Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...

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

SOE: Fachverband Physik sozio-ökonomischer Systeme

SOE 10: Data Analytics, Extreme Events, Nonlinear Stochastic Systems, and Networks (joint session DY/SOE)

SOE 10.2: Vortrag

Mittwoch, 18. März 2020, 15:30–15:45, ZEU 118

Constructing accurate and data-efficient molecular force-fields with machine learning — •Igor Poltavskyi, Grégory Fonseca, Valentin Vassilev-Galindo, and Alexandre Tkatchenko — University of Luxembourg, Luxembourg

Employing machine learning (ML) force-fields (FF) is becoming a standard tool in modern computational physics and chemistry. Reproducing potential energy surfaces of any complexity, ML models extend our horizons far beyond the reach of ab initio calculations. One can already perform nanosecond-long molecular dynamics simulations for molecules containing up to a few tens of atoms on a coupled-cluster level of accuracy, providing invaluable information about subtle details of intra-molecular interactions [1,2]. Next challenges are constructing ML FFs to molecules with 1000s of atoms and describing far-from-equilibrium geometries without losing accuracy and efficiency. To reach these goals, we developed methods for optimizing reference datasets and partitioning the problem of training global FFs into parts. By minimizing the prediction error for subsets of molecular configurations obtained by clustering, we can build ML FFs equally applicable for the entire range of reference data. Dividing the configuration space into sub-domains by physical and chemical properties, training corresponding ML models, and combining them into one global model enables highly-accurate FFs for molecules containing hundreds of atoms.

[1] Sauceda et al., J. Chem. Phys. 150, 114102 (2019).

[2] Chmiela et al., Nat. Commun., 9(1), 3887 (2018).