Berlin 2018 – scientific programme

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

MM: Fachverband Metall- und Materialphysik

MM 51: Topical Session (Symposium MM): Big Data in Materials Science - Managing and exploiting the raw material of the 21st century

MM 51.4: Talk

Thursday, March 15, 2018, 11:00–11:15, H 0107

Automatic Selection of Atomic Fingerprints for Machine-Learning Potentials — •Giulio Imbalzano — École polytechnique fédérale de Lausanne, Lausanne, Switzerland

Machine learning of atomic-scale properties is revolutionizing the way simulations are performed, making it possible to evaluate interatomic potentials with first-principles accuracy, at a fraction of the cost. The accuracy, speed and reliability of a machine-learning potential, however, depend strongly on the way atomic configurations are represented before being used as inputs. The raw Cartesian coordinates are typically transformed in "fingerprints" that are designed to better encode the symmetries of the problem and the physics of the interactions. I discuss an automatic protocol to select a reduced number of fingerprints out of a large set of candidates, based on the intrinsic correlations of the training data. This procedure can greatly simplify the construction of neural-network potentials that strike the best balance between accuracy and computational efficiency, and has the potential to accelerate by orders of magnitude the evaluation of Gaussian Approximation Potentials based on the Smooth Overlap of Atomic Positions kernel.