SKM 2023 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 29: Data Driven Materials Science: Big Data and Work Flows – Electronic Structure

MM 29.1: Vortrag

Mittwoch, 29. März 2023, 11:45–12:00, SCH A 251

Band Gap and Formation Energy Inference of Solids using Message Passing Neural Networks — •Tim Bechtel, Daniel Speckhard, and Claudia Draxl — Humboldt-Universität zu Berlin, Physics Department and IRIS Adlershof, Berlin, Germany

Graph-based neural networks and, specifically, message-passing neural networks have shown great promise in predicting physical properties of solids. Here, we target three tasks, formation energy regression, metal- non-metal classification, and band gap regression, using data from the AFLOW materials database [1]. In order to find optimal hyperparameters and model architecture, we perform a neural architecture search on the band gap regression task, using a random search algorithm. The model is based on a message passing neural network with edge updates [2], and provides users with uncertainty estimates via Monte- Carlo dropout. We analyze the domain of applicability of the model, for different space group symmetries, atomic species, and corrections applied to the underlying calculation. While we obtain overall excellent results, the model struggles to accurately predict oxide materials. We find that the uncertainty in different domains reflects the model’s predictive performance.

[1] S. Curtarolo et al., Comput. Mater. Sci., 58 (2012), pp. 227-235.

[2] P.B. Jørgensen et al., Preprint at arXiv:1806.03146 (2018).