Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 62: Methods in Computational Materials Modelling (methodological aspects, numerics)
MM 62.2: Vortrag
Donnerstag, 15. März 2018, 16:00–16:15, TC 006
Combination of machine learning and high-throughput DFT calculations for the prediction of thermodynamic stability — •Jonathan Schmidt1, Jingming Shi4, Pedro Borlido2, Liming Chen3, Silvana Botti2, and Miguel Marques1 — 1Institut für Physik, Martin-Luther-Universität Halle-Wittenberg, Germany — 2Institut für Festkörpertheorie und -optik, Friedrich-Schiller-Universität Jena and European Theoretical Spectroscopy Facility, Germany — 3Département MI, Laboratoire ICTT, Ecole Centrale de Lyon, France — 4Institut Lumière Matière, Université de Lyon, France
We perform a large-scale benchmark of machine learning methods for the prediction of the thermodynamical stability of solids. We start by constructing a data set that comprises density functional theory calculations of around 250,000 cubic perovskite systems. Incidentally, around 500 of these are thermodynamically stable but are not present in crystal structure databases. This data set is then used to train and test a series of machine learning algorithms to predict the distance to the convex hull of stability. In particular, we study the performance of ridge regression, random forests, extremely randomized trees (including adaptive boosting), and neural networks. We find that extremely randomized trees give the best results and use this method in combination with DFT to explore ternary compounds with the AB2C2 composition. By using machine learning we reduce the overall calculation cost by around 75% and find that there may be 10 times more stable compounds in these phases than previously known.