MASTER

Materials Agents for Simulation and Theory in Electronic-structure Reasoning: an active-learning framework in which LLM agents autonomously design, execute, and interpret density functional theory calculations to explore catalytic chemical space.

   
Affiliation Los Alamos National Laboratory, Theoretical Division, with the University of Connecticut (paper)
First introduced 2025-12 (arXiv:2512.13930, dated 2025-12-15)
Lifecycle stages Multi-stage (hypothesis about where to look → simulation design and execution → interpretation feeding the next choice)
Autonomy level Semi-autonomous — humans set the chemical target and success criteria; agents choose, construct, repair, and interpret each calculation
Domain focus Surface chemistry and electrocatalysis: CO adsorption on transition-metal adatoms on Cu(100), and on M–N–C single-atom catalysts
Availability Code on request from the corresponding authors; supporting data in the Supplementary Information

Approach

MASTER splits the problem into a simulation layer and a reasoning layer. The simulation layer is a multimodal system that translates natural-language specifications (“build a Cu(100) slab 4×4 with 6 layers and 15 Å vacuum; place an Ag adatom at the fourfold hollow site; place a CO molecule C-down bonded atop the adatom; fix the bottom layers”) into concrete DFT input geometries and workflows. A self-revision loop inspects the generated structure, diagnoses errors, and retries — addressing the standard bottleneck in which failed calculations require human intervention to repair.

The reasoning layer sits above it and decides which calculation to run next. Four strategies are compared: a single-agent baseline and three multi-agent designs — peer review (agents critique each other’s proposals), triage-ranking (candidates are ranked and the top ones advanced), and triage-forms (structured elicitation of each agent’s assessment). The framing is explicitly active learning: reasoning replaces trial-and-error enumeration over a chemical space of millions of configurations, and the paper’s central claim is that the resulting trajectories are chemically motivated rather than an artifact of stochastic sampling or semantic similarity.

Validation

Entirely in-silico. Two chemical applications — CO adsorption on transition-metal adatoms supported on Cu(100), and CO adsorption on M–N–C catalysts — with the geometry-generation subsystem benchmarked separately against subject-expert inspection across 18 representative transition-metal adatom configurations. Reasoning trajectories were audited against stochastic-sampling and semantic-bias null models.

Notable results

  • Reasoning-driven exploration reduced the number of atomistic simulations required to reach the chemical target by up to 90% relative to trial-and-error selection.
  • The self-correcting simulation agents reached a 97.8% success rate on adatom-only geometry construction, with most failures resolved on the first or second retry.
  • Recorded reasoning traces showed chemically grounded decisions — e.g. correcting a CO molecule placed oxygen-down on an osmium adatom — that could not be reproduced by stochastic sampling or semantic-similarity baselines.

Primary paper

Rothfarb, Davis, Holby, Matanovic, Li, Kort-Kamp, “Hierarchical Multi-agent Large Language Model Reasoning for Autonomous Functional Materials Discovery,” arXiv:2512.13930.

Other references

None yet.

Code

Not publicly released — available from the corresponding authors upon reasonable request.