Parts | Days | Selection | Search | Updates | Downloads | Help
DY: Fachverband Dynamik und Statistische Physik
DY 8: Artificial Intelligence in Condensed Matter Physics II (joint session TT/DY)
DY 8.1: Talk
Monday, March 18, 2024, 15:00–15:15, H 3025
Uncertainty-aware active learning for interatomic potentials generation and its applications for spin dynamics — •Valerio Briganti and Alessandro Lunghi — School of Physics, AMBER and CRANN Institute, Trinity College, Dublin 2, Ireland
In the last decade, the materials science community has increasingly exploited the potential of AI for various applications, ranging from the discovery of new materials to the generation of interatomic potentials (IP). Developments in the latter have enabled to perform molecular dynamics simulations with unprecedented timescales, with the promise of successfully overcoming the computational costs required by ab initio methods keeping a sufficiently high accuracy. Two of the main challenges in this field are the design of models to allow greater transferability and the optimal selection of data to be included in the training set. In this contribution, I will show how a linear regression model based on SNAP [1] augmented with an uncertainty aware active learning procedure [2] can efficiently lead to the generation of accurate IPs able to simulate the dynamics of organic and open-shell compounds at room temperature. In addition to this, I will also present the performance of machine learning IPs for prediction of phonons and spin-phonon relaxation time.
[1] A.P. Thompson et al., J. of Comp. Phys., 285 (2015) 316.
[2] V. Briganti, A. Lunghi, Mach. Learn.: Sci. Technol. 4 (2023) 035005.
Keywords: machine learning; interatomic potential; active learning; phonons; spin dynamics