Inverse Design of Catalytic Active Sites via Interpretable Topology-Based Deep Generative Models

This work present a persistent GLMY homology-based variational autoencoder framework (PGH-VAEs) designed to enable interpretable inverse design of catalytic active sites. PGH-VAEs integrate advanced topological algebraic analysis to mathematically quantify the three-dimensional structural sensitivity of active sites and establish intrinsic correlations with their adsorption properties. Using high-entropy alloys as a complex test case, PGH-VAEs demonstrate how the multi-channel encoding framework captures coordination and ligand effects at active sites, shaping the latent space of the generative model and significantly influencing the adsorption energies of key species. Building on the inverse design outcomes of PGH-VAEs, we propose strategies to optimize the composition and facet structures of high-entropy alloys to maximize the proportion of ideal active sites. This framework seamlessly combines catalyst design with topological analysis, offering a novel pathway for machine intelligence-driven development of efficient catalysts across diverse systems.

Keywords: Persistent GLMY homology, Deep learning Inverse design, Interpretability, Heterogeneous catalysis.

The data for this work comes from the article: Neural network-assisted development of high-554 entropy alloy catalysts: Decoupling ligand and coordination effects, please introduce the data used in this article.