Work

Physics-Informed Analytical Models for Interpretable and Deployable Hydrogen Storage Prediction in MOFs

Year: 2025

Type: article

Abstract: Storing hydrogen safely and compactly is a central challenge for clean transportation, grid support, and industrial decarbonization. Metal-organic frameworks (MOFs) are compelling candidates because of their ultrahigh internal surface areas and large pore volumes, yet the design space is effectively infinite, making exhaustive synthesis and simulation impractical. We introduce a physics-informed analytical modeling approach that uses deterministic sparsity-promoting symbolic regression with explicit physical constraints to produce a family of closed-form equations for fast, interpretable, and deployable prediction of usable hydrogen storage in MOFs. On a curated dataset of 88 400 MOFs, the method discovers compact, physically consistent structure-property relationships using as few as one property (void fraction for volumetric capacity) or two properties (void fraction and pore volume for gravimetric capacity) selected from seven descriptors: single-crystal density, gravimetric surface area, volumetric surface area, void fraction, pore volume, largest cavity diameter, and pore-limiting diameter. Equations using one to three properties achieve design-grade accuracy ( R 2 > 0.99 for gravimetric capacity and R 2 > 0.95 for volumetric capacity), while additional descriptors provide marginal gains for high-capacity edge cases, matching black-box machine learning performance and providing mechanistic insights into why some MOFs excel and others do not. Because evaluation is simulation-free and effectively costless, the models integrate directly into high-throughput screening, active learning workflows, and generative design to prioritize syntheses and avoid dead ends, establishing a generalizable path to interpretable and deployable surrogates for adsorption-based energy materials.

Source: PRX Energy

DOI: 10.1103/glnw-drn4

Author: Alauddin Ahmed

Institution: University of Michigan

Language: English

Topic: Machine Learning in Materials Science

Subfield: Materials Chemistry

Field: Materials Science

Domain: Physical Sciences

SDG: Affordable and clean energy

Open Access status: diamond

Funders: U.S. Department of Energy, University of Michigan, Office of Energy Efficiency and Renewable Energy, Sustainability Institute, Ohio State University