AMDEN: Amorphous Materials DEnoising Network

Disordered (amorphous) materials, such as glasses, are emerging as promising candidates for applications within energy storage, nonlinear optics, and catalysis. Inverse design aims to directly predict the composition and structure of new materials with targeted properties using machine learning models. This avoids the time-consuming trial-and-error process of traditional materials design and has the potential to significantly accelerate the discovery of new materials.In this work, we introduce AMDEN (Amorphous Material DEnoising Network), a diffusion model-based framework that generates structures of amorphous materials and can be conditioned on target properties.We demonstrate inherent challenges for diffusion models to generate relaxed structures. These low-energy configurations are typically obtained through a thermal motion-driven random search-like process that cannot be replicated by standard denoising procedures.We therefore introduce an energy-based AMDEN variant that implements Hamiltonian Monte Carlo refinement for generating these relaxed structures. We further introduce several amorphous material datasets with diverse properties and compositions to evaluate our framework and support future development.