Parts | Days | Selection | Search | Updates | Downloads | Help
MA: Fachverband Magnetismus
MA 39: SrTiO3: A Versatile Material from Bulk Quantum Paraelectric to 2D Superconductor II (joint session TT/KFM/MA/O)
MA 39.9: Talk
Thursday, March 21, 2024, 17:15–17:30, H 0104
Machine-learning-backed evolutionary exploration of the SrTiO3(110) surface phase diagram — •Ralf Wanzenböck1, Florian Buchner1, Michele Riva2, Jesús Carrete3, 1, and Georg K. H. Madsen1 — 1Institute of Materials Chemistry, TU Wien, A-1060 Vienna, Austria — 2Institute of Applied Physics, TU Wien, A-1040 Vienna, Austria — 3Instituto de Nanociencia y Materiales de Aragón (INMA), CSIC-Universidad de Zaragoza, 50009 Zaragoza, Spain
We use Clinamen2, a modern functional-style Python implementation of the covariance matrix adaptation evolution strategy (CMA-ES), to gain insights into the lesser-known regions of the complex SrTiO3(110) surface phase diagram. To speed up the process, we leverage the transferability of a neural-network force field (NNFF) implemented on top of the state-of-the-art JAX framework.
Starting from smaller reconstructions in well-explored phases, such as the 4× 1 surface reconstruction [Wanzenböck et al., Digit Discov 1, 703-710 (2022)], the NNFF is iteratively refined using an active learning workflow that relies on uncertainty estimation techniques [Carrete et al., J. Chem. Phys 158, 204801 (2023)]. We show how this workflow and the underlying uncertainty metric lead to a flexible NNFF, highlighted by the exploration of out-of-sample SrTiO3(110)-(2× n) reconstructions.
Keywords: SrTiO3; surface reconstruction; machine learning; evolutionary search