Diffusion Generative Models for the Inverse Design of Disordered Materials

Designing new materials with tailored properties is one of the ultimate goals of materials science. Traditionally, new materials are discovered through trial-and-error based experiments, guided by expert knowledge. However, recently a new approach, called inverse design, has become possible in which machine learning models are used to directly predict a new materials structure and composition given its target properties.<br/><br/>To enable inverse design of disordered materials such as glasses, we developed the AMDEN (Amorphous Materials DEnoising Network) framework. Our implementation is based on an equivariant graph neural network architecture and can be conditioned to generate material samples with specific properties. We developed an extensive data set of over 9000 oxide glasses containing up to three different glass formers and seven modifiers, for which we computed properties including elastic moduli, heat capacities and thermal expansion coefficients to benchmark our method.<br/><br/>Our results show that the model can generate high quality glass samples with a valid short-range order. By conditioning the model, samples with specific properties can be generated, even extending beyond the scope of the training data, indicating excellent extrapolation capabilities. Both composition and structure are modified by the model to obtain the desired properties, offering an exciting opportunity to explore property-structure relationships in disordered materials. Despite the model’s excellent performance in terms of structure and properties, challenges arise when it comes to generating low-energy, relaxed structures. To address this, we extended the traditional denoising scheme by integrating a Monte Carlo based relaxation procedure.