — Meeting attendees including guest speaker Robert Scheehl (sitting, far right) and Section Treasurer Pat Bauman (second from right) gathered for a group photo. District 22 — Central Valley December 12 Location: CTEC High School, Fresno, Calif. Summary: The CTEC High School Women Who Weld Contest was supported by incoming AWS Vice President Kerry Shatell, Past District 22 Director Rob Purvis, and Current District 22 Director Randy Emery. The AWS leaders acted as judges and technical support for this first-ever event that brought together 17 young women contestants. Two $1000 scholarships were awarded to the top two finishers. The contestants represented various high school welding programs CENTRAL VALLEY — Contestants and coaches are seen with judges Kerry Shatell, Rob Purvis, and Randy Emery. 58 | WELDING JOURNAL

SECTION NEWS from the Central Valley Section. The contest was hosted and developed by the faculty of CTEC High School. December 15 Location: College of the Sequoias, Tulare, Calif. Summary: The Section held its final 2022 meeting and voted on new executive Section members. Randy Emery, who was the previous Section chair, will now serve as this year’s District 22 Director and Section treasurer. Chris Huff, who served as the vice chair, is now Section chair. Jose Baca, who served as the previous second vice chair, will now serve as the vice chair. Charles Jeffress will be serving as Section secretary. CENTRAL VALLEY — Contest winner Emily Crawford is seen with former District 22 Director Rob Purvis and Welding Educator Jacob Cavazos. CENTRAL VALLEY — Proud contestants showed off their work. CENTRAL VALLEY — Former District 22 Director Rob Purvis (right) and current District 22 Director Randy Emery posed for a photo while recapping a great experience with the next generation of skilled professionals. CENTRAL VALLEY — Members gathered for a group photo during their last meeting of 2022. MARCH 2023 | 59

PERSONNEL The Welding Institute Appoints President The Welding Institute (TWI) has selected Professor Dame Julia Elizabeth King, Baroness Brown of Cambridge, as president. She is the first female president appointed since the formation of TWI 100 years ago in 1923. King graduated from the University of Cambridge with a master’s degree in natural sciences and a doctorate in fracture mechanics. From 1987 to 1994, she held a series of research and teaching positions at Cambridge. In 1994, she moved on to Rolls-Royce, where she held a number of senior positions, including head of materials, managing director of fan systems, and engineering director of the marine business. She was appointed chief executive of the Institute of Physics in 2002. From 2004 to 2006, she was principal of the Engineering Faculty at Imperial College London, after which she joined Aston University as vicechancellor, where she served until 2016 . She has been awarded the Grunfeld, John Collier, Lunar Society, Constance Tipper, Bengough, Kelvin, and Leonardo da Vinci medals as well as the Erna Hamburger Prize and the 2012 President’s Prize of the Engineering Professors’ Council. In addition, she has been involved in chairing the Science and Technology Select Committee, the Carbon Trust, STEM Learning Ltd., and the Adaptation Committee of the Climate Change Committee. Sonobond Ultrasonics President Retires Janet Devine, president of Sonobond Ultrasonics, West Chester, Pa., a wholly owned subsidiary of Inductotherm Group, has retired from the company. Born and educated in England, Devine joined the predecessor company to Sonobond after immigrating to the United States in 1959. She served as the company’s vice president and technical director. She was also a key member of the team that developed and patented the Wedge-Reed System. Additionally, she has been an active participant in the development of many commercial, medical, and industrial systems and applications for ultrasonic metal welding and ultrasonic bonding of nonwoven fabrics and synthetic materials. She became president of Sonobond in 1990, when it was rare for a woman to head a technology company. In 1991, she became a member of the committee that authored the “Ultrasonic Welding” chapter in AWS’s Welding Handbook, Volume 2, Eighth Edition. She also authored the “Ultrasonic Welding” chapter for the ASM Handbook, Volume 6A — Welding Fundamentals and Processes. Devine has been awarded multiple patents. Velo3D Hires Director of Solutions Engineering Velo3D, Campbell, Calif., a metal additive manufacturing technology company for mission-critical parts, has welcomed Robin Stamp as director of solutions engineering to help grow the company’s metal additive manufacturing technology with new customers and in new industries. He will work with customers to understand their needs, expand the manufacturing capabilities of Velo3D’s solutions, and educate customers on how to fully leverage various technology improvements. He will also oversee the development of new standards with regulatory agencies, qualification of new metal alloys for use in the Sapphire family of printers, and collaboration with partners and agencies. Stamp has extensive experience leading teams in the R&D of additive manufacturing technology. He was previously a principal engineer at SpaceX, where he worked on developing technology for space applications. He also spent more than a decade at Stryker, leading the R&D team responsible for creating new additive manufacturing processes for medical implants. Alcoa Restructures Executive Leadership Team Alcoa Corp., Pittsburgh, Pa., a provider of bauxite, alumina, and aluminum products, has restructured its executive leadership team to further improve the company’s focus on operational excellence, cost, and innovation. William F. Oplinger has been promoted to executive vice president (EVP) and chief operations officer. Molly Beerman has been appointed EVP and chief financial officer (CFO). She will also be the executive member to oversee the company’s IT and automation solutions team. Renato Bacchi will take on added responsibilities to become EVP and chief strategy and innovation officer while overseeing the company’s R&D technologies. Oplinger has served as Alcoa’s CFO since 2016. He has prior operational experience, including previously serving as the chief operating officer for Alcoa Inc.’s Global Primary Products division. Prior to his role with Alcoa Corp., he served three years as EVP and CFO of Alcoa’s former parent company, which he joined in 2000. Beerman served as the company’s vice president and controller from 2016 to 2019, when she was promoted to senior vice president and controller. She served as director of global shared services strategy and solutions from November to December 2016. In 2016, she held a consulting role with Alcoa Inc. From 2012 to 2015, she was vice president of finance and administration for a nonprofit organization focused on community issues. Prior to that, she worked for 11 years for Alcoa Inc. in a variety of roles in the finance function and was the J. E. King J. Devine R. Stamp 60 | WELDING JOURNAL

director of global procurement center of excellence from 2008 to 2012. Earlier in her career, she served in financial management positions at Carnegie Mellon University, PNC Bank, and Deloitte. Bacchi was previously senior vice president and treasurer of the company from 2019 to 2022. In addition to responsibility for global treasury activities, he also had accountability for the company’s corporate development function. Prior to that role, he was vice president and treasurer. Prior to that, he was assistant treasurer. Wall Colmonoy Welcomes Business Development Manager Wall Colmonoy (USA), Madison Heights, Mich., a materials engineering group, has added Josh Gardner as its business development manag er — Surfacing Products as part of the Alloy Products Americas’ sales team. He will be responsible for expanding product applications for new and existing surface coating customers. He will work alongside the sales team, performing technical inputs for surfacing and thermal spraying techniques. Gardner has more than 23 years of experience in thermal spraying; laser cladding; plasma transferred arc welding; high-velocity oxygen fuel; and sales applications for oil and gas, mining, agriculture, and aerospace industries. He has additional experience in training personnel and designing and installing thermal spray systems. WJ J. Gardner Do You Have a Resistance Welding Question? Email your submission to the Welding Journal’s Education Editor Roline Pascal at [email protected] so she can forward it to the RWMA Q&A authors. You may also send it to her attention at Welding Journal Dept. 8669 NW 36 St., #130 Miami, FL 33166 Your resistance welding question may be chosen for this column and help other individuals better understand how to solve a particular problem. MARCH 2023 | 61

As more companies are faced with the need for labor, it’s no secret that automation provides a potential solution for mitigating workforce challenges. More than 517,000 industrial robots were installed worldwide in 2021, a 31% increase from the previous year (Ref. 1). The first three quarters of 2022 saw a 24% increase over the same time period in 2021, setting up yet another record-breaking year for industrial robotics (Ref. 2). In most applications, automation serves as a viable solution for handling the labor-intensive jobs that are suffering from both a need for operators and a significant increase in labor costs. Weld grinding is an example of one of these applications because many companies are struggling to hire and retain operators due to skill levels and the highly physical demands of this process. Weld grinding automation systems can not only help take the burden off of labor challenges, but they can also lead to improvements in safety, output, cycle time, and part consistency. The key to fully optimizing and benefitting from an automation system is ensuring that the robotic cell is built around the most effective process for the application. Deconstructing Automated Weld Grinding When it comes to welding processes, it is helpful to review the two phases of automated grinding: weld preparation and weld finishing. THE AMERICAN WELDER AUTOMATING GRINDING Learn tips for automating weld prep and finishing BY NATHAN JACKSON A Primer for 62 | WELDING JOURNAL

Weld preparation typically involves removing paint or rust from the surface and adding a bevel to the workpiece to provide an opening for the weld filler. Since paint and rust are usually only found in job shop situations, the cleaning phase is not a common application for automation. However, using automation to grind a bevel ensures there is a clean, straight opening that will allow for a more constant, reliable weld. Flap discs or depressed center wheels are the best-suited abrasives for weld preparation. Weld finishing is the process of removing excess stock from the weld and polishing the welded area to match the surface finish of the surrounding area of the part. This is generally performed as a two-step process consisting of grinding and blending. For the grinding step, coarse-grit abrasives are chosen for high-stock-removal rates. Abrasive belts, flap discs, and depressed center wheels are the optimal tools for weld grinding. Since the purpose of the blending step is to provide the best possible finish, less-aggressive abrasive products in a fine grit are ideal. Flap wheels, nonwoven discs, wheels or belts, and paper discs on an orbital sander are the best products for weld blending. Approaching Automation As discussed previously, automating weld grinding processes has many benefits. However, before diving into automation, it is critical to define the objectives. First, consider which parts you will be automating, how much you are willing to spend on an automation system, what qualifies as an adequate cycle time, and the measurements used to determine if a finished part is acceptable. These considerations exemplify a concept in automated systems that is often referred to as the 80/20 rule, meaning, in many cases, that automating the last 20% of your process accounts for 80% of the cost. The higher cost components can be for a variety of reasons, such as the following: ■ Requiring additional fixtures for hardto-reach welds, ■ Unique tooling for odd-shaped parts, and ■ Additional vision and sensors needed to finish grinding inconsistent areas of the parts. It is important to consider how much of the process you are willing to pay to automate. In many cases, the best solution is to automate the bulk of the grinding operation while leaving the final steps to be performed by operators, making the best use of the operators’ time and skills. One of the easiest ways to simplify the weld grinding process is to improve the consistency of welds upstream. If welds are consistent, the same cycle of grinding will be able to produce the same finished product time after time. Large variations in the size or shape of the weld may require identifying the grind zone with a machine vision system, which will increase the price and complexity of the automation system. Determining Grinding Cell Composition A grinding cell has options for one of two types of systems: abrasive to part or part to abrasive. In the abrasive-to-part system, the part is mounted within the cell and the robot uses end of arm tooling (EOAT) to bring the abrasive to the part. For a weld grinding cell, the most common forms of EOAT are spindle motors, pneumatic tools, and custom-built belt grinding heads. This type of system is generally used when the part is too large for a robot to pick up safely or when the part has tight areas to be ground that a large abrasive belt or wheel cannot reach. Alternatively, in a part-toabrasive system, the robot picks up the part using a gripper or a custom fixture and brings it to a grinding head. This type of system is generally used for smaller, high-volume parts. Compliance Compliance is a key feature to have in any automated grinding cell, and it is especially critical when it comes to weld grinding. Compliance is defined as the adjustment of pressing force via motion modification. Essentially, a compliance device uses pneumatic pressure to allow an abrasive to maintain a constant force at any point within a specific mechanical stroke. Compliance makes The Norton Abrasive Process Solutions (APS) automation cell includes multiple grinding and finishing heads, enabling a wide variety of grinding applications. MARCH 2023 | 63

programming a robot simpler, as it reduces the accuracy requirement for points in the toolpath. It also helps accommodate for variations in the part, allowing the robot to still move its intended point rather than crashing in the event that there is extra material on the weld. Using compliance to ensure the grinding force is constant will give a more consistent finish while reducing cycle time and increasing the abrasive life. This is because maintaining ideal force provides an optimal materialremoval-to-abrasive-wear ratio. As such, it is recommended to use premium abrasives, which will provide the longest life in an automated cell, reducing abrasive changeover time and maximizing the uptime. The compliance in the grinding cell can either be on the grinding head or the end of the robot arm. In an abrasiveto-part system, the force control will come from a compliance device placed between the spindle or abrasive tool and the robot arm. In a part-toabrasive system, the compliance generally comes from a pneumatic cylinder on the grinding head, allowing the grinding force to remain constant regardless of the contact point between the part and the abrasive. Automation Expertise Although automating your weld grinding process may seem daunting, a proof-of-concept lab from an automation partner can greatly reduce your risk. By testing different parameters and being able to see a completed part firsthand, you can be sure that your automation system will be able to deliver the results you need. With proper planning and process development, automation can provide the perfect solution your company needs for improving your weld grinding processes. WJ References 1. International Federation of Robotics. 2022. ifr.org/ifr-press-releases/news/ wr-report-all-time-high-with-half-a-millionrobots-installed 2. Wessling, B. 2022. Robot sales on track to hit a new high in 2022. WTWH Media. therobotreport.com/robot-sales-on-trackto-hit-a-new-high-in-2022 NATHAN JACKSON ([email protected]) is an automation application engineer at Norton | Saint-Gobain Abrasives, Worcester, Mass. Robotic part-to-abrasive grinding of a welded component on a compliant belt head. 64 | WELDING JOURNAL

THE AMERICAN WELDER Excerpted from the Welding Handbook, Ninth Edition, Volume 3, Welding Processes, Part 2. High-frequency resistance welding (HFRW) includes a group of resistance welding process variations that use a high-frequency welding current to concentrate the welding heat at the desired location. The heat generated by the electrical resistance of the workpiece to high-frequency currents produces the coalescence of metals, and an upsetting force is usually applied to produce a forged weld. The principal application of HFRW is the manufacture of seam-welded pipe and tube. Induction Seam Welding Process For induction seam welding, the pipe or tube is formed from a metal strip in a continuous-roll forming mill with the edges to be welded slightly separated. The open edges of the pipe or tube are brought together by a set of forge pressure rolls in a vee shape until the edges touch at the apex of the vee (i.e., where the weld is formed). The weld point occurs at the center of the mill forge rolls, which apply the pressure necessary to achieve a forged weld. An induction coil, typically made of copper tubing or copper sheet with attached water-cooling tubes, encircles the tube (the workpiece) at a distance equal to 1 to 2 tube diameters ahead of the weld point. This distance, measured from the weld point to the edge of the nearest induction coil, is called the vee length. The induction coil induces a circumferential current in the tube strip that closes by traveling down the edge of the vee through the weld point and back to the portion of the tube under the induction coil — Fig. 1. The highfrequency current flows along the edge of the weld vee due to the proximity effect (i.e., the high-frequency current path at the surface of the workpiece is controlled by how close it is or its proximity to its own return path). The edges are resistance-heated to a shallow depth due to the skin effect. The term skin effect refers to when the high-frequency current in metal conductors tends to flow at the surface of the metal at a relatively shallow depth, which becomes shallower as the electrical frequency of the power source is increased. The geometry of the weld vee is such that its length usually is between 11/2 to 2 tube diameters long. The included angle of the vee generally is between 3 and 7 deg. If this angle is too small, arcing between the edges may occur, and it will be difficult to maintain the weld point at a fixed location. If the vee angle is too wide, the proximity effect will be weakened, causing dispersed heating of the vee edges, and the edges may tend to buckle. The best vee angle depends on the characteristics of the tooling design and the metal to be welded. Variations in vee length and angle will cause variations in weld quality. The welding speed and power source level are adjusted so that the two edges are at the welding or forge temperature when they reach the weld point. The forge rolls press the hot edges together, applying an upset force to complete the weld. Hot metal containing impurities from the faying surfaces of the joint is squeezed out of the weld in both directions inside and outside of the tube. Depending on the application for the tube being produced, the upset metal is normally trimmed off flush with the base metal on the outside of the tube, but it is sometimes trimmed from the inside. An impeder, which is made from a magnetic material such as ferrite, is generally required to be placed inside the tube. The impeder is positioned so that it extends about 1/16 to 1/8 in. (1.5 to 3 mm) beyond the apex of the vee and the equivalent of 1 to 2 workpiece diameters upstream of the induction coil. The purpose of the impeder is to increase the inductive reactance of the current path around the inside wall of the workpiece. This reduces the current that would otherwise flow around the inside of the tube and cause an unacceptable loss of efficiency. The impeder also decreases the magnetic path length between the induction coil and the tube, further improving the efficiency of power transfer to the weld point. The impeder must be cooled to prevent its temperature from rising above its Curie temperature, where it becomes nonmagnetic. For ferrite, the Curie temperature is typically between 340° and 650°F (170° and 340°C). WJ Fact Sheet Fig. 1 — Induction seam welding of a tube. Intro to Induction Seam Welding of Pipe and Tube 66 | WELDING JOURNAL

CLASSIFIEDS AD INDEX ALM Positioners 9 almmh.com Web contact only Arcos IBC arcos.us (800) 233-8460 AWS Administration 13, 15, 31 aws.org (800) 443-9353 AWS Events & Conferences 65, 68 aws.org/events (800) 443-9353 AWS Foundation 30, 35, 61, 67 aws.org/foundation (800) 443-9353 Bradford Derustit 22 derustit.com (714) 695-0899 Carestream 23 carestream.com/ndt (888) 777-2072 Cor-Met 7, 19 cor-met.com (800) 848-2719 Diamond Ground 11 diamondground.com (805) 498-3837 Fischer Engineering Co. 37 fischerengr.com (937) 754-1750 Flexovit 10 flexovitabrasives.com (800) 689-3539 Fusion Inc. 20 fusion-inc.com (800) 626-9501 G.A.L. Gage Co. 8 galgage.com (269) 465-5750 Genstar 27 gentec.com (909) 606-2726 Hobart Institute 22 welding.org (800) 332-9448 Höganäs 1 hoganas.com/arcx Web contact only Hypertherm 2 hypertherm.com/sync Web contact only Lincoln Electric OBC lincolnelectric.com/pipefab (888) 935-3877 Red-D-Arc 14 reddarc.com/cobots (678) 528-0538 Select-Arc IFC select-arc.com (800) 341-5215 Tokin America Corp. 37 tokinarc.com (513) 644-9743 Trumpf 5 trumpf.com Web contact only Weld Engineering Co. Inc. 21 weldengineering.com (508) 842-2224 MARCH 2023 | 69

Rapid Publication of Cutting-Edge Welding Research — WJRS Letters Dear Researchers, The Welding Journal Research Supplement (WJRS) is seeking manuscripts for rapid publication of emerging research that is cutting edge and novel. The manuscripts are relatively short in length and will be called WJRS Letters. They are aimed at allowing authors to quickly publish new research results on “hot” topic areas of interest to the WJRS. Editor approval is required for submission of WJRS Letters manuscripts. If you are interested, please submit an abstract summarizing your proposed manuscript. WJRS Letters are not intended to be data dumps or options for publications of low-quality research. Abstracts that do not describe new, cutting-edge research will be declined. If your abstract is accepted, you’ll be invited to submit a manuscript. Below are more details. (1) First, provide an abstract < 500 words, which includes author names, emails, and affiliations, to [email protected]. Note: It was recently discovered that this email was not working, but it’s been fixed. A single figure (not multipart) with a caption can also be provided with the abstract. Complete manuscripts submitted without approval will be declined immediately. Do not submit abstracts or manuscripts to the WJRS submission website. (2) If your abstract is accepted, you’ll be invited to write a manuscript of not more than 2000 words with three to four figures (with captions). All content must fit into five pages when laid out in the current WJRS format reflected in the template. Also, you will receive a template for your Letters after abstract approval. Manuscript content must fit into the template without format changes to be published. (3) Manuscripts will require succinct and focused communication in the following sections: • Introduction/Background/Objectives • Streamlined Procedures • Results • Substantive Discussion • Major Conclusions • Limited References If you have questions, please email [email protected]. Thank you for supporting this effort. Sincerely, Thomas J. Lienert, PhD, FASM, FAWS Review Editor

WELDING RESEARCH https://doi.org/10.29391/2023.102.004 Application of Machine Learning to Regression Analysis of a Large SMA Weld Metal Database A cluster analysis of a coalesced Fe-C-Mn high-strength steel dataset revealed that the ultimate tensile strength of weld metal can be related to austenite-toferrite transformation temperature in at least four ways BY R. VARADARAJAN AND K. SAMPATH Abstract A machine learning approach was used to perform a regression analysis of Evans’s shielded metal arc (SMA) weld metal (WM) database involving several groups of Fe-C-Mn high-strength steels. The objective of this investigation was to develop an expression for austenite-to-ferrite (Ar3) transformation temperature that also included the effects of principal and minor alloy elements (in wt-%) and weld cooling rate (in °C/s) and relate this expression with WM ultimate tensile strength (UTS). The Ar3 data from 257 records obtained from several selected sources were combined with Ar3 projections at extreme end points in Evans’s WM database. Subsequently, a cluster analysis was performed. The data in Evans’s database was filtered with the carbon equivalent number limited to 0.3 maximum, carbon content limited to 0.1 wt-% maximum, nitrogen content limited to 99 ppm (0.0099 wt-%) maximum, preassigned Ar3 values limited to 680°C minimum, and WM UTS limited to 710 MPa maximum. The results provided a good approximation to the expression for Ar3 transformation temperature in terms of elemental compositions and cooling rate. This allowed the Ar3 to correlate with WM UTS of Fe-C-Mn in at least four ways depending on the sign of correlation of the data clusters. The elemental combinations in the cluster with the highest negative correlation revealed highly predictable WM UTS. In particular, the new Ar3 expression helped to predict decreases observed in certain Ar3 experimental data on WMs with balanced Ti, B, Al, N, and O additions reported among 13 records with additional dilatometry results. This correlation between the new expression for the Ar3 temperature and UTS of Fe-C-Mn WM is expected to complement the Japan Welding Engineering Society artificial neural network model currently available to predict Charpy V-notch test temperature for 28 J absorbed energy based on WM chemical composition. It will thereby provide a pair of effective tools for efficient development and/or evaluation of high-performance welding electrodes based on an Fe-C-Mn system for demand-critical applications. Keywords ■ Fe-C-Mn High-Strength Steels ■ Shielded Metal Arc Welding (SMAW) ■ Database ■ Machine Learning ■ Cluster Analysis ■ Transformation Temperature Introduction In May 2017, Dr. Glyn M. Evans (formerly with Oerlikon, Switzerland) posted a large shielded metal arc (SMA) weld metal (WM) database on ResearchGate (Ref. 1). This database contains more than 900 WM compositions based on an Fe-C-Mn system. These WM compositions belong to 74 types or groups of Fe-C-Mn alloy systems and were derived from SUPPLEMENT TO THE WELDING JOURNAL, MARCH 2023 Sponsored by the American Welding Society MARCH 2023 | 31-s

the book Metallurgy of Basic Weld Metal by G. M. Evans and N. Bailey (Ref. 2). Each WM composition includes individual ranges of 16 (C, Si, Mn, P, S, Cu, Ni, Cr, Mo, Nb (Cb), V, Ti, B, Al, N, and O) alloy element additions to Fe, along with the respective values for six WM properties that include yield strength (YS), ultimate tensile strength (UTS), reduction of area (%RA), elongation (%El), and the test temperature required to achieve 100 J (T100J/°C) and 28 J (T28J/°C) absorbed energy during Charpy V-notch (CVN) impact testing. A recent research effort (Ref. 3) utilized a constraints-based model (CBM) as a simple and effective framework for organizing and analyzing the Fe-C-Mn SMA WM database (see aws.org/2021.100.036-database) to gain valuable insights. The CBM was built on the metallurgical principle that lowering relevant solid-state phase transformation (i.e., austenite decomposition) temperatures is beneficial in improving WM strength and fracture toughness while simultaneously reducing carbon content. The carbon equivalent number (CEN) developed by Yurioka et al. (Ref. 4) was also beneficial in improving the weldability of high-strength steels. Figure 1 (Refs. 5, 6) illustrates the relationship between the transformation temperature and UTS of ferritic-pearlitic, bainitic, and martensitic steels. The UTS of the ferriticpearlitic steels appeared to range between 400 and 550 MPa, while the corresponding transformation temperature appeared to range between 650° and 900°C. Interestingly, these ranges allow one to limit data selection when performing regression analysis of a large database such as Evans’s SMA WM database. To gain valuable insights into Evans’s database, the CBM used various statistical (regression) equations and obtained several calculated metallurgical characteristics (CMCs). The CMCs related the chemical composition of high-strength steel WM through Yurioka et al.’s CEN and selected solid-state phase transformation-start (TS ) temperatures, such as Ouchi et al.’s ferrite-start (Ar3) temperature (Ref. 7) and Steven and Haynes’s bainite-start (Bs ) and martensite-start (Ms ) temperatures (Ref. 8), through respective constitutive equations. The individual CMCs allowed classification and/or ranking of the WMs in the database. The CEN regression equation developed by Yurioka et al. (Ref. 4) is commonly used to evaluate the hydrogen cracking sensitivity of various types of modern structural, pipeline, and pressure vessel steels: = + {() × } where A(C) refers to the accommodation factor that is a function of C content, while EMU refers to a set of elemental multiplication units involving Si, Mn, Cu, Ni, Cr, Mo, V, Nb (Cb), and B. () = 0.75 + 0.25 ℎ [20 × ( − 0.12)] = {/24 + /6 + /15 + /20 + ( + + + )/5 + 5 × } Yurioka et al.’s aforementioned equation includes microalloy additions such as V, Nb, and B in addition to various principal alloy elements such as C, Mn, Cr, Ni, Mo, and Cu. Equations 4–6 indicate that all principal alloy elements decrease austenite decomposition temperatures with C affecting to a maximum extent, particularly when C content exceeds 0.12 wt-%. These equations were developed several decades ago and can be used to calculate or estimate the transformation temperatures when the cooling rate supposedly remains constant. !" (°) = 910 − 310() − 80() − 20() − 55() − 15() − 80() ! (°) = 830 − 270() − 90() − 37() − 70() − 83() ! (°) = 561 − 474() − 33() − 17() − 17() − 21() The classification and/or ranking of all WMs in Evans’s database using various CMCs obtained using the CBM approach (Ref. 3) reaffirmed that controlling the C content, CEN value, and calculated solid-state phase transformation temperatures, particularly the difference between the calculated BS and calculated MS temperatures, is critical to developing and identifying high-performance, high-strength steel welding electrodes. A dual approach that manipulated the contents of principal alloy elements, such as C, Mn, Cu, Ni, Cr, and Mo, based on Equations 4–6 and added balanced amounts of Ti, B, Al, N, and O appeared to offer the best means to lower (1) (2) Fig. 1 — Relationship between UTS (in MPa) and transformation temperature (in °C) of ferritepearlite, bainitic, and martensitic steels (Refs. 5, 6). (3) (4) (5) (6) 32-s | WELDING JOURNAL

relevant solid-state TS temperatures to produce WMs with high strength and exceptional fracture toughness. A part of Evans’s large SMA WM database used an independent scheme to build a total of 24 SMA welds, based on a TiBAlN series. These 24 welds included three subsets of eight welds, each at three levels of nitrogen content: normal (below 85 ppm or 0.0085 wt-%), intermediate (120 to 164 ppm or 0.012 to 0.0164 wt-%), and high (217 to 249 ppm or 0.0217 to 0.0249 wt-%). The primary intent of these three subsets of WMs was to identify and correlate the effects of Ti-B-Al-N-O microalloy additions on WM tensile strength, CVN impact toughness, and microstructure development in the fusion zone and reheated WM. Table 1 shows the chemical composition of 13 of 24 TiBAlN series of SMA WMs. The C content of these 13 WMs varied between 0.066 and 0.078 wt-%, while the Mn content varied between 1.4 and 1.66 wt-%. The Si content varied between 0.25 and 0.63 wt-%. Other principal alloy elements remained constant: Cr content at 0.03 wt-%, Ni content at 0.03 wt-%, Mo content at 0.005 wt-%, and Cu content at 0.03 wt-%. Microalloy additions V and Nb were held constant at 0.0005 wt-%. Compared to the aforementioned principal alloy elements and V and Nb microalloy additions, the Ti content varied between 0.0001 and 0.054 wt-%, B content ranged between 0.0001 and 0.0167 wt-%, Al content varied between 0.0001 and 0.058 wt-%, N content varied between 0.0041 and 0.0249 wt-%, and O content varied between 0.0282 and 0.0475 wt-%. Following weld mechanical testing that included all-weld metal tensile testing at ambient temperature and CVN impact testing over a wider temperature range from –120° to + 40°C, these 13 WMs were subjected to dilatometric evaluation (Refs. 9–11). These 13 WMs were selected to allow dilatometric evaluation of specific alloy additions relative to a range of Ti, B, Al, N, and O contents, particularly the effect of N content at three levels on both TS and transformation-finish (TF ) temperatures. The dilatometric evaluation studied the austenite-to-ferrite transformation during continuous cooling. Test specimens were machined to form hollow cylinders with the following dimensions: 10 mm long by 5 mm O.D. with 1 mm wall thickness. The axis of the test specimen was maintained parallel to the original welding direction. The specimens were subjected to the following controlled thermal cycle: austenitization at 1250°C for 2 min, followed by continuous cooling at a typical (weld) cooling rate of 13°C/s from 800° to 500°C (also known as Δt8/5; 13°C/s corresponds to about 23 s to cool from 800° to 500°C). The study determined the TS, 50% transformation (T50), peak rate transformation (TPRTT), and TF temperatures of the 13 (O, O2, X, X2, Y, Y2, Z, Z2, U, U2, V, V1, and V2) WMs. Table 2 shows the numerical values for the various transformation temperatures besides (TS–TF ) values, along with UTS and CVN test temperatures for 100 and 28 J absorbed energy of the 13 WMs (Refs. 9–11). The results found that weld V had the lowest TS temperature at 680°C and correspondingly the highest UTS at 732 MPa. A progressive increase in TS temperature occurred for welds U, Y2, Z2, Y, Z, O2, V2, X, X2, V1, O, and U2. Correspondingly, a progressive decrease in UTS from 644 to 528 MPa occurred in welds V1, Z, X2, O2, Y2, Y, V2, U, Z2, X, U2, and O. When WM N content was below 80 ppm (0.008 wt-%), the overall trend between TS temperature and UTS among all six WMs was found to be highly correlated, as shown in Fig. 2. Two of these six welds, welds Z (with 640 MPa UTS) and Y (with 594 MPa UTS), appeared closer and on either side of the trend line, indicating that their Ti, B, Al, N, and O additions were well or adequately balanced. The trendline equation showed the following: ( ) = 1814 − (1.6722 × ! °), ℎ " = 0.6368 thereby clearly confirming the metallurgical principle that lowering the transformation temperature aided to increase WM UTS. The trendline also indicated that a TS temperature greater than 680°C achieved a WM UTS less than 700 MPa (100 ksi). A recent review (Ref. 12) of the dilatometric results of the 13 Fe-C-Mn high-strength steel SMA WMs (see aws.org/2022.101.010-database) revealed that balanced Ti, B, Al, N, and O additions in welds Z and Y reduced the TS temperature. For example, weld Z containing microalloy additions as listed in Table 1 showed a total Ti, B, Al, N, and O content at 0.1133 wt-% that appeared to ensure effective deoxidation, formed complex inclusions, and distributed inclusions to enable development of highly fractureresistant refined WM microstructures. Depending on nominal WM chemical composition and actual effects during welding, these Ti, B, Al, N, O additions further lowered the actual TS temperatures, thereby promoting a cloudburst of austenite-to-ferrite phase transformation over a narrow (TS–TF ) temperature range. It may be wiser to avoid the rich and lean ends for these microalloy additions, except N, which should be held at the lean end, preferably much below 80 ppm (0.008 wt-%). As shown in Table 2, at a N content below 0.01 wt-%, the total Ti, B, Al, N, O additions were at 0.0833 wt-% in weld Y and 0.1133 wt-% in weld Z. These two welds offered nearly a 100°C shift in lowering CVN test temperature for either 28 or 100 J absorbed energy. Dilatometric evaluations of reheated WMs showed the following: 1) The balanced total (7) Fig. 2 — Effect of TS temperature on WM tensile strength. MARCH 2023 | 33-s

Table 1 — Chemical Composition of 13 TiBAlN Series Weld Metals (Ref. 1) Weld ID C (wt-%) Si (wt-%) Mn (wt-%) P (wt-%) S (wt-%) Cu (wt-%) Ni (wt-%) Cr (wt-%) O 0.074 0.25 1.4 0.007 0.008 0.03 0.03 0.03 02 0.073 0.27 1.66 0.008 0.009 0.03 0.03 0.03 X 0.069 0.45 1.47 0.006 0.005 0.03 0.03 0.03 X2 0.068 0.47 1.46 0.006 0.007 0.03 0.03 0.03 Y 0.07 0.45 1.57 0.01 0.006 0.03 0.03 0.03 Y2 0.069 0.36 1.51 0.008 0.007 0.03 0.03 0.03 Z 0.072 0.49 1.56 0.01 0.007 0.03 0.03 0.03 Z2 0.068 0.5 1.45 0.011 0.006 0.03 0.03 0.03 U 0.073 0.4 1.52 0.011 0.006 0.03 0.03 0.03 U2 0.066 0.36 1.4 0.012 0.007 0.03 0.03 0.03 V 0.078 0.6 1.44 0.007 0.006 0.03 0.03 0.03 V1 0.067 0.63 1.44 0.01 0.005 0.03 0.03 0.03 V2 0.069 0.6 1.42 0.012 0.006 0.03 0.03 0.03 Low 0.066 0.25 1.4 0.006 0.005 0.03 0.03 0.03 High 0.078 0.63 1.66 0.012 0.009 0.03 0.03 0.03 Range 0.012 0.38 0.26 0.006 0.004 0 0 0 34-s | WELDING JOURNAL

Table 1 — (continued) Mo (wt-%) Nb (wt-%) V (wt-%) Ti (wt-%) B (wt-%) Al (wt-%) N (wt-%) O (wt-%) 0.005 0.0005 0.0005 0.0001 0.0001 0.0006 0.0079 0.0475 0.005 0.0005 0.0005 0.0005 0.0005 0.0005 0.0235 0.0399 0.005 0.0005 0.0005 0.041 0.0002 0.0001 0.0077 0.0282 0.005 0.0005 0.0005 0.045 0.0002 0.0005 0.0249 0.0297 0.005 0.0005 0.0005 0.039 0.0039 0.0013 0.0083 0.0308 0.005 0.0005 0.0005 0.041 0.0044 0.0005 0.0232 0.0292 0.005 0.0005 0.0005 0.042 0.0048 0.016 0.0067 0.0438 0.005 0.0005 0.0005 0.047 0.0045 0.018 0.023 0.044 0.005 0.0005 0.0005 0.039 0.0158 0.0005 0.0084 0.029 0.005 0.0005 0.0005 0.039 0.0167 0.0005 0.0217 0.0297 0.005 0.0005 0.0005 0.054 0.0056 0.058 0.0041 0.044 0.005 0.0005 0.0005 0.048 0.0044 0.056 0.012 0.0473 0.005 0.0005 0.0005 0.043 0.0035 0.056 0.0235 0.047 0.005 0.0005 0.0005 0.0001 0.0001 0.0001 0.0041 0.0282 0.005 0.0005 0.0005 0.054 0.0167 0.058 0.0249 0.0475 0 0 0 0.0539 0.0166 0.0579 0.0208 0.0193 MARCH 2023 | 35-s

Ti, B, Al, N, O additions lowered the actual TS temperature by about 60°C compared to the calculated Ar3 transformation temperature obtained from Ouchi et al.’s constitutional relationship (Equation 4); 2) N more than 100 ppm (0.010 wt-%) effectively nullified the beneficial effects of Ti, B, and Al additions in lowering the transformation temperature; and 3) at N content much below 80 ppm (0.008 wt-%), both a lower TS temperature and a narrow start-to-finish (TS–TF) temperature range helped in achieving exceptional WM CVN impact toughness. The Japan Welding Engineering Society (JWES) offers a website (www-it.jwes.or.jp/weld_simulator/en/cal6.jsp) wherein one can calculate and predict the temperature for 28 J CVN impact toughness or absorbed energy of Fe-C-Mn WMs based on their chemical composition with certain minimum and maximum limits for all 16 (C, Si, Mn, P, S, Cu, Ni, Cr, Mo, Nb (Cb), V, Ti, B, Al, N, and O) alloy additions to Fe. The prediction is performed using artificial neural network (ANN) analysis by a software developed by D. J. C. MacKay at the University of Cambridge. The prediction is possible within the minimum and maximum limits set for the 16 alloy elements and uses Evans’s SMA WM database on low-alloy, high-strength steel WM as a basis. Evans’s database (Refs. 1, 2) has been made available to the University of Cambridge (phase-trans.msm.cam.ac.uk/ map/data/materials/). The database contains properties of all-weld metals obtained under the constant welding conditions of 1 kJ/mm arc energy and 200°C interpass temperature on 20-mm-thick plates. The ANN prediction gives maximum, minimum, and average values of transition temperature for 28 J CVN impact toughness or absorbed energy along with the degree of prediction error. When the difference between the maximum and minimum predicted values is more than 30°C, the prediction is considered unreliable. Figure 3 shows the predicted temperature for 28 J absorbed energy for welds Z (left) and Y (right). Table 3 shows a comparison of the actual test values and predicted results of CVN temperature for 28 J absorbed energy for all the 13 original welds. The predicted values for the 13 original welds were quite consistent with actual test results in both values and trend, and the error values associated with predictions were much less than 30°C for each of these welds. Furthermore, when the Ti, B, Al, N, and O contents of all 12 WMs were modified to the same values as in weld Z, the predicted temperature for 28 J absorbed energy for the welds decreased (Table 3) in all cases except weld Y with a total (Ti+B+Al+N+O) content at 0.0833 wt-% and weld V1 with a total (Ti+B+Al+N+O) content at 0.1677 wt-%. These findings clearly demonstrated that the UTS of Fe-C-Mn ferritic WMs increased with the decreasing TS, and a superior WM CVN toughness could be achieved by balancing Ti, B, Al, N, and O additions. As shown in Table 2, the balancing of Ti, B, Al, and O additions may be related to TS, with a decreasing TS requiring a lower amount of Ti, B, Al, and O additions. As revealed by weld V, when TS is at its low end, excessive amounts of Ti, B, Al, and O additions likely raised the CVN test temperature for 28 J absorbed energy, indicating the possibility to form numerous inclusions, which resulted in a so-called dirty weld. By contrast, as revealed by weld O, when TS is at its high end, disproportionate amounts of Ti, B, Al, and O additions likely raised the CVN test temperature for 28 J absorbed energy, indicating the possibility of free oxygen in solution. Interestingly, the JWES ANN template allows one to identify balanced Ti, B, Al, N, and O contents of WMs by manipulating their contents within the specified ranges mentioned in the JWES ANN template and achieve a CVN test temperature for 28 J absorbed energy colder than –60°C while ensuring Fig. 3 — The JWES neural network–predicted temperature for 28 J absorbed energy. Weld Z is on the left, and weld Y is on the right. 36-s | WELDING JOURNAL

that the error values associated with predictions were much less than 30°C for each of these welds. In other words, one could use the –60°C CVN test temperature for 28 J absorbed energy as a benchmark to distinguish welds with balanced Ti, B, Al, N, and O contents. While the JWES ANN template is available to predict CVN test temperature for 28 J absorbed energy based on WM chemical composition, a complementary relationship or ANN template involving WM UTS and WM chemical composition is currently unavailable. Modeling of WMs In recent years, there has been a growing interest to develop computer-based models on WM mechanical properties, particularly WM tensile and impact or fracture toughness properties based on WM chemical composition. For the most part, these modeling activities primarily involved ANNs, which are emerging as powerful tools with the capacity to reconstruct a database on weld properties through data selection and augmentation. These modeling activities implicitly recognize that suppressing austenite decomposition or lowering solid-state phase transformation temperatures induces greater nucleation rates and refines the resultant microstructural constituents (Ref. 13), thereby enhancing weld mechanical properties. However, there had been no explicit modeling efforts to correlate WM tensile and CVN impact toughness or fracture appearance transition temperature (FATT) with WM chemical composition through austenite (decomposition) transformation temperatures. Availability of such a correlation or relationship between WM UTS and WM chemical composition could complement the JWES ANN currently available to predict CVN test temperature for 28 J absorbed energy based on WM chemical composition, thus providing a pair of effective tools for efficient development Fig. 4 — Trellis plot of element amounts (in wt-%) against WM UTS in Evans’s dataset. Fig. 5 — The regression fit of UTS against the values reported in Evans’s database. MARCH 2023 | 37-s

Table 2 — Test Results of Transformation Temperature of 13 TiBAlN Weld Metals at 13°C/s Cooling Rate (Refs. 9–11) Weld ID Ti (wt-%) B (wt-%) Al (wt-%) N (wt-%) O (wt-%) O 0.0001 0.0001 0.0006 0.0079 0.0475 O2 0.0005 0.0005 0.0005 0.0235 0.0399 X 0.041 0.0002 0.0001 0.0077 0.0282 X2 0.045 0.0002 0.0005 0.0249 0.0297 Y 0.039 0.0039 0.0013 0.0083 0.0308 Y2 0.041 0.0044 0.0005 0.0232 0.0292 Z 0.042 0.0048 0.016 0.0067 0.0438 Z2 0.047 0.0045 0.018 0.023 0.044 U 0.039 0.0158 0.0005 0.0084 0.029 U2 0.039 0.0167 0.0005 0.0217 0.0297 V 0.054 0.0056 0.058 0.0041 0.044 V1* 0.048 0.0044 0.056 0.012 0.0473 V2 0.043 0.0035 0.056 0.0235 0.047 Low 0.0001 0.0001 0.0001 0.0041 0.0282 High 0.054 0.0167 0.058 0.0249 0.0475 Range 0.0539 0.0166 0.0579 0.0208 0.0193 *The transformation temperature for weld V1 was obtained from interpolation of graphical data reported in Ref. 11. 38-s | WELDING JOURNAL

Table 2 — (continued) (Ti + B + Al + N + O) (wt-%) UTS (MPa) CVN Test Temperature (°C) Transformation Temperature (°C) @ 100 J @ 28 J TS T50 TPRTT TF (TS–TF) 0.0562 528 –14 –42 762 658 650 554 208 0.0649 607 20 –16 754 630 606 534 220 0.0772 577 –61 –77 760 660 630 568 192 0.1003 631 –30 –58 760 650 638 572 188 0.0833 594 –82 –98 710 625 618 560 150 0.0983 605 –24 –56 703 612 606 510 193 0.1133 640 –83 –100 710 645 642 550 160 0.1365 583 13 –18 703 635 632 560 143 0.0927 586 –53 –80 700 626 620 531 169 0.1076 541 –52 –81 765 662 650 574 191 0.1657 732 –12 –46 680 598 596 507 173 0.1677 644 –64 –93 760 680 640 555 205 0.173 591 –45 –70 754 644 642 588 166 0.0562 528 –83 –100 680 598 596 507 143 0.173 732 20 –16 765 680 650 588 220 0.1168 204 103 84 85 82 54 81 77 MARCH 2023 | 39-s

of welding electrodes based on an Fe-C-Mn system for highperformance applications. Fujii and Ichikawa (Ref. 14) developed an ANN that could predict weld properties, including WM strength, FATT, and hardness. In addition, as one of the characteristics of their prediction, they identified that unreliability in estimated values can be displayed by the magnitude of the error bar. According to this system, the magnitude of the error bar was dependent on input conditions (test conditions) at the time; for example, where the data dispersion was large and reliability was low, the error bar was displayed as large and the computer itself was equipped with a function that could display the reliability of its prediction. The prediction of this error bar substantially extended the application scope of conventional neural networks and allowed the possibility of their application to the reconstruction of databases related to various properties. The University of Cambridge performed an ANN analysis of a vast and fairly general database assembled from publications on WM properties involving YS, UTS, elongation, and CVN impact toughness of ferritic steel WMs expressed as functions of chemical composition, heat input during welding, and postweld heat treatment (Refs. 15–18). This effort also used Evans’s SMA WM database on Fe-C-Mn WMs (Refs. 1–2). The outputs of the model were assessed in a variety of ways, including specific studies of model predictions for the Fe-C-Mn and Fe-2.25Cr-1Mo systems. Comparisons were also made with corresponding methods that used simple Fig. 6 — A regression of coalesced Ar3 values (Salganik et al., Trzaska, and Ouchi et al.) against WM UTS. Fig. 7 — A plot of new Ar3 values against WM UTS. 40-s | WELDING JOURNAL

physical metallurgical principles. The models appeared to capture vital metallurgical trends and emulate expectations from current physical metallurgy principles yet required much more systematic experimental data to improve the accuracy of their predictions. The U.S. Navy used an ANN technique developed by MacKay with a Bayesian framework wherein the probability of occurrence is interpreted as a reasonable expectation of a state of current knowledge but allows estimation of error bars like the ones introduced by Fujii and Ichikawa (Ref. 14). It also warns the user when data is sparse or locally noisy. This ANN was trained and tested on a set of data obtained from WMs of various steel types used for shipbuilding (Ref. 19). The input variables for the network used WM chemical composition and weld cooling rate. The output consisted of YS and UTS, elongation, and reduction of area. This effort created many models using different network configurations and initial conditions. The method revealed significant trends describing the dependence of WM mechanical properties on WM chemical composition and cooling rate. The U.S. Navy also used a similar ANN approach to model WM toughness characterized by CVN and dynamic tear tests Table 3 — Predicted CVN Test Temperature of 13 TiBAlN Series Weld Metals with Balanced TiBAlNO Additions Weld ID Measured CVN Test Temperature (°C) @ 28 J CVN Predicted Test Temperature (°C) @ 28 J Original Weld Balanced Weld 0 –42 –49 –82 O2 –16 –15 –88 X –77 –80 –90 X2 –58 –59 –89 Y –98 –102 –94 Y2 –56 –54 –89 Z –100 –94 — Z2 –18 –28 –89 U –80 –79 –92 U2 –81 –78 –84 V –46 –47 –95 V1 –93 –93 –90 V2 –70 –71 –90 MARCH 2023 | 41-s

of the same types of steels used for shipbuilding (Refs. 20, 21). The level of noise in the experimental data was perceived to be high, but it nevertheless allowed one to recognize reasonable trends and uncertainties when making predictions. For example, the WM toughness showed a nonlinear deterioration as the WM oxygen concentration increased, yet this behavior could be assessed quantitatively. Recently, Kim et al. (Ref. 22) performed an ANN analysis of Evans’s SMA WM database and investigated the effect of WM chemical composition on WM mechanical properties, including YS and UTS, and test temperature on CVN impact toughness testing to provide 100 J absorbed energy. Based on the data collected from previously performed experiments, Kim et al. developed correlations between related variables, analyzed the results, and offered predictive models. They prepared sufficient datasets using data augmentation techniques to overcome problems caused by insufficient data and enable better predictions. Finally, they developed closed-form equations based on the predictive models to evaluate WM mechanical properties according to WM chemical composition. Each ANN model developed in this study considered changes in the content of only two elements. The study is mainly useful to predict the relative increase or decrease according to the change in the content of any two elements. A recent research effort by Xiong et al. (Ref. 23) applied machine learning to predict mechanical properties of steels. The investigators selected 360 data on four mechanical properties (fatigue strength, tensile strength, fracture strength, and hardness) of both carbon steels and low-alloy steels from the National Institute for Materials Science (NIMS) database. They applied five machine learning algorithms on the 360 datasets to predict mechanical properties and determined that random forest regression provided the best correlation among the four most important features (tempering temperature and alloying elements of C, Cr, and Mo) for the mechanical properties of steels. They also used symbolic regression to generate mathematical expressions that explicitly predicted how each of the four mechanical properties varied quantitatively with the four most important features. Objectives The objective of the current effort was to use machine learning (Refs. 24, 25) for the following: 1) Determine a new expression for Ar3 temperature applicable to Evans’s SMA WM database that includes the effects of 16 principal and minor alloy elements (in wt-%) and weld cooling rate (in °C/s), 2) Use the new expression to develop a relationship with WM UTS, and 3) Perform a cluster analysis to gain additional insights. It is well known that the austenite-to-ferrite (Ar3) transformation temperature is a function of alloy additions, prior austenite grain size, cooling rate, and possibly prior thermomechanical processing history. However, Evans’s database didn’t have information on prior austenite grain size and weld cooling rate. Only the 13 WMs (Refs. 9–11) that were subjected to dilatometry evaluation recorded the weld cooling rate. Consequently, development of a new Ar3 expression based on Evans’s database is likely to have several limitations. However, as most material specifications and welding electrode specifications require the UTS to exceed a certain minimum value, one could still derive exceptional benefit from a new Ar3 expression that would be helpful in correlating Ar3 with WM UTS. Procedure A review of the metallurgical literature revealed numerous formulae on Ar3 expressions (Refs. 26–31). For example, Gorni (Ref. 32) in the Steel Forming and Heat-Treating Handbook documented a large collection of formulae that included regression equations for Ar3 temperatures of several types of steels. A few of these equations include various microalloy additions along with cooling rates or cooling times from 800° to 500°C, or Δt8/5. However, many times, the Ar3 formulae reported in the literature are applicable only over limited ranges of element compositions and other predictors such as grain size, strain rate, etc. In support of the current research effort, some Ar3 equations, including Ouchi et al.’s expression (Equation 4), seemed to be suitable starting points. Still, they are limited in chemical composition range and predictive capabilities because they do not include all the 16 alloying elements reported in Evans’s database. In addition to Ouchi et al.’s Ar3 regression expression, two other Ar3 regression equations, mentioned below, were considered for use in the current research effort. The equation by Salganik et al. is as follows (Ref. 26): !"(°) = 735.6 + 180.1( + ) + 1206.9( + ) − 10.9( + + + + ) + 755.3( + ) − 328.8( + + ) The equation by Trzaska is as follows (Ref. 27): (8) Fig. 8 — A plot of UTS against new Ar3 values (Equation 17) in clusters with a strong negative correlation. 42-s | WELDING JOURNAL

!"(°) = 857 − 257() + 23() − 69() +26() − 38() − 20() − 20() + 34() + 0.07(#) − 17$.&' where TA is the austenitizing temperature in °C, and the CR is the cooling rate in °C/min. The sign of Si in Salganik et al.’s Ar3 expression seemed to interfere with the correct evolution of the regression coefficient for Si in the current work. It was set to +10.9, and the Ar3 expression by Salganik et al. was modified as follows: !"(°) = 735.6 + 180.1( + ) + 1206.9( + ) + 10.9 − 10.9( + + + ) + 755.3( + ) − 328.8( + + ) Similarly, the Cu term in Trzaska’s Ar3 expression seemed to interfere with the correct evolution of the Cu coefficient sign in the current work. So, it was changed to –10(Cu), and Trzaska’s Ar3 expression was modified as follows. !"(°) = 857 − 257() + 23() − 69() − 10() − 38() − 20() − 20() + 34() + 0.07(#) − 17$.&' The data ranges for the applicability of these equations are given in Gorni’s handbook (Ref. 32). Trzaska’s equation for Ar3 seemed to be applicable over almost the entire range of Evans’s database, although it did not include all the alloying elements. Salganik et al.’s expression for Ar3 was applicable only to a limited number of records in Evans’s database because of the narrow data range of this expression, though it included several minor alloying elements. The above expressions were used in this work as discussed below. The initial data preparation involved an examination of Evans’s database and addition of columns to include various CMCs, including a CEN (Ref. 4) and selected austenite decomposition temperatures, such as TS or Ar3 (Ref. 7) or Bs and Ms (Ref. 8), based on WM compositions using corresponding constitutive equations. Following the aforementioned computation using Microsoft Excel, certain records that did not obey the relationship on the required ordering of these temperatures (i.e., Ar3 > Bs > Ms ) were eliminated. Also, some records containing no information in some columns were excluded. This first phase filtering yielded 858 records. Then the records were limited to the high wt-% of elements indicated by Trzaska’s Ar3 data range. This filtering yielded 809 records. Then two indicators were set on the data to indicate if the records were in Salganik et al.’s data range or in Trzaska’s data range. Then Ar3 values from the modified Salganik et al. (Equation 10) and modified Trzaska (Equation 11) equations were added to the records as appropriate. A coalesced Ar3 column was also added to the data. Salganik et al.’s modified Ar3, Trzaska’s modified Ar3, and Ouchi et al.’s Ar3 (Equation 4) were coalesced in this order. The above steps were taken to identify and select valuable records for regression analysis. These Ar3 values provided tentative values that appeared in certain records that were added as high-end guideposts to regression. Subsequently, the data was further restricted to certain WM compositions wherein the CEN was limited to 0.3 maximum, carbon content was limited to 0.1 wt-% maximum, and nitrogen content was limited to 99 ppm (0.0099 wt-%) maximum. Also, records with coalesced Ar3 values less than 680°C that appeared as strong outliers in an UTS vs. Ar3 regression were excluded. (9) (10) (11) Fig. 9 — A plot of UTS against new Ar3 values (Equation 18) in clusters with a moderate negative correlation. MARCH 2023 | 43-s

Data records with UTS values greater than 710 MPa were excluded. All these data filters reduced the number of records to 595 out of more than 900 records in Evans’s database. To study potential regression equations for Ar3, the 13 WM records in the TiBAlN series that had experimental Ar3 (i.e., TS ) values (Refs. 9–11) were extracted and combined with Ar3 predictions for 21 other records that were within the range of Salganik et al.’s (Ref. 26) modified regression equation and 75 records found in the Mintz et al. dataset (Ref. 28). To make the new expression for Ar3 applicable over the entire range in Evans’s dataset, 23 records with highest wt-% of elements and fractions of the highest wt-% in Evans’s dataset were also added with Ar3 values predicted using the Ar3 coalescing logic described earlier. Then additional data on Ar3 and respective element compositions and cooling rates found in Refs. 29–43 were collected. These experimental datasets included a wide variety of low-alloy steel compositions and cooling rates and respective Ar3 values. Also, 20 sets of numerical records on various steels with random element composition values at low range were included to improve the regression intercept and accuracy. A composite dataset created in this manner was used to derive a new regression equation for Ar3, including the effects of major and minor alloy elements (in wt-%) and weld cooling rate (in °C/s). Appendix I (aws.org/2023.102.004.appendix) provides the experimental datasets extracted from selected sources and used in this investigation. A standard machine learning approach implemented in R (Refs. 24, 25) using linear model (lm()) and model identification using regsubsets() was used in the investigation, and the final formula was hand tuned as well. A data frame of linear and nonlinear predictors with the experimental or coalesced temperatures was prepared, and the model was generated as follows: model = lm(Ar3 ~., data = df) or model = regsubsets(Ar3~., data = df , method = ‘backward’) where df is the data frame. Several functions were invoked on the model output to print out various properties of the model. The function summary(mdl) is typically used to indicate the fit coefficients in the case of lm(), and the summary() function provides a high-level summary of various models generated when regsubsets() is used. Additional functions were invoked to probe the contents of the models further. The output of regsubsets() contained various models tried, their R2, adjusted R2, Cp (a variant of Akaike information criterion or AIC, developed by Colin Mallows), and BIC (Bayesian information criterion) values for all models. Cp, BIC, and adjusted R2 are the model metrics typically used to select an appropriate model. Many linear and nonlinear predictors were provided to the models by adding respective columns in the data frame. In the beginning, many square and square root terms were added to the data frame for this purpose along with the main linear terms. Many models suggested by model identification were examined and modified. Usually, a model having the least Mallows’s Cp or BIC merits selection. The current investigation also evaluated if the model was good in reproducing the decrease in Ar3 temperatures for a few selected experimental records with the balance Ti, B, Al, N, and O content (Refs. 9–11). As a result, a few models recommended by automated model selection having the lowest Mallows’s Cp were examined, and a manual selection was made based on a few required model properties — such as negative coefficients for primary alloying elements, a positive coefficient for Si, and a negative coefficient for cooling rate — and the model satisfactorily predicting the observed Ar3 decreases in the five experimental records related to the presence of balanced Ti, B, Al, N, and O content. Overall, 257 data records on WM or low-alloy steels were combined in an R data frame to determine a new expression for Ar3. The rest of the analysis used 595 data records obtained after applying various filters to Evans’s database. Results and Discussion Various data properties and the results of correlations with the new regression formulae for the Ar3 temperature are described below. Initially, the elemental levels in Evans’s database were plotted as shown in Fig. 4. These Trellis plots revealed that carbon content in the data was largely centered around 0.075 wt-% while manganese content seemed to form two major clusters. Cr, N, Mo, Cu, Nb, Al, and B contents largely clustered close to 0 wt-% though much higher numbers were found in some data records. Si, Ti, N, O, S, and P contents seemed to form largely single clusters with few outliers. V content seemed to cluster largely around two values. These clusters and outliers would create similar clusters or spreads in dependent variables such as UTS, YS, and CVN at 28 J. CVN test temperatures at 28 and 100 J vs. UTS showed wide scatter. A formal analysis following data preparation allowed an assessment of the effect of certain predictors on WM UTS and Ar3 temperature. Fig. 10 — A plot of UTS against new Ar3 values (Equation 19) in clusters with a strong positive correlation. 44-s | WELDING JOURNAL

UTS vs. Predictors Initially, a regression of UTS using multiple linear terms was obtained as follows: () = 294 + 844() + 112.8() + 79.2() + 72.3() + 32.4() + 83.3() + 151.5() + 1773.3() + 995.1() + 855.3() + 289() + 3638.2() Figure 5 shows the UTS fit given by Equation 12 plotted against reported WM UTS values in Evans’s database. It is interesting to note a near linear relationship with less scatter in UTS between the 450 and 600 MPa range, which is consistent with the data on ferritic-pearlitic steels shown in Fig. 1. This UTS regression was performed over 595 records in Evans’s database obtained after applying several filtering conditions as described earlier. Simple regression that included most of the elements except P and S showed an adjusted R2 of 0.9118. This is a pretty good fit despite a few key unknowns about the WM, such as (average) prior-austenite grain size and cooling rate. The p-values of all the coefficients were well less than 0.05. The coefficients of B, O, and S were not significant. The coefTable 4 — Minimum and Maximum Limits for Elemental Composition and Cooling Rate Element Minimum (wt-%) Maximum (wt-%) C 0.024 0.792 Si 0 2.04 Mn 0 2.52 P 0 0.11 S 0 0.046 Cu 0 2.04 Ni 0 3.49 Cr 0 2.8 Mo 0 1.11 Nb 0 0.098 V 0 0.099 Ti 0 0.069 B 0 0.02 Al 0 1.55 N 0 0.0270 O 0 0.118 Cooling Rate (°C/s) 0.001 30 (12) MARCH 2023 | 45-s

ficient for P was barely significant, and its effect was small; so, it was excluded from the equation. The regression shown in Equation 12 for WM UTS appeared similar to the following regression equation by Mesplont (Ref. 29), which is suitable for high-strength bainitic steels with C content below 0.8 wt-% and the following range of compositions: Mn < 2 wt-%, Si < 1.8 wt-%, Cr < 2 wt-%, Mo < 0.8 wt-%, Cu < 1.6 wt-%, Ti < 500 ppm, P < 700 ppm, Nb < 800 ppm, and B < 30 ppm. () = 288 + 803() + 178() + 83() + 1326() + 60() + 122() + 320() + 2500() + 180() + 36000() The adjusted R2 of the fit improved when the grain size was approximately calculated and added to the model. This can be done using the equation for YS that contains grain size (Ref. 5) and further making an assumption about dissolved nitrogen. UTS is also found to be linearly correlated with YS. Regression expressions for CVN at 28 and 100 J could not be identified. Ar3 vs. UTS Subsequently, the coalesced Ar3 values assigned to 595 records in the analysis set were correlated with experimental values for WM UTS reported in Evans’s database. A regression of computed Ar3 against UTS showed a poor fit with an adjusted R2 of only 0.304, as shown in Fig. 6. This poor fit is understandable as Ouchi et al. and Trzaska’s equations for Ar3 do not include all principal and minor alloy additions, cooling rate, or (average) prior-austenite grain size. During the filtering process, records having a computed Ar3 value lower than 680°C were excluded. If these records were included, they would appear as strong outliers in Fig. 6. The correlation factor between the initial computed Ar3 and WM UTS was –0.551. As mentioned earlier, Evans’s WM dataset did not provide a direct correlation between computed values of Ar3 and experimental values of WM UTS. Consequently, the proposed new expression for Ar3 was formulated as discussed below. Ar3 Regression Using Ilman’s Experimental Data The 13 experimental Ar3 (i.e., TS ) values provided separately by Ilman et al. (Refs. 9–11) are well explained by cooling rate and material compositions, although certain elemental weight percentages remain constant. To mitigate the adverse effects of multicollinearity issues, the data was augmented using a few records reported in Lolla et al. (Ref. 35), Vega et al. (Ref. 36), and Deva et al. (Ref. 38). A good Ar3 regression model for this small dataset was as follows: !"(°) = 1185.95 − 6.55() − 534.02() + 394.05() − 330.45() + 1586.63() + 64477.78() + 3093.8() + 4606.92() + 7827.45() − 24156.63( #) − 842.71>√@ − 1502.42>√@ − 1653160.12( × ) − 79158.05( × ) − 91474.94( × ) where CR is the cooling rate in °C/s. The cooling rate for Illman et al.’s experimental weld data is set at 13 °C/s. The adjusted R2 of the above regression is 0.95. This regression expression indicated that cooling rate and a few of the major elements (in wt-%) are good predictors of Ar3. The decreases observed in some experimentally determined values of TS were especially well predicted by this model. The prediction error is less than 1% for all the data included in the model. It was mentioned earlier that a balanced addition of Ti, B, Al, N, and O significantly decreased the measured TS vs. the calculated Ar3 values. This can be observed in the regression model without the inclusion of the cooling rate as a predictor. As shown in Table 2, the sum of Ti, B, Al, N, and O minor alloy additions decreased as the experimental TS value increased in this small dataset. This seemed to be a good indicator. But the adjusted R2 of the model, including only the alloying elements, was not high even though it offered good predictive capabilities. Once the cooling rate was added as a predictor and the effects of the multicollinearity issue were mitigated, the adjusted R2 of the model improved significantly. Ar3 Regression Using Composite Data It is quite apparent that more experimental data is needed to refine the Ar3 regression from Ilman et al.’s data to achieve higher reliability and to make the formula applicable over a much wider data range. The 75 experimental records provided by Mintz et al. (Ref. 28) were included first. Additionally, 21 records in Evans’s dataset that were in Salganik et al.’s (Ref. 26) data range were selected and assigned Ar3 values using Salganik et al.’s modified Ar3 expression (Equation 10). Another 23 records from Evans’s dataset with coalesced Ar3 values were added to constrain the regression at applicable data boundaries and interior points and to guide the regres- (13) (14) Fig. 11 — A plot of UTS against new Ar3 values (Equation 20) in another cluster with a very strong negative correlation. 46-s | WELDING JOURNAL

sion process. Additional experimental data reported in Refs. 29, 30, and 33–43 were also included. Appendix I provides details of the respective datasets. Additionally, 20 numerical steel records with random element concentrations at a very low range around 0.01 for primary elements and around 0.0001 for minor elements were added using Ouchi et al.’s expression (Equation 4). Overall, 257 records were collected to obtain a new Ar3 expression using a multiple linear regression. The experimental data along with coalesced Ar3 values at selected extreme and interior points in Evans’s dataset served as guideposts to regression in lieu of full experimental data on WM transition temperatures in Evans’s database. The regression expression for Ar3 from this composite dataset was obtained as follows: !"(°) = 906.49 − 2.78() − 439.3() + 34.17() − 36.7() − 8.5() − 51.2() − 27.08() − 63.48() − 1765.95() − 520.29() − 2401.12() − 1784.44() + 21.89() + 5300.15() − 420.96() + 297.07(#) − 16.4(#) + 11668.54(#) + 458.21(√) − 1142.45(√) + 298.91(√) where CR is the cooling rate in °C/s. The adjusted R2 of this fit was 0.9087. The standard error of the residuals was 24.89. The intercept value of this new Ar3 expression was close to 910°C, the Ae3 (equilibrium austenite-ferrite transformation) temperature for pure iron. Several nonlinear terms were included to cover a wide range of 14 elemental compositions (except P and S) in Evans’s dataset. Most of the p-values of the intercept and the coefficients of C, Mn, Ni, Cr, Mo, Si, Ti, Nb, N, C2, Mn2, Nb2, Al, V, √Ti, √O, and √N were very significant or marginally significant. The coefficients of Cu and O also did not have p-values below 0.05, but they were left as indicated by regression because their values were in the expected range. The new Ar3 regression equation predicted the decreases observed in measured values of TS reported in the experimental data in Table 2 within a 2 to 3% error range in most cases. This equation is also likely to predict Ar3 values over the entire range of Evans’s dataset reasonably well. The above new Ar3 regression equation for WM UTS appears similar to Miettinen’s (Ref. 44) Ar3 regression equation, which also includes several nonlinear terms. It is also partly similar to the complex expressions for transition temperatures reported by Kasatkin et al. (Ref. 31). Other nonlinear terms and cross terms are recommended in the context of critical temperature and CCT diagram models by Miettinen et al. (Ref. 45). The regression equation by Miettinen et al. (Ref. 45) also refers to the cross term Cu × B, among others. The model from the current investigation also predicted the significance of this cross term. The quadratic term Nb2 is also recommended by Yuan et al. (Ref. 46). For the sake of simplicity and to assess the impact of major and minor elements in Evans’s database, the new regression expression for Ar3 shown in Equation 15 was felt sufficient for the intended purpose. The new Ar3 regression formula is applicable over almost the entirety of Evans’s database. The related data limits for various elements and weld cooling rate are shown in Table 4. Figure 7 shows a plot of UTS against the proposed new Ar3. A linear regression of UTS against the proposed new Ar3 is indicated by the straight line in Fig. 7. The line passes through the data better compared to the one indicated in Fig. 6. The adjusted R2 of the fit of the new Ar3 against UTS is 0.499 compared to 0.3058 in the previous fit in Fig. 6. This indicates that the new Ar3 is likely to better predict UTS. The regression relation is the following: () = 1242.93 − 0.91 !"(°) Cluster Analysis The scatter in Fig. 7 is attributed to various inherent metallurgical characteristics of different clusters of experimental data in Evans’s dataset. Many of these metallurgical characteristics include the amount of free nitrogen; size and composition of the inclusions and precipitates; effect of tempering caused by the deposition of over-lying runs; microphase morphology (e.g., the form of the carbides); relative grain sizes of the coarse-grained, fined-grained, and the intercritical regions; etc. (Ref. 47). The individual experimental clusters were indicated by weld ID tags in the data with different colors in Fig. 7. The scatter can be better explained when each experimental cluster is examined and combined with similar individual experimental clusters. Evans’s database contains 73 weld experimental clusters. The pruned dataset of 595 records contains 55 weld clusters. These individual weld clusters exhibit various trends in UTS vs. the new Ar3 values. Some individual weld clusters have strong down trends, several of them have even up trends, and many of them do not have strong trends in UTS vs. Ar3. Some weld clusters have only one or two data points. The rest have three or more data points. Table 5 shows the statistics on selected individual weld clusters and the regression of UTS vs. the new Ar3 of each weld cluster using the R2 value, intercept and slope of the line, and trend indicator. Table 5 also shows the minimum and maximum WM UTS (in MPa) in the respective weld series and the corresponding number of data points in the weld series. The above individual clusters can be grouped further. This was accomplished manually using TIBCO Spotfire® (Ref. 48). When the individual weld clusters are grouped, the information can be condensed in a few charts to explain the scatter in Fig. 7, and the impact of combining various microelements on UTS can also be understood more clearly. In general, the trend between UTS and Ar3 temperature should be as indicated in Fig. 1. However, Evans’s WM data contains at least four types of composite weld clusters: a cluster of clusters with a strong or moderate negative correlation between UTS and Ar3, a cluster of clusters with a positive correlation between UTS and Ar3, and a composite cluster of clusters with a very strong negative correlation. The respective data are illustrated in Figs. 8–11. (15) (16) MARCH 2023 | 47-s

Table 5 — Regression and Summary Statistics of Selected Weld Series in Evans’s Database Weld ID R2 Intercept Slope Trend Minimum UTS (MPa) Maximum UTS (MPa) Count AB 0.251 81.238 0.493 Up 454 499 12 ACoNb 0.876 1600.655 –1.390 Down 481 705 9 ACoV 0.910 1355.749 –1.076 Down 481 590 10 ACrTi 0.766 1473.870 –1.276 Down 494 597 11 ACuTi 0.623 4176.206 –4.508 Down 492 686 6 Al 0.658 –629.981 1.520 Up 529 570 9 AlN 0.657 –627.983 1.518 Up 529 570 9 AlO 0.616 –517.523 1.375 Up 524 581 26 AlTi 0.341 1122.213 –0.747 Down 514 622 33 AlTiN 0.407 –1588.414 3.030 Up 583 622 6 AMoTi 0.840 2333.428 –2.248 Down 492 694 20 ANiTi 0.714 884.939 –0.475 Down 492 562 15 AoPlus 0.281 942.227 –0.531 Down 494 619 8 AO-Ti 0.382 1436.715 –1.177 Down 447 523 15 BN 0.196 1112.977 –0.786 Down 521 561 7 CB 0.453 1397.047 –1.135 Down 521 617 44 CMn 0.898 1610.519 –1.382 Down 462 602 12 CoPlus 0.740 1413.219 –1.177 Down 526 658 5 Cplus 0.473 1296.832 –0.965 Down 533 672 5 CrMo 0.899 1570.725 –1.357 Down 501 657 5 48-s | WELDING JOURNAL

In general, the experimental series indicated by weld IDs in the data can be joined to form a few major clusters. However, the conventional cluster analysis, such as K-means clustering (Ref. 24), cannot be performed for this dataset. The weld series can be combined when their UTS vs. Ar3 regression slopes are nearly the same and if the distances between two regression lines are small or their combined regression has nearly the same characteristics, such as adjusted R2 before and after adding in the new weld series to the composite group. The trends of the weld series in the composite clusters were manually identified, and the weld series were grouped Table 5 — (continued) Weld ID R2 Intercept Slope Trend Minimum UTS (MPa) Maximum UTS (MPa) Count MnCr 0.854 1554.366 –1.330 Down 466 676 8 MnMo 0.904 2190.718 –2.089 Down 466 623 9 MnNb 0.843 1506.371 –1.219 Down 494 685 17 MnNi 0.692 1156.997 –0.823 Down 466 569 10 MnOx 0.881 1086.432 –0.748 Down 475 555 15 MnSi 0.301 1362.136 –1.023 Down 453 639 14 MnTi 0.777 1215.296 –0.879 Down 492 577 17 MnV 0.847 1504.494 –1.210 Down 494 681 20 S 0.393 –2332.078 3.723 Up 523 546 5 Tab 0.205 1133.158 –0.756 Down 404 638 45 Ti 0.678 1822.660 –1.669 Down 537 654 9 TiB 0.474 1349.502 –1.066 Down 521 617 39 TiBAlN 0.554 4341.670 –5.254 Down 594 660 4 TiBN 0.474 1349.502 –1.066 Down 521 617 39 TiN 0.517 1335.580 –1.041 Down 528 597 10 TiOX 0.466 –560.604 1.473 Up 539 596 7 Zn 0.242 1442.055 –1.186 Down 528 545 4 MARCH 2023 | 49-s

using Spotfire. The underlying weld series are indicated in the legends in Figs. 8–11. The respective simple regression equations of the four major clusters between UTS and Ar3 are as follows: Strong negative correlation () = 1528.9 − 1.25!"(°) Moderate negative correlation () = 1330.97 − 1.05!"(°) Strong positive correlation () = −475.1 + 1.33!"(°) Very strong negative correlation () = 2363.94 − 2.29!"(°) The R2 values of the regression lines in Figs. 8–11 were 0.727, 0.814, 0.638, and 0.913, respectively. Most of the remaining weld series can be combined into another composite cluster. The adjusted R2 value of the regression of UTS vs. Ar3 in this cluster was 0.55. Figures 8, 9, and 11 indicate that UTS decreased with Ar3 in general. Figure 10 indicates that UTS increased with Ar3 for a small number of cases, but this increase strongly depended on WM elemental compositions. The element combinations in the first cluster (Fig. 8) are likely to yield quite predictable weld strengths. The element combinations in the second cluster (Fig. 9) with moderate negative correlation may yield stronger welds though UTS may be less predictable. Ar3 values below 680°C seemed to provide UTS above 640 MPa. Figure 10 indicates that aluminum and nitrogen contents created an unexpected positive correlation between UTS and Ar3, and the respective UTS values were below 580 MPa. It can be noted that the charts indicate different ranges of UTS for different combinations of the underlying experimental clusters. Figures 8 and 11 show clusters that can yield a maximum UTS close to 700 MPa. Weld experimental data on certain groups such as TiBAlN also indicate the possibility of achieving UTS greater than 710 MPa. These high UTS values from the TiBAlN series were not included in the analysis. So, they do not appear in Figs. 8 or 9. The above correlation or relationship between WM UTS and the new expression for Ar3 temperature that includes the effects of principal and minor alloy elements (in wt-%) and weld cooling rate (in °C/s) are expected to complement the JWES neural network currently available to predict CVN test temperature for 28 J absorbed energy based on WM chemical composition, thus providing a pair of effective tools for efficient development of welding electrodes based on an Fe-C-Mn system for high-performance applications. Conclusions Several new relationships have been obtained using the selected SMA WM dataset in Evans’s database. A new equation for Ar3 temperature in low-alloy steels containing several minor elements was derived using a novel approach. The analysis was carried out using a machine learning approach for multiple linear regression in R. Initially, all the records in this database were examined from various perspectives. The previously discussed charts and regression results indicate salient perspectives of this data and the new results reported. The SMA WM database is quite large compared to typical datasets reported in metallurgical literature. Understanding the message conveyed by this large database is quite daunting. Machine learning approaches, including automated model finding using multiple regressions, improve our productivity in discovering the relationships embedded in the data. The analysis reported here indicates that WM UTS can be correlated linearly with WM elemental composition. This regression equation improved notably after the inverse square root of approximate grain size was added as another predictor. The WM UTS had a linear correlation with YS. The test temperature for CVN at 28 or 100 J absorbed energy did not correlate fully well with elemental compositions, UTS, YS, and Ar3 temperature. Only a poor-quality regression could be obtained even after including nonlinear terms in the regression equation of CVN at 28 J. Evans’s database is supported by only a small dataset of experimentally determined Ar3 (i.e., TS ) values, and this database does not provide critical temperatures for all the reported WMs in the database. This was rectified in our effort. The 13 experimental values of records containing Ar3 (or TS ) were combined with various experimental data indicated in Appendix I, a small number of records in Evans’s dataset that obeyed data range for Salganik et al.’s expression for Ar3 (Ref. 26), and 23 extreme high-end points in Evans’s dataset. Twenty numerical records on various steels were also added at very low element concentrations, and their Ar3 values were set using Ouchi et al.’s expression (Equation 4). Using this composite dataset, a new Ar3 regression relation applicable to the entirety of Evans’s database was derived. This expression includes several nonlinear predictors, and it was designed to include a large data range in Evans’s dataset and to accommodate some nonlinearities expected in the fit and in the underlying physical phenomena. The expression also improves prediction accuracy for five records in question in Ilman’s experimental data. This expression should also predict Ar3 for all records in Evans’s database to a very good approximation. The relation between UTS and Ar3 exhibits a significant scatter and low-regression quality. Closer examination using a cluster analysis revealed that individual weld series in Evans’s database contributed to the scatter, and they can be combined into at least four clusters, namely, one cluster having a strong negative correlation between UTS and Ar3, another having a moderate negative correlation between UTS and Ar3, a third small cluster having a strong positive correlation between UTS and Ar3, and a fourth cluster having (17) (18) (19) (20) 50-s | WELDING JOURNAL

a very strong negative correlation between UTS and Ar3. The second and fourth clusters seemed to provide clues to create the strongest welds with UTS approaching 700 MPa. Future research may generate more experimental data to improve the regression result for the Ar3 expression applicable to Evans’s WM dataset. Various WM transition temperatures and cooling rates could be determined experimentally and documented for a few additional points in Evans’s WM dataset, including those with elemental composition in the high wt-% range. A full factorial design does not appear to be necessary. Acknowledgments Krishna Sampath expresses his gratitude to Dr. Glyn M. Evans for his continued encouragement, technical guidance, and support. The authors also remain thankful to the AWS peer reviewers for their insightful comments that enabled the authors to further improve the organization and content of the manuscript. References 1. Evans, G. M. 2015. Database — Weld metal composition and properties. DOI: 10.13140/RG.2.1.3628.1764 2. Evans, G. M., and Bailey, N. 1999. Metallurgy of Basic Weld Metal. Abington, UK: Abington Publishing. 3. Sampath, K. 2021. Analysis of a high-strength steel SMAW database. Welding Journal 100(12): 410-s to 420-s. DOI: 10.29391/2021.100.036 4. Yurioka, N., Suzuki H., Ohshita, S., and Saito, S. 1983. Determination of necessary preheating temperature in steel welding. Welding Journal 62(6): 147-s to 153-s. 5. Bramfitt, B. L. 1998. “Structure/Property Relationships in Irons and Steels.” In Metals Handbook Desk Edition, Second Edition, edited by Davis, J. R., 153–173. Materials Park, Ohio: ASM International. DOI: 10.31399/asm.hb.mhde2.a0003090 6. Pickering, F. B. 1978. Physical Metallurgy and the Design of Steels. London, UK: Applied Science Publishers Ltd. 7. Ouchi, C., Sampei, T., and Kozasu, I. 1982. The effect of hot rolling condition and chemical composition on the onset temperature of γ-α transformation after hot rolling. Transactions of the Iron and Steel Institute of Japan 22(3): 214–222. DOI: 10.2355/isijinternational1966. 22.214 8. Steven, W., and Haynes, A. G. 1956. The temperature of formation of martensite and bainite in low-alloy steels. Journal of the Iron and Steel Institute 183(8): 349–359. 9. Ilman, M. N., Cochrane, R. C., and Evans, G. M. 2012. Effect of nitrogen and boron on the development of acicular ferrite in reheated C-Mn-Ti steel weld metals. Welding in the World 56: 41–50. DOI: 10.1007/BF03321394 10. Ilman, M. N., Cochrane, R. C., and Evans, G. M. 2014. Effect of titanium and nitrogen on the transformation characteristics of acicular ferrite in reheated C–Mn steel weld metals. Welding in the World 58: 1–10. DOI: 10.1007/s40194-013-0091-x 11. Ilman, M. N., Cochrane, R. C., and Evans, G. M. 2015. The development of acicular ferrite in reheated Ti–B–Al–N-type steel weld metals containing various levels of aluminium and nitrogen. Welding in the World 59: 565–575. DOI: 10.1007/s40194-015-0231-6 12. Sampath, K. 2022. Metallurgical design rules for high-strength steel weld metals. Welding Journal 101(5): 123-s to 143-s. DOI: 10.29391/2022.101.010 13. Bhadeshia, H. K. D. H. 2007. Frontiers in the modelling of steel deposits. Journal of the Japan Welding Society 76(2): 24–30. DOI: 10.2207/jjws.76.102 14. Fujii, H., and Ichikawa, K. 2001. Estimation of weld properties by Bayesian neural network. Welding International 15(12): 935–939. DOI: 10.1080/09507110109549469 15. Lalam, S. H. 2000. Modelling of mechanical properties of ferritic weld metals. PhD diss., Cambridge University. 16. Lalam, S. H., Bhadeshia, H. K. D. H., Evans, G. M., and MacKay, D. J. C. 2000. Estimation of mechanical properties of C-Mn weld metals. Paper read at IIW Subcommittee IIA Autumn Meeting, Sandvik, Sweden. 17. Lalam, S. H., Bhadeshia, H. K. D. H., and MacKay, D. J. C. 2000. Estimation of mechanical properties of ferritic steel welds. Part 1: Yield and tensile strength. Science and Technology of Welding and Joining 5(3): 135–147. DOI: 10.1179/136217100101538137 18. Lalam, S. H., Bhadeshia, H. K. D. H., and MacKay, D. J. C. 2000. Estimation of mechanical properties of ferritic steel welds. Part 2: Elongation and Charpy toughness. Science and Technology of Welding and Joining 5(3): 149–160. DOI: 10.1179/136217100101538146 19. Metzbower, E. A., DeLoach, J. J., Lalam, S. H., and Bhadeshia, H. K. D. H. 2001. Neural network analysis of strength and ductility of welding alloys for high strength low alloy shipbuilding steels. Science and Technology of Welding and Joining 6(2): 116–124. 20. Metzbower, E. A., DeLoach, J. J., Lalam, S. H., and Bhadeshia, H. K. D. H. 2001. Analysis of toughness of welding alloys for high strength low alloy shipbuilding steels. Science and Technology of Welding and Joining 6(6): 368–374. 21. Metzbower, E. A., DeLoach, J. J., Lalam, S. H., and Bhadeshia, H. K. D. H. 2002. “Secondary Effects in Neural Network Analysis of the Mechanical Properties of Welding Alloys for HSLA Shipbuilding Steels.” In Mathematical Modelling of Weld Phenomena 6, edited by Cerjak, H., and Bhadesia, H. K. D. H., 231–242. London, UK: Maney Publishing. 22. Kim, J.-H., Jung, C.-J., Park, Y. I., and Shin, Y.-T. 2022 Development of closed-form equations for estimating mechanical properties of weld metals according to chemical composition. Metals 12(3): 528. DOI: 10.3390/met12030528 23. Xiong, J., Zhang, T., and Shi, S. 2020. Machine learning of mechanical properties of steels. Science China Technological Sciences 63(7): 1247–1255. DOI: 10.1007/s11431-020-1599-5 24. James, G., Witten, D., Hastie, T., and Tibshirani, R. 2013. An Introduction to Statistical Learning: with Applications in R. Berlin, Germany: Springer Nature. 25. Weisberg, S. 1980. Applied Linear Regression. Hoboken, N.J.: John Wiley & Sons Inc. 26. Salganik, V. M., Chikishev, D. N., Pozhidaeva, E. B., and Nabatchikov, D. G. 2016. Analysis of structural and phase transformations in low-alloy steels based on dilatometric studies. Metallurgist 59(9): 766–773. DOI: 10.1007/s11015-016-0172-3 27. Trzaska, J. 2015. Empirical formulae for the calculation of austenite supercooled transformation temperatures. Archives of Metallurgy and Metals 60(1): 181–185. DOI: 10.1515/amm-2015-0029 28. Mintz, B., Banerjee, J. R., and Banks, K. M. 2011. Regression equation for Ar3 temperature for coarse grained as cast steels. Ironmaking & Steelmaking 38(3): 197–203. DOI: 10.1179/ 030192310X12827375731429 29. Mesplont, C. 2002. Phase transformations and microstructuremechanical properties relations in complex phase high strength steels. PhD diss., University of Lille. DOI: 1854/333 30. Lotter, U. 1988. Aufstellung von Regressionsgleichungen zur Beschreibung des Umwandlungsverhaltens beim thermomechanischen Walzen, 17–18. Technical report, Thyssen Stahl. MARCH 2023 | 51-s

31. Kasatkin, O. G., Vinokur, B. B., and Pilyushenko, V. L. 1984. Calculation models for determining the critical points of steel. Metal Science and Heat Treatment 26(1): 27–31. DOI: 10.1007/BF00712859 32. Gorni, A. A. 2011. Steel Forming and Heat-Treating Handbook, 24. São Vicente, Brazil. DOI: 10.13140/RG.2.1.1695.9764 33. Schindler, I., Nemec, J., Kawulok, P., Kawulok, R., Sevcak, V., Opela, P., and Rusz, S. 2018. The combined effect of chemical composition and cooling rate on transformation temperatures of hypoeutectoid steels. Kovové Materiály 56(3): 163–170. DOI: 10.4149/km_2018_3_163 34. Zhang, C., Yang, Z., Enomoto, M., Chen, H., Yang, Z., and Zhang, C. 2016. Prediction of Ar3 during very slow cooling in low alloy steels. ISIJ International 56(4): 678–684. DOI: 10.2355/isijinternational. ISIJINT-2015-602 35. Lolla, T., Babu, S. S., Lalam, S., and Manohar, M. 2013. Comparison of simulated heat affected zone microstructures of niobium microalloyed steels subjected to multi-pass weld thermal cycles. Proceedings of International Seminar on Welding of High Strength Pipeline Steels, 281–297. Araxá, Brazil: CBMM, Pittsburgh, Pa.: TMS. 36. Vega, M. I., Medina, S. F., Chapa, M., and Quispe, A. 1999. Determination of critical temperatures (Tnr, Ar3, Ar1) in hot rolling structural steels with different Ti and N contents. ISIJ International 39(12): 1304–1310. DOI: 10.2355/isijinternational.39.1304 37. Rodrigues, K. F., and Lúcio de Faria, G. 2021. Characterization and prediction of continuous cooling transformations in rail steels. Materials Research 24(5). DOI: 10.1590/1980-5373-MR-2020-0519 38. Deva, A., De, S. K., Kumar, V., Deepa, M., and Jha, B. K. 2013. Influence of boron on the hardenability of unalloyed and low alloyed steel. International Journal of Metallurgical Engineering 2(1) 47–51. DOI: 10.5923/j.ijmee.20130201.07 39. Symmach, S. M. 2016. Mechanical properties of steels with a microstructure of bainite/martensite and austinite islands. Master’s thesis, Colorado School of Mines. hdl.handle.net/11124/170090 40. Zhang, Z., and Farrar, R. A. 1995. An Atlas of Continuous Cooling Transformation (CCT) Diagrams Applicable to Low Carbon Low Alloy Weld Metals. London, UK: Institute of Materials. DOI: 10.1201/9781003070405 41. Naik, A. K., Roshan, R., Arora, K. S., Shajan, N., and Mishra, S. C. 2021. Continuous cooling transformation diagram and mechanical properties in weld coarse-grain heat-affected zone of API-X70 steel. Sadhana 46(2): 88. DOI: 10.1007/s12046-021-01623-2 42. El-Shennawy, M., Farahat, A. I., Masoud, M. I., and Abdel-Aziz, A. I. 2016. Effect of boron content on metallurgical and mechanical characteristics of low-carbon steel. International Journal of Mechanical Engineering 5(2): 1–14. 43. Zavdoveev, A., Poznyakov, V., Baudin, T., Rogante, M., Kim, H. S., Heaton, M., Demchenko, Y., Zhukov, V., and Skoryk, M. 2021. Effect of heat treatment on the mechanical properties and microstructure of HSLA steels processed by various technologies. Materials Today Communications 28: 102598. DOI 10.1016/j.mtcomm.2021.102598 44. Miettinen, J. 1996. Simple semiempirical model for prediction of austenite decomposition and related heat release during cooling of low alloy steels. Ironmaking & Steelmaking 23(4): 346–356. 45. Miettinen, J., Koskenniska, S., Somani, M., Louhenkilpi, S., Pohjonen, A., Larkiola, J., and Komi, J. 2021. Optimization of CCT curves for steels containing Al, Cu and B. Metallurgical and Materials Transactions B 52: 1640–1662. DOI: 10.1007/s11663-021-02130-9 46. Yuan, X. Q., Liu, Z. Y., Jioa, S. H., Ma, L. Q., and Wang, J. D. 2006. The onset temperatures of γ to α-phase transformation in hot deformed and non-deformed Nb micro-alloyed steels. ISIJ International 46(4): 579–585. DOI: 10.2355/isijinternational.46.579 47. Evans, G. M. 2022. The effect of micro-alloying elements on the ultimate tensile stress of C-Mn multi-run welds. IIW Doc. II-2445- 2022 (II-C-619-2022), Tokyo, Japan. 48. TIBCO Spotfire®. tibco.com/products/tibco-spotfire RAJAN VARADARAJAN is with Norfolk Southern, Atlanta, Ga. KRISHNA SAMPATH ([email protected]) is with Creative Concepts, Marietta, Ga. 52-s | WELDING JOURNAL

WELDING RESEARCH https://doi.org/10.29391/2023.102.005 Effects of Filler Wire Intervention on the Gas Tungsten Arc: Part III — Process Stability Control of Wire-Filled GTAW A sensing method that collects global and local arc information was used to monitor the stability of metal transfer and weld surface height BY S. ZOU, Z. WANG, Y. CAO, AND S. HU Abstract Stability control of the welding process is necessary to guarantee weld quality. In this study, a sensing method that collects both global and local arc information was proposed to conveniently monitor metal transfer stability during gas tungsten arc welding. This sensing method was also used to monitor the stability of the weld surface height by sensing the change in global arc length. The stability factor (f mt) was calculated to quantify the metal transfer stability. The characteristic signal (U*), which represents the average global arc voltage in the presence of a liquid bridge, was extracted to characterize the change in arc length by decoupling the dynamic interference signal of metal transfer. Both a fuzzy controller and a proportional integral derivative controller were designed to control the metal transfer stability and the weld surface height. The preliminary control experiments proved the effectiveness and potential of the proposed sensing and control strategies. Keywords ■ Metal Transfer ■ Process Stability ■ Arc Voltage ■ Arc Sensing ■ Feedback Control ■ Wire-Filled Gas Tungsten Arc Welding (GTAW) Introduction Automated gas tungsten arc welding (GTAW) has matured into a well-established manufacturing technique that is frequently used in aerospace, nuclear power, etc. Its process stability has long been one of the main concerns in the welding community (Ref. 1). For GTAW in complex environments, real-time assurance of the stability of the weld surface height and metal transfer is of great significance to the quality of weld shaping. Real-time monitoring for these purposes often requires the establishment of process sensing and control systems; therefore, this work proposed a practical arc sensing method to achieve real-time feedback control of the weld surface height and metal transfer in GTAW. In general, the process stability of GTAW partly depends on the stability of the weld surface height. This height will vary for many reasons, such as alignment/fitup errors, welding torch movement or positioning errors, the accumulated unevenness from the previous pass, etc. An unduly convex weld surface is likely to cause short-circuiting contact and tungsten inclusion when the filler metal is added. Additionally, if the weld surface is too far from the tungsten electrode, the arc stability will be weakened, and this surface will also face insufficient gas shielding. The uneven height of the weld surface will increase the risk of weld defects and may impact subsequent welding. Thus, it makes sense to monitor the change in this height. A direct monitoring method has been used to capture images of the weld surface height. This method involves mounting a camera with optical filters (i.e., passive vision sensing), where the distance from the tip of the nonconsumable tungsten electrode to the weld surface can be extracted by image processing to characterize the variation in the weld surface height (Refs. 2–4). A mirror imaging method has MARCH 2023 | 53-s

also been proposed to capture the reversed reflection of the tungsten electrode under an appropriate exposure (Refs. 5, 6). This method used the mathematical specular reflection model to obtain the distance between the tungsten electrode tip and its reversed image, which was used as a variable to estimate the weld pool surface height. Another visionbased method that has received much attention involves active structured-light vision sensing, which can monitor the dynamic evolution of the surface height by processing the two-dimensional structured-light reflection images (Refs. 7, 8) or by reconstructing the three-dimensional (3D) weld pool (Refs. 9–12). Based on the reconstructed 3D coordinates of the weld pool surface, Zhang et al. (Ref. 13) as well as Liu and Zhang (Ref. 14) extracted several geometrical parameters — such as the weld surface convexity (i.e., average height of the weld pool), length, and width — to estimate the backside bead geometry. Liu and Zhang (Ref. 15) also realized the predictive control of the weld pool surface by establishing a linear model between the weld surface convexity, length, and width and the welding current and speed. In another study (Ref. 16), the empirical response of the human welder to the real-time state of the topside weld pool was modeled as a neuro-fuzzy inference system, which was used for the feedback control of the weld pool geometry. There are more nonvisual sensing methods available, including arc acoustic sensing (Refs. 17–19) and arc light sensing (Refs. 20, 21). These two methods can assess the surface height by acoustoelectric or photoelectric measurement of the arc length. All of the aforementioned are noncontact sensing methods, which means that a sensing system independent of the welding circuit needs to be established. As a result, limitations can arise when spatial accessibility, welding posture, and so forth are taken into account. The arc sensing method, which takes the welding arc itself as a sensor to extract electrical signals directly from the welding circuit without the demand for complex external systems, is easier to implement and more cost-effective in engineering applications. Based on the inherent relationship between arc voltage and arc length in GTAW (Ref. 22), this method has been applied in weld joint penetration control (Refs. 23, 24) and automatic voltage control (AVC) (Refs. 25, 26). Although it is generally believed that the change in the weld surface height can be characterized by arc voltage due to the change in arc length caused by its change, using the arc voltage to characterize the change in the weld surface height or arc length still has its limitations because the arc voltage is also affected by the electrical noise, dynamic behaviors of the liquid filler metal (if filler metal is added), etc. The filter is usually designed to filter out the highfrequency noise in the arc voltage (Ref. 27) so that arc sensing remains relatively reliable during autogenous GTAW. However, in wire-filled GTAW, the dynamic behaviors of the liquid filler metal will affect the welding arc, and the resultant nonhigh-frequency signal components will be coupled into the arc voltage (Refs. 28, 29). This coupling makes it difficult to remove the nonhigh-frequency signal components by filtering, thus reducing the monitoring accuracy of the weld surface height or arc length by arc sensing. Additionally, the molten metal at the end of the filler wire (herein referred to as welding wire) is always expected to transfer smoothly into the weld pool so as to minimize the oscillation, spatter, or burning loss caused by its interaction with the welding arc, whereas metal transfer stability is sometimes difficult to guarantee in complex conditions (e.g., variable welding posture, narrow space, or high deposition rate). Furthermore, there are relatively few convenient methods for monitoring and characterizing metal transfer stability. The following raises two issues about the process stability monitoring of wire-filled GTAW. One is how to conveniently monitor the metal transfer and the other is how to better monitor the change in the weld surface height or arc length under the interference of metal transfer based on arc sensing. To address the two issues, this work proposed a global and local arc sensing method, and two related characteristic signals were extracted and analyzed. Based on this method, additional welding experiments were conducted to present the results of a feasibility study on controlling the metal transfer stability and the weld surface height. Arc Sensing and Metal Transfer The conventional arc sensing method usually only measures the arc voltage between the base metal and tungsten electrode, which includes the global voltage drop between the cathode and anode of the welding arc. In this work, this method was extended to a dual-path arc sensing method in which the global arc voltage between the base metal and tungsten electrode (denoted as Ug) as well as the local arc voltage between the base metal and welding wire (denoted as Ul ) were measured simultaneously. For brevity and convenience of description, the sensing for Ug and Ul is referred to herein as global arc sensing and local arc sensing, respectively. Figure 1 presents a schematic diagram of the arc sensing method, in the direct-current welding mode, proposed in this work. When a current flows through the welding circuit, the measured arc voltage depends on the length and electrical characteristics of the measured arc. In global arc sensing, the weld surface height changes Ug by affecting the arc length, while the intervention of the welding wire affects Ug by changing the arc characteristics (Refs. 28, 29). The influence of Fig. 1 — Schematic diagram of the global and local arc sensing method. 54-s | WELDING JOURNAL

the welding wire on Ug is usually not static because the liquid metal droplet generated at its end may completely or partially go through a transition process (including growth, oscillation, detachment, and transfer to the weld pool) that will affect the arc characteristics to varying degrees. In addition, the influence of the weld surface height and metal transfer on Ug is superimposed, so it is difficult to distinguish them with a single Ug. This superimposition indicates the necessity of mining other sensing data. In local arc sensing, the consumable welding wire is used as a measuring terminal to measure the arc voltage in the local area of the arc, and the metal droplet exactly hangs on this consumable measuring terminal. Generally speaking, the metal transfer modes in GTAW can be categorized as uninterrupted bridging transfer, interrupted bridging transfer, and free-flight transfer. Figure 2 presents the possible states of the molten metal/droplet at the end of the welding wire at a given moment. If the molten metal at the end of the welding wire is always in contact with the weld pool, as shown in Fig. 2A, in a given period of time, it is considered to be in the uninterrupted bridging transfer mode. Since the liquid bridge is always existent during this time, it can be inferred that Ul will always tend to zero. If the state of the molten metal alternates between what is shown in Fig. 2A and B in a given period of time, it can be classified to be in the interrupted bridging transfer mode. In this mode, Ul will change periodically between a near-zero value range and a higher nonzero value range because the liquid bridge break will cause a significant arc voltage drop between the two local arc sensing measuring terminals. Additionally, if the liquid bridge never appears during metal transfer in a given period of time, it is considered to be in the free-flight transfer mode, as presented in Fig. 2C. In this mode, Ul will vary in a nonzero value range, and when the liquid droplet with a volume detaches from the welding wire, Ul will increase rapidly because the detachment causes the welding wire to be far from the weld pool. In addition, as shown in Fig. 2D, the pendent metal droplet at the end of the welding wire will oscillate due to internal and external forces as its volume increases, resulting in different degrees of Ul fluctuation. The aforementioned analyzed theoretical change process of Ul from bridging transfer to free-flight transfer is shown schematically in Fig. 3. Broadly speaking, it kept at a long-term near-zero value range; began to produce a positive pulse; presented a waveform similar to a square wave; generated an apparent fluctuation with a negative pulse; and, finally, fluctuated at a nonzero value range without any pulse. The whole process corresponds to the change of the metal transfer process from stable to unstable, and in this change, the existing time of the liquid bridge became shorter and shorter, and the droplet oscillation became more and more apparent. In short, the metal transfer process was expected to be reflected in the measured Ul to some extent. Experimental Design Based on the proposed sensing method, the experimental system was established. The welding system of this work mainly consisted of a Fronius MagicWave 4000 GTAW power supply that worked in the constant-current mode, a GTAW Fig. 2 — The possible states of the liquid metal/droplet: A — Bridging state; B — growing state; C — free-flight detachment state; D — oscillation state. A C D B MARCH 2023 | 55-s

torch, and a workpiece as the base metal. The sensing system was mainly composed of a computer, a USB-4711A data acquisition card, a Hall current sensor (used to determine that the voltage change was not caused by the current fluctuation), and two Hall voltage sensors. The signal processing and feedback controllers were realized by MATLAB® programming. The original sampling frequency of the electrical signals was 1024 Hz, at which the detailed information of metal transfer was obtained. The collected electrical signal data was filtered by a twelfth-order Butterworth low-pass filter with a passband frequency of 100 Hz, a cutoff frequency of 150 Hz, and a 40-dB attenuation. The experimental materials mainly included 4-mm-thick Q235B mild steel plates, an AWS ER70S-6 welding wire with a diameter of 1.2 mm, shielding gas of pure argon (99.99%), etc. Bead-on-plate welding experiments were conducted with the welding torch moving and the workpiece staying stationary. Table 1 shows the overall arrangements of these welding experiments. Experiment A1 was mainly used to explain the feature extraction of arc voltage signals. Experiments B1 to B3 were used for signal analysis. Experiments C1 and C2 were aimed at the control of metal transfer. Experiments D1 to D4 plus Experiments E1 and E2 were designed for the control of the arc voltage and weld surface height, respectively. In some of the designed experiments, the height of the welding torch was adjusted to change the electrode-to-workpiece distance (ETWD) gradually (e.g., A1, C1) or quickly (e.g., step change in C2). In some other experiments, to analyze the signal changes during welding or verify the effectiveness of the feedback control, a concave pit or surface step was preset on the workpiece (e.g., B1 to B3, E1 to E2) to purposely alter the workpiece surface height. Feature Extraction and Analysis Signal Extraction Figure 4 shows the arc voltage data collected under the gradual change in the ETWD in Experiment A1. Due to the inherent relationship between the arc length and arc voltage, the measured global arc voltage Ug first increased and then decreased gradually as the welding torch height changed. With the increase in ETWD, the overall fluctuation amplitude of Ug became significantly larger, a result that mainly stemmed from the influence of metal transfer. Meanwhile, the liquid bridge formed and broke with varying frequency, which caused the measured local arc voltage Ul to fluctuate violently and switch back and forth between a high value range and a low value range. It can be seen from the partial enlarged view that at the moment when the liquid bridge formed or broke, the change in Ul was as large as 4 to 5 V, far exceeding the change in Ug. This information was useful to determine whether the liquid bridge exists at a sampling moment. If the liquid bridge existed for a longer time, the metal transfer mode was more inclined to be an uninterrupted bridging transfer (i.e., the metal transfer tended to be stable). If the liquid bridge existed for a shorter time or even zero, the interrupted bridging transfer or even the free-flight transfer accompanied by pendent droplets with a large volume was more likely to occur (i.e., the metal transfer developed toward instability). Thus, the metal transfer stability in a period of time could be quantified by the ratio of the time when the liquid bridge was absent to this period of time. This ratio was defined as the stability factor of metal transfer (denoted as f mt) herein, and its expression is as follows: Fig. 3 — Schematic image of a typical change trend in local arc voltage Ul under different metal transfer modes. 56-s | WELDING JOURNAL

fmt = k∕N where N is the length of the sampling window, and k is the total number of sampling points collected when the liquid bridge did not exist in this sampling window. Because of the impedance in the welding circuit, when the liquid bridge existed, Ul was actually not equal to zero, and a threshold was set to 2.9 V to represent that the liquid bridge was considered to form between the welding wire and molten pool if Ul was less than 2.9 V. The comparison between Ul and the threshold could be conducted in the sampling window, and the total number of Ul larger than the threshold, which was the value of k, was obtained. In this way, the calculated f mt was between 0 and 1. To characterize the change in arc length more accurately, it was necessary to decouple the dynamic interference of the metal transfer on Ug as far as possible. Since the specific moments when the liquid bridge existed could be obtained by the threshold judgment of Ul , the arc length could be characterized by using Ug collected only at these moments (denoted as Ug ’ ). As can be seen in Fig. 4, compared to Ug, the signal fluctuation generated with the gradual increase in ETWD was significantly suppressed in Ug ’ . This way of removing the signal interference from Ug is different than some other filtering methods in that the latter are likely to only smooth the interference signal components related to the unstable metal transfer instead of decoupling them. Ug ’ could be further smoothed afterwards. If all of the Ug ’ collected in the sampling window of length N are expressed as [Ug ’ (1), Ug ’ (2), …, Ug ’ (N–k)], the arithmetic mean can be obtained as follows: U * = $Ug' (1) + Ug' (2) +... Ug' (N – k)./(N – k) U* can then be used for the characterization of arc length. The schematic diagram of signal extraction is shown in Fig. 5. Herein, N was set as 256, so the stability factor f mt and arc voltage U* were calculated every 0.25 s. Figure 6 shows the f mt and U* extracted from the arc voltage data measured in Experiment A1. The distribution of f mt was in line with the actual change of metal transfer. The closer f mt was to zero, the more stable the metal transfer became. Due to the shaking of the welding torch at some moments and the deformation of the workpiece in this experiment, U* did not change strictly with the change in the nominal ETWD. Signal Analysis Figure 7 shows the arc voltage U* collected in Experiments B1 and B2. In these two experiments, the welding arc was purposely extinguished before its center had crossed the preset step on the workpiece surface, and the arc crater was formed just before this step. Although the weld surface height was relatively consistent, U* had still shown an upward trend before the arc extinction in Experiment B1 and a downward trend before the arc extinction in Experiment B2. Since the welding arc was bell shaped and its anode was an area rather than a point, U* was not only affected by the weld pool surface height directly below the tungsten electrode but also by the surface height of the entire anode enveloped by the welding arc. Before the arc center crossed the step area, since this area had been enveloped by the arc, U* produced a change. Figure 8 presents the arc voltage U* in Experiment B3. Similarly, due to the preset pit on the welding path, if the arc was extinguished when the weld pool was about to cross the pit, U* also showed an upward trend. After the arc crossed the pit, U* dropped and became relatively flat. The change in workpiece surface height produced a new flow in the weld pool. (1) (2) Fig. 4 — Arc voltage signals measured in Experiment A1. MARCH 2023 | 57-s

Table 1 — Welding Experiment Design and Parameters Experiment No. Welding Current I (A) Welding Speed v (cm · min–1) Wire Feed Speed vf (cm · min–1) ETWD* (mm) Gas Flow Rate Q (L · min–1) Remarks A1 155 7 140 [5, 7.5] 12 Gradual change in the ETWD B1 165 10 180 5.5, 6.5 12 Preset a surface step B2 165 10 180 6.5, 5.5 12 Preset a surface step B3 160 10 150 5.5 12 Preset a surface pit C1 155 7 — [5.5, 7.5] 12 Gradual change in the ETWD C2 165 7 — 6, 7 12 Step change in the ETWD Experiment No. Welding Current I (A) Welding Speed v (cm · min–1) Reference Arc Voltage Ur (V) ETWD (mm) Gas Flow Rate Q (L · min–1) Remarks D1 170 8 17.4 6.5 10 — D2 170 8 17.2 6.5 10 — D3 170 8 17.3, 17.1 6.5 10 Negative step of Ur D4 170 8 17.2, 17.5 6.5 10 Positive step of Ur E1 170 8 17.3 — 10 Preset a surface step E2 170 8 17.1 — 10 Preset a surface pit 58-s | WELDING JOURNAL