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

DY: Fachverband Dynamik und Statistische Physik

DY 10: Modeling and Simulation of Soft Matter (joint session CPP/DY)

DY 10.1: Vortrag

Montag, 5. September 2022, 15:00–15:15, H39

Machine Learning of consistent thermodynamic models using automatic differentiation — •David Rosenberger¹, Kipton Barros², Timothy Germann², and Nicholas Lubbers² — ¹Freie Universität Berlin, Berlin, Germany — ²Los Alamos National Laboratory, Los Alamaos, NM, USA

Instead of fitting suitable analytical expressions to thermophysical data, we propose to combine automatic differentiation and artificial neural networks (ANNs) to obtain complex equations of state (EOS) for arbitrary systems. Rather than training directly on the properties of interest, we train an ANN on a model free energy whose partial derivatives match the thermophysical properties measured in experiment. We show that this method is advantageous over direct learning of thermodynamic properties, in terms of both accuracy and the exact preservation of the Maxwell relations. Furthermore, the method can implicitly solve the integration problem of computing the free energy of a system without explicit integration given appropriate data to learn from.