Multi Element Glass Dataset for Benchmarking Generative Models for Disordered Materials

Diffusion generative models originally became popular for image generation but have more recently also found adoption by the materials science community. By conditioning such models on material properties, new materials can be predicted that match a set of given target properties. In the domain of crystalline materials, significant progress has been made with the development of new model architectures, driven by the availability of large databases, such as the Materials Project. Unfortunately, data availability for disordered materials such as glasses is extremely limited.<br/><br/>We therefore developed the Multi Element Glass Dataset (MEGD) to train and benchmark generative models for amorphous materials. The molecular dynamics dataset consists of over 9000 samples of oxide glass structures together with properties including thermal expansion coefficients, heat capacities and elastic constants. As the glass structure depends on the thermal history, we developed a dedicated simulation workflow to generate the structural samples. The workflow is based on the BMP1) interatomic potential and determines a suitable initial temperature for each composition that ensures proper melting while avoiding evaporation events. The MEGD includes a wide range of compositions including up to three glass formers (Si, P, Al) and up to four out of seven modifiers (Li, Be, K, Ca, Ti, Ba, Zn).<br/><br/>We highlight our own benchmarking results based on the MEGD using our recently developed AMDEN (Amorphous Materials DEnoising Network) framework.