Data-Driven Generation of Zeolites: Preliminary Results on Coordination Control

Porous materials are central to thermodynamic processes such as gas separation and gas-stream dehydration, where performance depends on adsorption energetics and transport properties. Large experimental structure databases enable data-driven exploration of next-generation porous materials. Diffusion models, which generate structures by reversing a noise-adding process, offer a promising route for this exploration. However, current approaches applied to Si–O zeolites still struggle to capture realistic coordination environments.

In this study, we present preliminary results from the generation of 2000 Si–O zeolite structures conditioned on target volume fractions. The volume fractions of the generated samples follow the target values with a relative standard deviation of about 10%, showing that the model captures global geometric constraints. We also measured distributions of energetic descriptors, namely the Henry coefficient and the isotropic heat of CH₄ adsorption, which reflect those of the training database.
Structural features such as Si–O ratios, coordination numbers, and framework connectivity were analyzed. Although these properties are peaked around expected ideal values, their broad distributions include many poorly connected or non-physical structures, highlighting current limitations.

To address these limitations, we propose methodological advances including dynamic control of local coordination during diffusion and the use of an autoencoder to reduce the diffusion input space and improve training efficiency. This workflow will be coupled with high-throughput screening to establish structure–property relationships, including adsorption uptake and diffusion coefficients, and with molecular dynamics to relax generated structures and assess their mechanical and chemical stability. We also discuss challenges in extending this approach to more complex materials and other adsorbates such as CO₂.