Skip to content

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

🕒 Published (v1): 2026-06-17 14:57 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AdsMind is a closed-loop, five-agent framework that couples an LLM planner with machine-learning force field (MLFF) relaxations to autonomously discover stable surface–adsorbate configurations. Unlike open-loop predecessors, it feeds physical relaxation diagnostics back into the planner, enabling error detection and correction. It achieves 100% and 98.8% success rates on two benchmarks while requiring only ~4–5 MLFF relaxations per case.

Problem

Identifying the lowest-energy adsorption configuration on heterogeneous catalyst surfaces is a combinatorial bottleneck: exhaustive DFT is prohibitively expensive (hundreds to thousands of CPU-hours per configuration), and open-loop LLM agents (e.g., Adsorb-Agent) that generate hypotheses in a single pass have no mechanism to detect or correct geometrically or chemically erroneous placements. These open-loop pipelines silently propagate site biases and produce qualitative errors—including incorrect adsorption-energy signs—for molecular adsorbates on complex surfaces.

Method

AdsMind implements a closed-loop multi-agent architecture with five specialized agents:

  1. Planner — proposes an adsorption hypothesis (site type: ontop/bridge/hollow; binding atom symbols; adsorbate binding indices) as JSON, consuming accumulated search history and feedback signals.
  2. Validator — enforces pre-simulation schema constraints (valid heavy-atom indices, site-type compatibility, no-repeat check); failures redirect the Planner without consuming the relaxation budget.
  3. Executor — converts the symbolic hypothesis to atomic coordinates via a surrogate-SMILES geometry layer, samples up to 8 surface sites × 4 conformers, screens with single-point MACE-MP-0 energies, then relaxes the top candidate with BFGS (\(f_\text{max} = 0.10\) eV/Å, 200 steps).
  4. Analyzer — parses the relaxed trajectory into chemically meaningful diagnostics: connectivity via \(d_{ij} \leq \gamma(R_i + R_j)\) with \(\gamma \in \{1.30, 1.35, 1.45\}\), site classification, dissociation flags, and Chemical Slip events (planned vs. relaxed binding mode mismatch).
  5. Summarizer — produces structured Markdown reports including per-iteration tables, Planner reasoning, visualizations, and convergence curves upon run termination.

Three closed-loop mechanisms drive error correction: Chemical Slip detection (flags binding-mode mismatches post-relaxation), FORBID directive (converts slipped/unstable sites into explicit prohibitions for subsequent Planner iterations), and TERMINATE directive (early stopping when the history converges within 0.05 eV). Disabling all three and restricting to one attempt yields the 1-Shot open-loop ablation. The primary MLFF is MACE-MP-0 (small); four LLM backends are tested (Gemini-2.5-Pro, GPT-5.4, Claude-Sonnet-4.6, Grok-4) at temperature 0.

Key Contributions

  • Closed-loop multi-agent harness with physics-grounded feedback replacing open-loop single-pass LLM planning for adsorption configuration search.
  • Three novel control mechanisms (Chemical Slip, FORBID, TERMINATE) that convert MLFF relaxation outcomes into actionable planning constraints.
  • Validator gate that blocks schema-malformed hypotheses before any physical simulation, conserving relaxation budget.
  • DFT (VASP/PBE) validation demonstrating that open-loop baselines produce qualitative adsorption-energy sign errors for molecular adsorbates, while AdsMind preserves sign in all tested cases.
  • OCD-GMAE62 benchmark spanning intermetallic, monometallic, oxide, and compound surfaces across two evaluation tiers (Tier 1: full matrix; Tier 2: stability audit with \(N=3\) repeats on 12 subsampled cases).

Results

  • Success rate (AA20): 100% (Full) vs. lower in 1-Shot ablation (across 4 LLM backends × 20 cases = 400 runs).
  • Success rate (OCD-GMAE62): 98.8% (Full) across 1240 runs (62 cases × 5 configs × 4 backends).
  • Relaxation efficiency: 4.11 relaxations/case (AA20) and 4.67 relaxations/case (OCD-GMAE62) — approximately 14× fewer than heuristic enumeration (median 56 and 61.5 sites/case, respectively).
  • DFT validation (6 AA20 systems): AdsMind (MACE-MP-0, GPT-5.4 Full) MAE vs. PBE = 1.55 eV; Adsorb-Agent (EquiformerV2) MAE = 1.84 eV; Adsorb-Agent exhibits adsorption-energy sign errors on CuPd₃–NNH and OH/Pt(111)/Pt(100); AdsMind preserves correct sign in all 6 cases.
  • Full configuration reduces cross-backend energy dispersion relative to 1-Shot ablation.

Limitations

  • DFT ground-truth validation covers only 6 of 20 AA20 systems; no DFT validation on OCD-GMAE62.
  • Absolute MAE vs. DFT remains large (1.55 eV), inheriting MACE-MP-0 (small) force-field accuracy limits.
  • Benchmarks are relatively small (82 surfaces total); generalization to highly complex multi-component or solvated surfaces is not demonstrated.
  • All LLM backends run at temperature 0, which may reduce diversity of hypotheses and understate stochastic variance.
  • MACE-MP-0 dispersion is disabled in primary benchmarks; sensitivity tests suggest up to ~0.01 eV effects, but dispersion-critical physisorption cases are not evaluated.
  • The ~14× relaxation reduction is relative to heuristic enumeration, not DFT-cost-equivalent methods.

Relevance to Harnesses / Meta-Harnesses

AdsMind is a concrete example of a domain-specific multi-agent harness with a well-articulated closed-loop control architecture: the five agents (Planner, Validator, Executor, Analyzer, Summarizer) map directly onto the plan–validate–execute–analyze–report pipeline pattern common in scientific automation harnesses. The three feedback mechanisms (Chemical Slip, FORBID, TERMINATE) are harness-level control-flow primitives—they govern iteration, early stopping, and constraint propagation without LLM hallucination, demonstrating how physics-grounded signals can replace purely model-driven self-reflection. The Validator pre-simulation gate and the Summarizer's structured Markdown reporting are archetypal harness components (input validation, standardized output), and the ablation study systematically isolates each mechanism's contribution—a methodology directly applicable to evaluating analogous modules in general-purpose meta-harnesses.