MOFFieldBench: A Machine Learning Benchmark for Continuous Electronic Fields in Metal-Organic Frameworks

Abstract

Machine-learning models for metal-organic frameworks (MOFs) predominantly rely on atomic structures, whereas electron-density (ED) and electrostatic-potential (ESP) fields remain underexplored as scalable learning representations. These continuous fields encode complementary spatial and electronic information relevant to structure-property relationships, particularly pore geometry and guest-framework interactions, but costly first-principles calculations, storage requirements and the scarcity of systematically aligned datasets limit their broader use. Here we introduce MOFFieldBench, a multimodal benchmark comprising approximately 46,000 MOFs with one-to-one aligned crystallographic structures, ED fields and ESP fields. A unified workflow integrates field generation, error-controlled compression and sparse point-cloud representation, supporting field reconstruction, prediction of 13 material properties, multimodal property prediction and pairwise cross-modal retrieval. ED and ESP provide competitive representations for pore geometry, adsorption and separation properties. Multimodal fusion improves several prediction tasks relative to the best unimodal reference, supporting complementary predictive roles for atomic structures and electronic fields. Learned ED reconstruction closely approximates first-principles reference densities, while predicted-density initialization reduces self-consistent-field calculation time by 13.59\% on average in paired tests. These results support electronic fields as practical, informative learning modalities and establish MOFFieldBench as a resource for studying representation choice, multimodal learning and electronic-field-driven MOF modelling.