Regensburg 2022 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 25: Data Driven Materials Science: Computational Frameworks / Chemical Complexity

MM 25.2: Vortrag

Mittwoch, 7. September 2022, 16:15–16:30, H46

Efficient parameterization of the atomic cluster expansion — •Anton Bochkarev, Yury Lysogorskiy, Matous Mrovec, and Ralf Drautz — Atomistic Modelling and Simulation, ICAMS, Ruhr-Universität Bochum, D-44801 Bochum, Germany

The atomic cluster expansion (ACE) is a machine learning model with a complete basis set representation that can be used for constructing interatomic potentials. These potentials can be both, general-purpose as well as potentials designed for a specific application. The former are usually more reliable and accurately describe materials in various conditions, but building such models often requires a materials specific expertise and extensive training datasets. Purpose-specific potentials have only limited ranges of applicability, but are also less demanding in terms of training data. Here we demonstrate a complete, efficient and largely automated framework for constructing quantum accurate ACE models for various applications. Our framework includes automated data generation, model parameterization and validation. Efficient implementations on CPU and GPU hardware enable large scale simulations.