Atomistic Structure Search of Nanoscale Systems Using Generative Modeling and Evolutionary Algorithm

Knowledge of the exact atomistic structure is crucial for computational analysis of unknown phases observed in experiments and for discovering novel materials. We previously developed the FANTASTX (Fully Automated Nanoscale To Atomistic Structure from Theory and eXperiments) package, a multi-objective evolutionary algorithm to determine atomistic structure from experimental characterization data. This approach relies on sampling the potential energy landscape of thermodynamically stable structures to find the best match between the simulated phase and observed experimental data. In this work, we develop a method for efficient sampling of candidate structures based on generative modeling.<br/><br/>Using a variational autoencoder (VAE), we transform the original, structure-oriented search space to a low-dimensional, property-oriented one. Structures are represented using an invertible point-cloud representation, and an auxiliary network is used to steer the latent space distribution according to the desired property. In this talk, we will discuss key components of this approach, including network optimization and sampling of the resulting latent space. We report the results of preliminary tests performed on CdTe and IrOx datasets, which utilize molecular dynamics and density functional theory for structure relaxation, respectively. This VAE-augmented approach enables more efficient sampling by refining the search space to meet expected property criteria, such as energy or mismatch to experimental data.