Catalyst Diffusion Transformer: Generative Inverse Design of Heterogeneous Catalysts

Traditional catalyst discovery relies on screening predefined libraries, which limits the exploration of the vast chemical space of potential materials. This study addresses the need for a unified, scalable generative framework capable of inverse design—generating new catalyst structures that satisfy specific functional constraints, such as binding energy and composition, rather than simply searching through existing databases.

CatDiT employs a two-stage latent diffusion architecture. First, a Variational Autoencoder (VAE) with an SE(3)-equivariant encoder (EquiformerV2) compresses catalyst structures into a compact, chemically meaningful latent space. Second, a Diffusion Transformer (DiT) performs generation within this latent space using flow matching. The model incorporates property conditioning—such as adsorbate identity and catalyst class—via classifier-free guidance (CFG), allowing researchers to steer the generative process toward desired targets.

CatDiT demonstrates state-of-the-art performance in generating physically plausible slab-adsorbate structures, outperforming existing models like CatGPT and CatFlow in structure validity and uniqueness. The framework successfully extends inverse design beyond simple alloys to include complex oxide surfaces. In a practical application for the nitrogen reduction reaction (NRR), the model generated 28 DFT-relaxed candidates that satisfy target activity windows, showing a 1.5-fold enrichment over the source distribution. The model also exhibits strong extrapolation capabilities, generating novel compositions not present in the training data.

By enabling property-directed inverse design, CatDiT significantly reduces the computational burden of catalyst discovery. Its ability to handle diverse catalyst classes and multi-property conditioning makes it a versatile tool for researchers aiming to discover high-performance catalysts for specific chemical reactions, effectively expanding the searchable chemical space while maintaining high structural feasibility.