AdsMind: A Physics-Grounded Multi-Agent System for Self-Correcting Discovery of Adsorption Configurations on Heterogeneous Catalyst Surfaces

2026-06-17 · Source: Artificial Intelligence · Field: Science & Research — Physical Sciences & Chemistry, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

AdsMind, a physics-grounded multi-agent system, significantly improves the discovery of lowest-energy surface-adsorbate configurations crucial for heterogeneous catalysis modeling. It addresses the computational burden of ab initio calculations and the lack of physics-grounded feedback in open-loop large language model (LLM) agents. AdsMind employs a closed-loop framework with machine-learning force field (MLFF) relaxation feedback, enabling autonomous error correction. Across four LLM backends, it achieved 100% success on the AA20 benchmark and 98.8% on OCD-GMAE62, while using only 4.11 and 4.67 MLFF relaxations per case, respectively. This represents an approximately 14-fold reduction compared to heuristic enumeration baselines. Density functional theory (DFT) validation confirmed AdsMind's ability to correct qualitative adsorption-energy sign errors found in open-loop Adsorb-Agent outputs, preserving correct signs and showing closer quantitative agreement.

Key takeaway

For research scientists modeling heterogeneous catalysis, AdsMind offers a robust approach to overcome computational bottlenecks in adsorption configuration discovery. You should consider integrating closed-loop MLFF relaxation feedback into your LLM-driven workflows to achieve higher reliability and significantly reduce ab initio calculation costs. This method corrects qualitative errors and accelerates the identification of low-energy configurations, enabling more efficient and accurate autonomous chemistry.

Key insights

AdsMind uses MLFF feedback in a multi-agent system for self-correcting, efficient discovery of catalytic adsorption configurations.

Principles

• Closed-loop feedback enhances LLM agent reliability.
• Physics-grounded correction improves accuracy.
• Reduced relaxations accelerate configurational search.

Method

AdsMind employs a closed-loop multi-agent system that integrates LLM-generated initial guesses with machine-learning force field (MLFF) relaxation feedback for autonomous error correction and iterative refinement of adsorption configurations.

In practice

• Apply MLFF feedback to refine LLM-generated structures.
• Integrate multi-agent systems for complex chemical searches.
• Validate LLM outputs with DFT for critical applications.

Topics

• Heterogeneous Catalysis
• Adsorption Configurations
• Multi-Agent Systems
• Machine Learning Force Fields
• Large Language Models
• Density Functional Theory

Best for: AI Scientist, Research Scientist

Related on AIssential

• AdsMind: A Physics-Grounded Multi-Agent System for Self-Correcting Discovery of Adsorption Configurations on Heterogeneous Catalyst Surfaces — AdsMind is a physics-grounded multi-agent system using MLFF feedback for self-correction in heterogeneous catalysis configuration discovery.
• AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations — AutoDFT automates complex DFT calculations via a closed-loop multi-agent LLM framework, adapting dynamically to emergent results.
• MIND: AI Co-Scientist for Material Research — MIND is an LLM-driven multi-agent framework for automated, experimentally verified hypothesis validation in materials science.
• Exascale Multi-Task Graph Foundation Models for Imbalanced, Multi-Fidelity Atomistic Data — Exascale multi-task graph foundation models accelerate materials discovery by enabling billion-scale screening and data-efficient fine-tuning.
• LEAP: A closed-loop framework for perovskite precursor additive discovery — Coupling a domain-specialized LLM with active learning and expert review accelerates perovskite additive discovery under low-data conditions.
• Battery-Sim-Agent: Leveraging LLM-Agent for Inverse Battery Parameter Estimation — LLM agents can reframe inverse scientific problems as reasoning tasks, outperforming traditional black-box optimization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.