Regensburg 2019 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 10: Methods in Computational Materials Modelling (methodological aspects, numerics)

MM 10.7: Talk

Monday, April 1, 2019, 17:30–17:45, H45

Machine learning for training lattice based models — Mattias Ångqvist, Erik Fransson, J. Magnus Rahm, and •Paul Erhart — Chalmers University of Technology, Department of Physics, Gothenburg, Sweden

Creating lattice based models for studying configurational and vibrational effects has now become possible with little technical effort. The basic idea is that the models will be trained using reference data which commonly comes from expensive density functional theory calculations. If the training goes well the model will both accurately predict the training data but may also predict unseen data and do so at a fraction of the computational cost. The training is evaluated by cross validation by splitting up the reference data into testing and training sets. The number of optimization algorithms that one may choose to do the training with are very numerous. In this talk I will present how the different optimization algorithms perform when training these so called lattice based models both with respect to error estimation and to the prediction of e.g. thermodynamic quantities.