Skip to content

SimMOF: AI agent for Automated MOF Simulations

🕒 Published (v1): 2026-03-31 02:08 UTC · Source: Arxiv · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

SimMOF is an LLM-based multi-agent framework that automates end-to-end MOF (metal-organic framework) computational simulation workflows from natural language queries. It orchestrates four simulation engines (Zeo++, RASPA, LAMMPS, VASP) through modular agents handling planning, input construction, execution, error recovery, and result interpretation. The system mirrors expert-researcher decision-making, including dependency-aware parallelism and closed-loop self-debugging.

Problem

MOF computational simulations require expert knowledge across workflow construction, parameter selection, force-field assignment, structure preparation, tool interoperability, and error diagnosis — collectively creating a high barrier for non-specialists. Existing LLM-based automation efforts address only isolated steps (input generation or single-tool execution) and target narrow task types (e.g., CO₂ adsorption only), leaving no flexible, general-purpose automation for complete end-to-end MOF research workflows.

Method

SimMOF decomposes the simulation pipeline into three functional modules:

  1. Task Management Module — a query agent parses natural-language intent and identifies the target system; a working agent constructs dependency-aware execution plans with explicit parallelization constraints (independent jobs run concurrently, dependent jobs are sequenced); a response agent translates outputs back to natural language.

  2. Simulation Module — four specialist agents wrap Zeo++ (geometry/topology), RASPA (GCMC adsorption), LAMMPS (MD: diffusion, mechanics), and VASP (DFT: electronic structure, binding energies). The working agent orchestrates invocation order at runtime.

  3. Assistance Module — a RAG agent retrieves setup parameters and protocols from a MOF literature database; a structure agent resolves material identifiers to CoRE MOF/CSD/PubChem entries; an error agent performs both pre-execution input validation and post-execution log inspection with iterative corrective feedback (closed-loop recovery). An optional screening agent applies hierarchical low-cost filters (geometric descriptors, MLIP-based Henry constants) before expensive simulations.

For high-throughput screening, the framework applies sequential filters to shrink candidate sets (e.g., 3,786 → 1,878 → 1,000 via structure validation, atom-count limits, geometric accessibility, then Henry constant ranking) before invoking parallel RASPA evaluations.

Key Contributions

  • End-to-end natural-language-to-simulation pipeline covering geometry, adsorption, MD, and DFT analyses for MOFs.
  • Dependency-aware parallel job scheduling encoded explicitly in working plans, enabling automatic sequencing (e.g., geometry optimization → complex optimization → binding energy).
  • Closed-loop error recovery: pre-execution validation and post-execution log inspection with automatic input regeneration (demonstrated by detecting CH₄→COâ‚‚ force-field mismatch and adding Coulombic terms).
  • RAG-grounded parameter selection consulting MOF literature databases to initialize force fields, protocols, and screening heuristics.
  • Hierarchical screening agent that applies low-cost descriptor cascades before expensive GCMC/DFT, preserving top-10 candidates exactly versus exhaustive evaluation.
  • Multi-tool integration across four simulation codes with a unified agent interface.

Results

Workflow generation accuracy (Table 1): - Zeo++: 5/5 tasks (100%) - RASPA: 4/4 tasks (100%) - LAMMPS: 4/5 tasks (80%; failure: Young's modulus, requires structural deformation preprocessing) - VASP: 4/5 tasks (80%; failure: nudged elastic band, requires intermediate image construction)

Quantitative accuracy vs. reference values (Table 2): - Zeo++ geometric descriptors (LCD, PLD): 0.00% error on all tested MOFs - RASPA adsorption (N₂, CH₄, H₂ uptake): errors of 0.01–6.11% - LAMMPS diffusivity (O₂): 1.06–2.60% error - VASP band gaps: 0.07–7.50% error - VASP binding energies: 0.07–17.85% error (worst case H₂O binding energy, absolute deviation 0.05 eV, within DFT typical error)

High-throughput screening: Top 10 CHâ‚„ uptake candidates from screened subset of 1,000 (out of 3,786) were identical to top 10 from exhaustive RASPA evaluation of the full database.

Limitations

  • Young's modulus (LAMMPS) and nudged elastic band (VASP) workflows unsupported due to missing structure preprocessing (deformation routines, NEB image construction).
  • No quantitative evaluation of multi-step case studies (parallel VASP, LAMMPS with error correction) beyond qualitative demonstration.
  • Relies on external structure databases (CoRE MOF, CSD) whose coverage constrains accessible materials.
  • RAG quality is bounded by the indexed MOF literature corpus; novel materials or unusual protocols may lack retrieval support.
  • VASP binding energy errors up to ~17% for some systems suggest sensitivity to initial configuration sampling that is not yet fully resolved.
  • Parallelization is limited to within-plan job-level concurrency; cross-plan or cross-user scheduling is not addressed.

Relevance to Harnesses / Meta-Harnesses

SimMOF is a domain-specific meta-harness: it demonstrates the pattern of a coordinator agent decomposing a high-level intent into a dependency graph of tool-agent invocations, then monitoring, error-recovering, and synthesizing results — the canonical harness loop applied to computational chemistry. Its three-module architecture (task management / simulation / assistance) is a concrete instantiation of the plan→dispatch→verify→report harness pattern, where the working agent acts as the meta-orchestrator. The closed-loop error recovery mechanism — where an error agent iteratively corrects and re-invokes simulation agents — is a practical implementation of self-healing harness execution that could generalize to other domain-specific orchestration contexts. The hierarchical screening agent illustrates how a harness can insert adaptive filtering stages to manage computational cost before expensive sub-agents are invoked.