Interpretable Inverse Design of Metal-Organic Frameworks with Large Language Model Agents

A research team has introduced LLM4MOF, a framework that deploys large language model agents to perform interpretable inverse design of metal-organic frameworks, proposing and testing chemistry hypotheses in a closed-loop simulation environment [1]. The system, detailed in a preprint posted to the arXiv repository on June 28, 2026, uses two cooperating language-model agents [1]. One agent proposes design hypotheses covering metal nodes, organic linkers, pore geometry, and functional chemistry. A second agent translates those hypotheses into constraints that select candidate metal-organic frameworks, or MOFs, each built from a metal node, an organic linker, and a matching topology [1]. arXiv, which hosts the paper, is an open-access repository of electronic preprints that has grown to a submission rate of about 24,000 articles per month as of late 2024 [10]. Each hypothesis is tested through four diagnostic beams that apply different subsets of its constraints, allowing researchers to compare whether geometry, chemistry, or metal choice drives performance [1]. The loop operates autonomously over ten iterations, refining its hypotheses based on simulation results [1]. Across six adsorption, separation, and electronic-structure tasks, LLM4MOF concentrated its search on top-performing structures within 400 property evaluations, even without access to the global property landscape of existing databases [1]. The framework also generates new MOFs de novo and validates them in live simulation, adapting geometry to each requested condition. The authors report that it outperformed both random search and a genetic algorithm at a cost of roughly $1 per campaign [1]. The work demonstrates that language-model agents can conduct simulation-grounded inverse design without training a separate machine-learning model for each objective, a departure from conventional approaches that often require expensive property labels and yield opaque models [1]. The preprint appears within arXiv’s machine learning category and is accompanied by the arXivLabs framework, a community-innovation space that allows collaborators to develop and share experimental tools directly on the article record page [1][9]. arXivLabs, launched in 2020, sets guidelines ensuring that partners share arXiv’s values of openness, community, excellence, and user data privacy [9].