ADVANCED MATERIALS & PROCESSES | NOVEMBER/DECEMBER 2025 30 must edit the drafts. Keep in mind that the same prompt can produce different answers. Watch for hallucinations. Manage the process. At The Timken Company, we have large language models (LLMs) behind a firewall that can be used to gain insights. In our Quality Lab there are 30 people who can pull in data. To mitigate “junk in/junk out,” AI helps to provide a continuity to the quality of the data without worrying that someone “fat fingered” it. This gives us more confidence in the model. There may be value in mining the wealth of historical notebooks, but each company must determine the cost/benefit analysis of spending the time and money. What is the value proposition? Do you go back five years? Or determine what the best data is and only digitize that. Don’t overwhelm yourself with data. The goal is to develop optimi- zation with limited data. Maniruzzaman (Caterpillar Inc.): At Caterpillar, AI/ML activities are centered on simulation-driven development for engineering applications. We are also applying predictive analytics to optimize heat treatment, alloy development, and structural analysis. We have been using ML for durability predictions and process optimization in materials engineering. There are still challenges to validating data. Start small and validate what AI provided. If it gives two answers, find out the confidence level of each. Then build your own confidence by keeping the human in the loop to make the final determinations. Create your own specialized model based on your own data. Use tribal knowledge and convert it to PDFs or a database where AI can understand it. Focus on digitizing quality information not quantity. Hodges (ALC Inc.): Using internal AI tools and the help of a partnering company, ALC Inc. created “George” to simulate George Bodeen’s professional voice and leadership philosophy. Bodeen was a well-respected president of Lindberg Steel Treating Co. and past president of both ASM International and MTI. “George” was posed a question about his career at Lindberg and the answer came from a Bodeen-like voice. The bot had been trained on archival data but with conversational output. The demo showcased how AI preserves industrial legacy knowledge and illustrates human-AI collaboration in professional storytelling. There are various levels of AI agents. They can range from a simple co-pilot model to something like “George” that has more guardrails and parameters, and up to fully autonomous AI agents. Most seasoned employees have a built-in “bologna detector.” We can spot when something is off. It’s difficult to train the next generation of workers how to spot hallucinations and inaccurate information. Sebastian (QuesTek In- novations LLC): The evolution of modeling started with ML in the late 2000s. There was limited data available. But ML could be used to get to potential options quicker and then staff could follow up with atomistic experiments. Modeling has grown with the entrance of AI. QuesTek uses advanced methods such as integrated computational materials engineering (ICME). It is a physics-based approach augmented with AI. ML and AI can eventually help us develop a database that can predict all alloys based on the periodic table; but we are still filling in the gaps. Most of us have found that AI can write a summary of complex topics. But currently, it is better at interpreting than extracting meaning. The process still needs a human behind the operation for decision making and intuition. Regardless, these are exciting times for data science and materials science. Lurger (QATM, a Verder Group): At QATM, AI is used in hardness testing operations (e.g., for Vickers, Knoop, and Brinell) for image analysis and evaluation. AI is capable of template recognition, which provides increased automation; it allows for independence of the sample surface for increased accuracy; and it reduces operator influence, which saves time. QATM achieved 40% lower standard deviation on reference material thanks to AI. In addition, we achieved an increase in detection on etched surfaces from approximately 5% to 95%. Yet, limitations exist. It is still difficult to ensure quality through retraining and blind trust in the result. For example, when new versions of software are released, it is challenging to trust and verify the system. We need customer cooperation to achieve the best solution. ~AM&P GET ENGAGED, GET INVOLVED, GET CONNECTED The ASM Materials Informatics Committee is an active group of ASM members working together to drive industry-wide efforts in implementing artificial intelligence and machine learning methods to accelerate the design, discovery, and deployment of new materials, processes, and analysis techniques. Committee projects include developing a new ASM Handbook volume and organizing IMAT conference programming. Members with similar interests are welcome to join. For more information, contact staff liaison Scott Henry at matinfo@asminternational.org. Moderators and panelists from left: Larissa Vilela Costa, Lee Rothleutner, Lucas Lurger, Craig Hodges, Trisha Rouse, Jason Sebastian, and Mohammed Maniruzzaman.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==