Skip to content

Large language model agents accelerate inverse design of metal-organic frameworks for gas separation

🕒 Published (v1): 2026-07-12 04:18 UTC · Source: Arxiv · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

LEMO Agent is a closed-loop LLM agent framework for inverse design of metal–organic frameworks (MOFs) in MOFid string space, coupling language-based candidate generation with grammar/chemical validity checking, Transformer-based surrogate property prediction, structured memory, and multi-island exploration. It outperforms generative, optimization, and vanilla agent baselines on CH4/N2 and CO2/N2 separation tasks in both predicted selectivity and structural diversity. Selected candidates proceed through GCMC simulation, experimental down-selection, wet-lab synthesis, and SEM characterization.

Problem

High-throughput computational screening of MOFs is limited to pre-enumerated databases and does not solve the inverse design question of generating new, chemically valid, task-optimized MOFs beyond the initial candidate pool. Conventional deep generative models (RL, diffusion, flow matching) explore learned distributions but lack iterative feedback, explicit chemical-constraint enforcement, and memory of failure modes. Vanilla LLMs can produce MOFid-like strings but do not enforce grammar, verify chemical validity, incorporate property feedback, or accumulate design knowledge across iterations.

Method

LEMO Agent operates on MOFid strings—compact symbolic encodings of organic linkers, metal nodes, and topology labels—as the design space. Each iteration executes a generate–validate–evaluate–remember cycle:

  1. Planner analyzes the accumulated design context (target objective, MOFid grammar rules, prior successes and failures) and formulates a next-round search strategy.
  2. Generator (LLM) proposes new MOFid candidates under those constraints.
  3. Action module runs MOFid grammar checking and RDKit validity checking, then queries Transformer-based surrogate property predictors for CH4/N2 selectivity, CH4 uptake, CO2/N2 selectivity, and CO2 uptake (trained on MOFid–property pairs from existing databases).
  4. Memory module partitions feedback into three stores: knowledge memory (grammar rules, domain knowledge, objectives), success memory (valid high-performing candidates and structural motifs), and failure memory (invalid or low-performing cases).

A multi-island strategy runs multiple independent agent instances, each with its own local strategy and candidate pool, exploring different regions of MOFid space in parallel. Island selection probabilities are computed from accumulated island scores to balance exploitation and exploration. After optimization, selected candidates are reconstructed from MOFid, relaxed, and evaluated by GCMC simulation; an experimental down-selection workflow applies property screening, chemical feasibility assessment, PubChem/MolPort ligand purchasability checks, and GPT-assisted synthesis planning.

Key Contributions

  • Closed-loop LLM agent framework (LEMO Agent) for MOF inverse design directly in MOFid symbolic space.
  • Structured three-component memory (knowledge / success / failure) enabling agents to learn from both successful and failed design trajectories across iterations.
  • Multi-island exploration mechanism that maintains performance–diversity balance by organizing search into multiple complementary trajectories.
  • Transformer-based MOFid surrogate property predictors (\(R^2 = 0.94, 0.93, 0.84, 0.88\) for CH4/N2 selectivity, CH4 uptake, CO2/N2 selectivity, CO2 uptake respectively) providing fast per-iteration feedback without full GCMC simulation.
  • End-to-end validation pipeline from symbolic generation through GCMC simulation, experimental down-selection, wet-lab synthesis, and SEM characterization.

Results

  • Transformer MOFid predictors on held-out test sets: CH4/N2 selectivity \(R^2 = 0.94\), MAE \(= 0.26\); CH4 uptake \(R^2 = 0.93\), MAE \(= 0.07\); CO2/N2 selectivity \(R^2 = 0.84\), MAE \(= 7.78\); CO2 uptake \(R^2 = 0.88\), MAE \(= 0.15\).
  • Over 50 iterations on CH4/N2, LEMO Agent continuously improves best-so-far selectivity and substantially outperforms the vanilla agent, which saturates quickly at a lower level.
  • Candidate validity rate for CH4/N2 increases from 9.5% (iteration 0) to 30.2%, 52.3%, and 61.6% at iterations 10, 25, and 50.
  • LEMO Agent achieves highest selectivity and working capacity among all baselines (vanilla agent, LLM-only, MOFGPT) on both CH4/N2 and CO2/N2 tasks while maintaining structural diversity comparable to MOFGPT.
  • MOFGPT achieves high diversity but substantially lower selectivity and working capacity than LEMO Agent.
  • Ablation (LEMO w/o MI, removing multi-island): both property performance and diversity decrease, confirming the multi-island mechanism is necessary for the performance–diversity balance.
  • Selected candidates reconstructed from MOFid, evaluated by GCMC, and subjected to wet-lab synthesis and SEM characterization as initial experimental validation.

Limitations

  • Property evaluation during optimization relies on Transformer surrogates, not ground-truth GCMC; surrogate errors propagate to agent decisions.
  • Physical structural descriptors (PLD, framework density) are estimated from ML surrogates rather than geometry-optimized CIF structures.
  • Validity rate at convergence (~62%) means ~38% of generated candidates per iteration are still chemically invalid, representing wasted generation budget.
  • Wet-lab validation is preliminary (initial synthesis and SEM only); no adsorption performance data from experiments are reported.
  • The MOFid symbolic representation may not fully capture structural nuance (e.g., framework flexibility, interpenetration), potentially limiting the agent's reasoning about geometry-sensitive separation properties.
  • Paper text is truncated before the physical reasonableness and diversity analysis results are fully reported.

Relevance to Harnesses / Meta-Harnesses

LEMO Agent is a domain-specific meta-harness that orchestrates a heterogeneous tool stack—LLM generator, grammar checker, cheminformatics validator, surrogate predictors, and persistent memory—within a deterministic iterative pipeline, making it a concrete instantiation of the closed-loop agent harness pattern. The multi-island mechanism is itself a meta-level coordination layer that routes generation across parallel sub-agents and aggregates their outcomes, directly paralleling multi-agent orchestration strategies studied in general harness research. The structured three-component memory (knowledge/success/failure) is a harness-managed shared state store that persists information across agent turns, demonstrating how harness-level state architecture shapes agent behavior and convergence. This work extends harness design from software/task domains into scientific discovery, showing that harness choices—memory partitioning, island selection policy, validation gating—are primary drivers of end-task performance rather than incidental scaffolding.