Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 18: SYMD contributed
MM 18.2: Talk
Wednesday, March 19, 2025, 10:30–10:45, H23
Physics-Based Generative Models: Enhanced Structure-Property Sampling in Inverse Materials Design — •Patricia König1, Nicolas Bergmann1, Piero Coronica2, Chiara Panosetti1, Hanna Türk1, Karsten Reuter1, and Christoph Scheurer1 — 1Fritz-Haber-Institut der MPG, Berlin — 2Max Planck Computational and Data Facility, Munich
Data-driven approaches for the inverse design of novel materials with desired properties have become a key tool in materials discovery. Here, we introduce a framework using physics-based Generative Adversarial Networks for enhanced structure-property sampling via latent space design.
We are interested in sampling structures of two chemical systems associated with different relevant physical quantities, like the work function in the electrochemical adsorption of iodide and hydroxide on copper surfaces, and the oxygen chemical potential in the CO to CO2 conversion over an amorphous RuO2 catalyst. As part of our framework, we track and evaluate the structural diversity and convergence of our generator with machine-learning interatomic potentials and quantitative metrics. This enables a high throughput and cost-effective evaluation of structural guesses and their related properties to leverage the full potential of generative models. Concluding, we are showing on two model systems how to explore a vast chemical space of datasets with sparse areas, particularly structures with high free energies in transition states and diverse amorphous surface structures, thereby advancing the understanding and design of novel materials.
Keywords: Generative Adversarial Networks; Inverse Design; Enhanced Structure-Property Sampling; Dataset Engineering; Electrochemistry