Deixis Online

Three different stable configurations of sulfate electrolytes (red and yellow spheres) to layered surfaces of copper (brown spheres) and gold (gold and medium-brown spheres) determined using experiments and simulations. Electrochemical surfaces are important in understanding batteries, fuel cells, sensors and catalysts. (Reprinted with permission from Langmuir 2024, 40, 4914−492. Copyright 2024 American Chemical Society.)

Tristan Maxson had an intriguing question: Could a high-performance machine learning model trained only on the physics of liquid water also accurately simulate water’s ice and gas phases? The authors of a 2023 paper concluded it couldn’t. For Maxson, it was the perfect opportunity to test a new machine learning interatomic potentials (MLIP) model at the limits of the possible.

“The model accurately predicted the gas, liquid and ice phases,” says Maxson, a computational chemical engineering Ph.D. student at the University of Alabama. “We were really surprised that it worked for the crystalline ice phase. The model was able to learn the physics behind it without being directly taught this.”

The results, published in 2024 in the Journal of Physical Chemistry Letters, are an important proof-of-concept of MLIPs as new, powerful simulation tools to accelerate research in areas from biology to more efficient and durable energy materials.

An interatomic potential is a set of mathematical rules that describes the complex dance of forces between atoms — how atomic bonds are made, broken and vibrated. It’s the bedrock of modeling in chemistry and materials science.

Yet computational researchers have faced a conundrum in usefully applying interatomic potentials, Maxson says. Models based in classical physics are computationally efficient but miss much of the physics by simply treating atoms as balls connected by springs. Quantum mechanical methods, such as density functional theory (DFT), capture atoms’ true electronic structure but are too computationally slow for large systems.

“The promise of MLIPs is that you get quantum accuracy at classical speeds,” Maxson says. “It’s revolutionary in terms of what you can imagine you could model.”

In a recent paper, Maxson reported modeling an atomic system 10 times larger than traditional DFT simulations and for significantly longer. “In the time a normal DFT model wouldn’t be able to do one short simulation, with MLIPs we did hundreds of much longer simulations.”

Maxson’s path to MLIPs began with a Purdue University summer internship that grew into an eight-year, part-time research assistant staff position in chemical engineer Jeffrey Greeley’s lab. The job provided invaluable experience in creating and troubleshooting high-throughput computational workflows on Purdue’s high-performance clusters.

It also led to his first paper, The Atomic Simulation Environment (2017), already a cornerstone in the field. “It’s a Python software package that helps you set up atomic simulations for all kinds of atomistic codes,” says Maxson, who in parallel earned undergraduate and master’s degrees in chemistry at Ball State University doing extensive experimental research and teaching.

At home sick with COVID in 2020, he tweeted that he was looking for a Ph.D. advisor in a chemical engineering program. Based on that first paper’s influential following, he received responses from more than a dozen professors from Britain to Korea. He chose to join University of Alabama chemical engineer Tibor Szilvási’s new lab group.

In the first year of his Ph.D. program, Maxson explored several topics before landing on MLIPs, a hot emerging field. He also received the Department of Energy Computational Science Graduate Fellowship (DOE CSGF). In 2022, at his first CSGF review meeting — a chance for fellows, alumni and DOE researchers to connect — Maxson met 2020–2023 CSGF recipient Albert Musaelian, a MLIP expert at Harvard University’s School of Engineering and Applied Sciences.

Musaelian described Allegro, the MLIP he developed during his fellowship. It was an ideal addition to Maxson’s PhD research — a cutting-edge MLIP to be tested.

“I work on the applied side of these models,” Maxson notes. “My goal is to figure-out how to best use the method.”

In the water-phase research, he demonstrated that Allegro has remarkable “chemical intuition,” or the ability to accurately interpolate from first principles to accurately predict known values it hasn’t been trained on. (This in contrast to the large-language-model phenomenon of hallucinations in which a machine learning model extrapolates fictional results.)

Subsequently, Maxson explored whether Allegro could simulate physical experimental results. He and Szilvási chose to model supported metallic nanoparticles, pillow-shaped metallic particles that are critical in catalysis and sensor applications. Until then, computational modelers had been able to simulate only idealized scaled-down versions.

“We were able to take an experimental result and demonstrate that machine learning potentials could reproduce it,” says Maxson of research published in 2025 in Angewandte Chemie Novit. “Now we’re really getting to the point where we can model the real nanoparticle.”

The journal’s advisory committee highlighted the paper as a “valuable contribution to realistically simulating supported nanoparticles.”

During a summer 2024 CSGF internship with Chris Mundy and Greg Schenter at Pacific Northwest National Laboratory, Maxson extended his MLIP research to include heterogenous interfaces — for example, an aqueous-solid interface. It’s work he’s carried over into Szilvási’s group.

Besides the fellowship, Maxson has received other DOE support, in 2025 and 2026 for solo research time on the Perlmutter supercomputer at the National Energy Research Scientific Computing Center. There, he has used MLIP to simulate how molecules at high concentrations are absorbed on supported nanoparticle surfaces.

Having now worked for extensive periods in two chemical engineering research groups, Maxson says he’s energized to one day lead his own group at a university or national laboratory.

“My dream job is to simply do good, rigorous and meaningful research wherever I find myself.”