SKM 2023 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 31: Data Driven Materials Science: Big Data and Work Flows – Machine Learning

MM 31.3: Vortrag

Mittwoch, 29. März 2023, 16:15–16:30, SCH A 251

Learning to Spell Materials - Coordinate-free Discovery with Natural Language Processing — •Konstantin Jakob, Karsten Reuter, and Johannes T. Margraf — Fritz Haber Institute, Berlin, Germany

Over the last decade, computational screening with structure-based machine learning models has led to some advances in the discovery of novel inorganic materials. Unfortunately, the overwhelmingly large space of possible compositions and atomic configurations together with the exceeding rarity of well-suited candidates ultimately poses a limit to the applicability of this approach. In contrast, purely composition-based representations neglect differences in the chemical properties of different crystal polymorphs and thus lack accuracy. A middle ground between full structural and simple compositional representations has been established for organic molecules using string representation such as SMILES. While these have proven highly advantageous for molecular discovery when combined with natural language processing models, analogous representations for the more complex class of inorganic materials are still missing. Bridging this gap, we investigate the performance of recurrent neural networks (RNNs) in predicting crystallographic properties by reading a materials composition element by element. Their striking accuracy suggests that symmetry- or prototype-based string representations could be generated with little computational effort at a large scale. The invertibility of these intermediate representations via restricted structure searches is investigated, paving the way to their application for conditional generative models.