Bryan Goldsmith, Associate Professor and Associate Chair of Chemical Engineering, College of Engineering
Watch RecordingBryan Goldsmith did not arrive at AI through a sudden conversion. He arrived through a slow, deliberate accumulation of domain-driven decisions, each one pushing his research a little further toward data-centric methods while staying anchored in the chemical engineering problems that had motivated him from the beginning. His journey is a case study in how to integrate machine learning into a field without losing sight of what the field actually needs to know.
The foundational strategic choice was made before he arrived at Michigan. Goldsmith had trained as a chemical engineer — quantum mechanical modeling, catalysis, materials — and he recognized that his competitive advantage would never lie in developing new algorithms. “I can’t compete with the computer scientists. I can’t compete with the statisticians.” What he could do was bridge both fields — use their methods in his domain to answer questions that neither could address alone. That framing shaped everything downstream: who he hired, who he collaborated with, which problems he chose to work on, and how he taught.
His first foray into data-driven methods came at the Fritz Haber Institute in Berlin, where he worked with a machine learning researcher to apply subgroup discovery — an algorithm for finding local patterns in data rather than fitting a global model — to catalysis problems. The collaboration gave him both a method and a template: find an ML expert, learn from them, and apply their tools where you have domain leverage. He carried that template back to Michigan, hiring graduate students and postdocs with machine learning backgrounds to cover areas he lacked, collaborating with professors around the world, and teaching a machine learning course that forced him to stay current with a rapidly moving field.
One of the recurring workhorses in Goldsmith’s lab is Bayesian optimization — a technique that starts with a small initial dataset, trains a probabilistic model, estimates uncertainty across the unexplored space, and selects the next data point to evaluate based on an optimization criteria. In collaboration with Prof. Suljo Linic, they applied BO to the discovery of iridium-aluminum oxide electrocatalysts for water splitting — a material with potential in the renewable hydrogen economy. This loop allowed the team to survey doped iridium oxide alloys across the periodic table without computing every combination individually. The process identified tantalum, tungsten, and molybdenum doping as promising, and drew specific attention to iridium-molybdenum oxide, which had been underexplored in the literature. The resulting synthesized catalyst proved more active and more stable than baseline iridium — and cheaper. The Bayesian loop did not replace chemical intuition. It directed it toward regions of the search space that intuition alone would not have prioritized.
More recently, the lab has turned to foundation models — large pre-trained models for atomic simulation — to run molecular dynamics on systems that quantum mechanical methods cannot reach in terms of length and time scales. The question, in one project, was about catalyst stability: why do platinum nanoparticles on metal oxide supports migrate, coalesce, and lose surface area over time — a process called Ostwald ripening — and which support materials resist it? Foundation model simulations allowed the team (with Prof. Suljo Linic) to screen hundreds of supports, extract physical descriptors using interpretable machine learning, and predict that barium oxide — an understudied material — would outperform the state-of-the-art. The combination of large-scale simulation and interpretable downstream analysis was deliberate: simulation opens the search space; interpretability closes it by extracting the mechanistic insight that makes the result actionable. “I’m not trying to make the most accurate method. I’m not trying to make the next algorithm,” Goldsmith said. “Where could I make a contribution?” That question, asked honestly at the beginning and revisited throughout, has kept the work both technically ambitious and scientifically grounded.