Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 3: Data-driven Materials Science: Big Data and Worksflows
MM 3.9: Talk
Monday, March 17, 2025, 12:30–12:45, H10
Learning Disorder in Generative Materials Discovery - Bridging Prediction and Experiment — •Konstantin Jakob1, Aron Walsh2, Karsten Reuter1, and Johannes T. Margraf1,3 — 1Fritz-Haber-Institut der MPG, Berlin, Germany — 2Imperial College London, London, UK — 3University of Bayreuth, Bayreuth, Germany
In recent years, generative machine learning (ML) models have demonstrated tremendous potential for the design and discovery of new materials. This has led to extensive predictions of previously unknown, potentially stable inorganic materials. However, current models suffer from the fact that the underlying training data is purely based on density functional-calculations for small, ideal crystals. As a consequence, many of the supposedly new materials are in fact experimentally known as disordered crystals. In this work, we address this issue by performing a thorough analysis of crystal disorder in the experimental structures of the Inorganic Crystal Structure Database (ICSD). Based on this, we develop disorder classification models and representations that can predict the likelihood of disorder across chemical space. Eventually, these concepts will allow us to extend current generative models to realistic crystal systems and bridge the gap between prediction and experiment.
Keywords: Materials Discovery; Crystal Disorder; Machine Learning; Language Models