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 104: Topical Session: Data Driven Materials Science - Machine Learning Applications (joint session MM/CPP)
CPP 104.4: Vortrag
Donnerstag, 19. März 2020, 18:15–18:30, BAR 205
Transferable Gaussian Process Regression for prediction of molecular crystals harmonic free energy. — •Marcin Krynski1 and Mariana Rossi1,2 — 1Fritz Haber Institute of the Max Planck Society, Berlin, Germany — 2MPI for Structure and Dynamics of Matter, Hamburg, Germany
Organic molecular crystals are a large group of compounds with properties tied strongly to the crystallographic structure of their numerous polymorphs.
While thermodynamic free energies are necessary for obtaining a reliable polymorph energy ranking[1,2], their inclusion in large-scale simulations for polymorph screening is challenging, because dispersion-corrected DFT accuracy is needed in order to capture the complex charge rearrangement and bond-softening.
Therefore, to predict harmonic Helmholtz free energies, we devised a framework that employs the transferable Gaussian Process Regression model with Smooth Overlap of Atomic Positions[3] descriptors representing local atomic environments.
We developed strategies based on farthest point sampling to minimize the size of the training set and to ensure statistical diversity.
We benchmark our framework on a set of 444 hydrocarbon crystal polymorphs.
Superior performance and high prediction accuracy, with mean absolute deviation below 0.1 meV/atom is achieved by a hyperparameter optimisation performed on empirical-potential models, ensuring sensitivity to longer-range structural patterns.
[1] J. Nyman and G. Day, CrystEngComm 17, 5154 (2015);
[2] M. Rossi, P. Gasparotto, M. Ceriotti, PRL 117, 115702 (2016);
[3] A. P. Bartók, R. Kondor, G. Csányi, PRB 87, 184115 (2013).