Multi-modal conditional diffusion model using signed distance functions for metal-organic frameworks generation

The design of porous materials with user-desired properties has been a great interest for the last few decades. However, the flexibility of target properties has been highly limited, and targeting multiple properties of diverse modalities simultaneously has been scarcely explored. Furthermore, although deep generative models have opened a new paradigm in materials generation, their incorporation into porous materials such as metal-organic frameworks (MOFs) has not been satisfactory due to their structural complexity. In this work, we introduce MOFFUSION, a latent diffusion model that addresses the aforementioned challenges. Signed distance functions (SDFs) are employed for the input representation of MOFs, marking their first usage in representing porous materials for generative models. Using the suitability of SDFs in describing complicated pore structures, MOFFUSION exhibits exceptional generation performance, and demonstrates its versatile capability of conditional generation with handling diverse modalities of data, including numeric, categorical, text data, and their combinations.