Researchers at Tohoku University's Advanced Institute for Materials Research developed an AI agent called MLP Copilot to streamline the creation of atomic models in materials science. The framework learns from quantum calculations to predict atomic energies and forces, enabling larger simulations without requiring every calculation to run from first principles.
Hao Li, distinguished professor at WPI-AIMR, described the tool as similar to autopilot for pilots, designed to make materials science more efficient while keeping humans in control of decisions.
The team tested MLP Copilot on aluminum nanoparticle oxidation. Over approximately 80 active-learning iterations, they assembled 88,833 atomic configurations. Uncertain configurations prompted additional quantum calculations and model retraining. Researchers validated the models against reference results, thermodynamic properties, and independent cluster tests before using them in oxidation simulations.
The simulations revealed two behaviors in particles with an oxide shell. At moderate temperatures, temporary nanoscale pathways allowed oxidation to proceed while limiting material movement, a phenomenon the researchers called a "breathing mode." At higher temperatures, the shell ruptured and exposed the aluminum core to oxygen, accelerating oxidation. Outward aluminum transport emerged as the dominant pathway under the studied conditions.
Throughout the workflow, AI agents organized project knowledge and retrieved evidence, but researchers decided which models to accept and how to interpret results.
The work, published in the Proceedings of the National Academy of Sciences on Oct. 15, 2026, reflects a shift toward human-agent collaboration in research. Future workflows could support recursive self-improvement, in which feedback refines both system capabilities and planning abilities, with each change assessed by researchers.