Dismai-Bench: benchmarking and designing generative models using disordered materials and interfaces

Generative models have received significant attention in recent years for materials science applications, particularly in the area of inverse design for materials discovery. However, these models are usually assessed based on newly generated, unverified materials, using heuristic metrics such as charge neutrality, which provide a narrow evaluation of a model's performance. Also, current efforts for inorganic materials have predominantly focused on small, periodic crystals (≤20 atoms), even though the capability to generate large, more intricate and disordered structures would expand the applicability of generative modeling to a broader spectrum of materials. In this work, we present the Disordered Materials & Interfaces Benchmark (Dismai-Bench), a generative model benchmark that uses datasets of disordered alloys, interfaces, and amorphous silicon (256-264 atoms per structure). Models are trained on each dataset independently, and evaluated through direct structural comparisons between training and generated structures. Such comparisons are only possible because the material system of each training dataset is fixed. Benchmarking was performed on two graph diffusion models and two (coordinate-based) U-Net diffusion models. The graph models were found to significantly outperform the U-Net models due to the higher expressive power of graphs. While noise in the less expressive models can assist in discovering materials by facilitating exploration beyond the training distribution, these models face significant challenges when confronted with more complex structures. To further demonstrate the benefits of this benchmarking in the development process of a generative model, we considered the case of developing a point-cloud-based generative adversarial network (GAN) to generate low-energy disordered interfaces. We tested different GAN architectures and identified reasons for good/poor performance. We show that the best performing architecture, CryinGAN, outperforms the U-Net models, and is competitive against the graph models despite its lack of invariances and weaker expressive power. This work provides a new framework and insights to guide the development of future generative models, whether for ordered or disordered materials.