Dresden 2020 – wissenschaftliches Programm
Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 68: Topical Session: Data Driven Materials Science - Descriptors (joint session MM/CPP)
CPP 68.1: Vortrag
Mittwoch, 18. März 2020, 11:45–12:00, BAR 205
Evaluating representations of atomistic systems for machine learning — •Marcel Langer1 and Matthias Rupp1,2 — 1Fritz Haber Institute of the Max Planck Society, Berlin, Germany — 2Citrine Informatics, Redwood City, CA, USA
Interpolating between computationally expensive first-principles calculations with fast machine-learning surrogate models increases the feasible scope of exploration when a large space of potentially similar structures is sampled, for instance in the search for novel materials or the exploration of phase diagrams.
The choice of representation of the atomistic systems under consideration is important for the accuracy of such surrogate models. We present a rigorous empirical comparison of the Many-Body Tensor Representation [1], Smooth Overlaps of Atomic Positions [2], and Symmetry Functions [3] for energy predictions of molecules and materials. In this, we control for data distribution, hyper-parameter optimization, and regression method. We also investigate the relationship between predictive performance and computational cost, and discuss how to assess predictions beyond mean errors, which cannot fully describe model behaviour in practice. [4,5]
[1] H. Huo and M. Rupp, arXiv, 1704.06439 (2017)
[2] A. Bartók, R. Kondor., G. Csányi, Phys. Rev. B 87, 184115 (2013)
[3] J. Behler, J. Chem. Phys. 134, 074106 (2011)
[4] C. Sutton et al., ChemRxiv, 9778670 (2019)
[5] Z. del Rosario et al., arXiv, 1911.03224 (2019)