Materials

uMOF: A Universal Database, Benchmark, and Machine Learning Interatomic Potentials for Metal-Organic Frameworks

Théo Jaffrelot Inizan, Prathami Divakar Kamath, Alin Marin Elena, Kristin A. Persson

Featured September 9, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

By creating a huge, high-quality dataset and using smart AI to find real-world data, the authors built new computer models that are much better at predicting how spongy materials called MOFs capture gases, even outperforming models trained on much more data.

In depth

The paper introduces uMOF, a comprehensive resource for Metal-Organic Frameworks (MOFs) comprising the largest DFT dataset at the accurate r

\textsuperscript 2

SCAN-D4 level, a novel LLM-mined experimental benchmark, and two fine-tuned machine learning interatomic potentials (MLIPs). These uMOF models demonstrate superior accuracy in predicting complex, dynamics-sensitive properties like gas adsorption enthalpies, significantly outperforming existing baselines by leveraging high-fidelity training data and diverse sampling strategies, including finite-temperature molecular dynamics.

Key Takeaways

1
The authors release the uMOF dataset, the largest and most accurate density functional theory (DFT) dataset for MOFs, computed at the r $\textsuperscript 2$ SCAN-D4 level across 85,524 configurations and 79 elements, crucially including finite-temperature molecular dynamics (MD) simulations.
2
A novel LLM-mined experimental benchmark for MOFs is introduced, comprising 3,986 verified property values (3,146 experimental) extracted from 626 papers via a multi-stage pipeline with deterministic hallucination filtering.
3
Two universal MLIPs, uMOF-MH and uMOF-POLAR, fine-tuned on the uMOF dataset, achieve over 80% error reduction on gas adsorption enthalpies compared to specialized baselines, demonstrating that high-fidelity data and diverse sampling (especially MD) are more critical than raw data volume for complex MOF properties.

Conceptual Flow

HIGH LEVEL

Methodology (The Logic)

The authors built a super-detailed digital library of materials, then used smart computer programs to find real-world test results, and finally trained new AI models to predict how these materials behave.

Material Recipes

Real-World Papers

Create & Collect

Digital Material Data

Verified Test Results

Results (The Impact)

Their new AI models, trained on this special data, are much better at predicting how materials capture gases, especially for tricky situations, beating older models by a lot.

Old Prediction Models

New AI Models

Compare Performance

Better Gas Capture Predictions

More Accurate Material Design

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.