Inverse Design of Materials Using Diffusion Probabilistic Models

The so-called chemical space includes all the molecules and materials we already know and those that have not yet been discovered. The total number of compounds is larger than the number of stars in the universe and so far, only less than a drop of water has been explored in the ocean of chemical space. Exploring this space to find useful materials for a given application is thus a huge and expensive task. The objective of the present proposal is to develop and harness the power of new machine learning (ML) methods to help materials chemists with this task by prioritizing the available options at the materials design stage. We will focus on the specific case of glass-ceramic materials, which contain a dispersion of ordered crystals in a disordered glass matrix. Glass ceramics offer superior mechanical properties, but existing ML methods have not yet been successfully used to create novel materials. To this end, we will here develop diffusion probabilistic models to first learn the connection between chemistry/structure and material properties and then allow an inverse design of new materials with desired properties. This framework will enable materials chemists to test design hypotheses against the ML predictions trained on all available data.