Generative Intelligence Explores Chemical Space at Ten Million Catalysts

Abstract

Discovery of catalytic materials requires systematic exploration of vast chemical spaces. However, the scope of exploration that can be achieved using conventional theoretical and experimental methods is very limited. Herein, we present a scalable framework based on Distributed Generative Transformers, which integrates a Transformer-based generative model with a distributed, parallel generation-screening workflow. Leveraging a pretrained model followed by task-specific fine-tuning, the tailored conditional generation strategy achieves >90% validity for target adsorbates such as CH₃, thereby enabling focused exploration of methane conversion catalysts up to ten million candidates, which expands the accessible design space by two orders of magnitude relative to existing generative models. By coupling dimensionality reduction with a machine-learning-potential (MLP) model for performance prediction, we efficiently identifies 26 known active catalysts and more than 1,200 previously unreported candidates. Subsequent subgroup discovery (SGD) analysis reveals synergistic elemental effects that modulate surface electronic structure, tuning CH₃ binding energies into the optimal window for CH₄ activation. This work establishes a generalizable paradigm that seamlessly bridges generative exploration with high-throughput performance mapping, dramatically accelerating catalyst discovery across unprecedented chemical spaces.