Technical Summary: AdsMind

Problem Statement

Identifying the lowest-energy surface–adsorbate configuration is a fundamental bottleneck in computational heterogeneous catalysis. The configuration space grows combinatorially with site types (ontop, bridge, hollow), adsorbate orientations, and conformational flexibility, often yielding hundreds to thousands of candidates for a single system. While Density Functional Theory (DFT) provides high accuracy, its computational cost (hundreds to thousands of CPU-hours per relaxation) renders exhaustive exploration prohibitive.

Machine Learning Force Fields (MLFFs) offer a speedup of orders of magnitude but do not solve the upstream search strategy problem: deciding which configurations to evaluate. Existing approaches fall into two categories with distinct limitations:

1. Open-loop Deep Learning: Methods like graph neural networks or generative models propose candidates in a single pass without physical feedback. If an initial proposal misses a stable configuration or is geometrically inconsistent, the pipeline cannot recover.
2. Stochastic Search: Genetic algorithms or basin hopping reduce evaluations but still require significant expert oversight and scale poorly across composition spaces.

Furthermore, recent Large Language Model (LLM) agents, such as Adsorb-Agent, have shown promise in generating adsorption hypotheses but operate as open-loop systems. They lack a physics-grounded feedback mechanism to correct erroneous initial guesses, leading to "silent" failures where the realized physical structure differs from the planned intent (e.g., site slips or dissociation) without the agent being informed.

Methodology: The AdsMind Framework

The authors propose AdsMind (Adsorption configuration discovery with Machine intelligence and relaxation feedback), a closed-loop multi-agent system designed to enable autonomous error correction through MLFF relaxation feedback. The framework integrates an LLM planner with iterative structural relaxation using the MACE-MP-0 force field.

Core Architecture

AdsMind consists of five specialized agents:

1. Planner: Generates structured adsorption hypotheses (JSON format) specifying site types, binding atoms, and indices. Crucially, its prompt includes accumulated search history and feedback signals from previous iterations.
2. Validator: Enforces schema constraints before simulation (e.g., valid binding indices, geometric compatibility with site types, and deduplication) to prevent wasting resources on malformed hypotheses.
3. Executor: Converts symbolic hypotheses into concrete atomic structures, standardizes the surface slab, and performs structural relaxation using the BFGS algorithm with the MACE-MP-0 force field.
4. Analyzer: Parses relaxation trajectories to extract chemically meaningful diagnostics, including adsorption energies, bond integrity, and Chemical Slip detection (mismatches between planned and relaxed binding sites).
5. Summarizer: Handles terminal reporting and early-stopping logic, generating structured Markdown reports with visualizations.

Closed-Loop Mechanisms

The system relies on three feedback mechanisms to drive iterative correction:

• Chemical Slip Detection: Identifies when the relaxed binding mode or site type differs from the Planner's initial intent.
• FORBID Directive: Converts slipped or unstable site intents into explicit prohibitions for subsequent iterations, preventing the agent from repeating known failures.
• TERMINATE Directive: Uses convergence signals (e.g., energy degeneracy within 0.05 eV of the best result) to enable early stopping, optimizing the relaxation budget.

Experimental Setup

The framework was evaluated on two benchmark datasets:

• AA20: 20 surface–adsorbate cases involving intermetallic and monometallic surfaces.
• OCD-GMAE62: 62 cases spanning intermetallics, monometallics, and chemically diverse compound surfaces (chalcogenides, alloys).

Four LLM backends were tested: Gemini-2.5-Pro, GPT-5.4, Claude-Sonnet-4.6, and Grok-4. The primary MLFF calculator was MACE-MP-0 (small). The study included ablation variants (1-Shot, w/o Term, w/o Slip, w/o Forbid) and comparisons against a heuristic enumeration baseline and the open-loop Adsorb-Agent. DFT validation was performed on six representative AA20 systems using VASP/PBE.

Key Results

1. DFT Validation and Physical Fidelity

On six representative AA20 systems, AdsMind demonstrated superior qualitative agreement with DFT compared to the open-loop Adsorb-Agent:

• Sign Correctness: AdsMind preserved the correct exothermic sign for adsorption energies in all tested cases. In contrast, the reported Adsorb-Agent outputs exhibited qualitative sign errors (predicting endothermic binding for exothermic systems) for molecular adsorbates (OH and NNH) on Pt and CuPd3 surfaces.
• Quantitative Agreement: AdsMind achieved a Mean Absolute Error (MAE) of 1.55 eV against DFT, compared to 1.84 eV for Adsorb-Agent, with no sign mismatches observed.

2. Search Reliability and Efficiency

• Success Rates: AdsMind (Full configuration) achieved a 100% success rate on AA20 and 98.8% on OCD-GMAE62. The open-loop 1-Shot baseline dropped to 89.5% on OCD-GMAE62.
• Relaxation Budget: AdsMind drastically reduced the number of MLFF relaxations required per case.
  • AA20: 4.11 relaxations per case (vs. median 56 for heuristic enumeration).
  • OCD-GMAE62: 4.67 relaxations per case (vs. median 61.5 for heuristic enumeration).
  • This represents an approximately 14-fold reduction in computational cost compared to heuristic baselines while maintaining higher reliability.

3. Ablation and Convergence Analysis

• Feedback Necessity: The 1-Shot (open-loop) variant showed significantly higher energy deviations and lower success rates. The "Chemical Slip" analysis revealed that 61.5–76.9% of valid 1-Shot attempts on intermetallic surfaces relaxed to a different site type than intended, highlighting the failure mode of open-loop agents.
• Cross-Backend Agreement: The closed-loop Full configuration reduced cross-backend energy dispersion. On AA20, the mean energy range across four LLMs dropped from 0.473 eV (1-Shot) to 0.195 eV (Full).
• Stability: Tier 2 stability audits showed that 72.9% of Full-configuration runs returned selected energies with a run-to-run range below 0.001 eV, indicating that feedback constraints effectively dampen LLM stochasticity.

4. Trade-offs

The study acknowledges a reliability–exploration trade-off. While AdsMind prioritizes reliable convergence and budget control, the FORBID directive can occasionally over-constrain the search, causing the system to miss lower-energy local basins in rare edge cases (e.g., one instance where 1-Shot found a structure 0.99 eV lower than Full). However, the authors note this is a backend-specific edge case rather than a systematic failure.

Significance and Claims

The paper positions AdsMind as a foundational step toward interpretable and self-correcting autonomous scientific discovery in computational catalysis. Its primary contributions are:

1. Closing the Loop: It demonstrates that feeding physical relaxation outcomes back into an LLM planner allows the system to detect, diagnose, and recover from its own reasoning errors (e.g., geometric slips), a capability absent in open-loop agents.
2. Reliability over Breadth: AdsMind establishes a "reliability-first" operating envelope. It sacrifices the exhaustive breadth of heuristic enumeration to achieve chemically audited convergence with a fraction of the computational cost.
3. Interpretability: By quantifying "Chemical Slip" and generating diagnostic artifacts (FORBID constraints, bond-integrity reports), the system makes the search process inspectable, bridging the gap between symbolic LLM planning and physical reality.
4. Robustness: The framework delivers consistent performance across diverse LLM backends and surface chemistries, reducing sensitivity to the specific model used and providing a stable basis for downstream comparisons.

The authors conclude that while AdsMind does not guarantee global minima in all complex cases, it provides a practical, physics-grounded workflow that supports more informed autonomous chemistry by ensuring that proposed configurations are not just statistically probable, but physically consistent and self-verified.