Abstract
The vastness of chemical space presents a long-standing challenge for the exploration of new compounds with pre-determined properties. In materials science, crystal structure prediction has become a mature tool for mapping from composition to structure based on global optimisation techniques. Generative artificial intelligence now offers the means to efficiently navigate larger regions of crystal chemical space informed by structure-property datasets of materials. Here, we introduce a model, named Chemeleon, designed to generate chemical compositions and crystal structures by learning from both textual descriptions and three-dimensional structural data. The model employs denoising diffusion techniques for compound generation using textual inputs aligned with structural data via cross-modal contrastive learning. The potential of this approach is demonstrated for multi-component compound generation, including the Zn-Ti-O ternary space, and the prediction of stable phases in the Li-P-S-Cl quaternary space of relevance to solid-state batteries.