Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 9: Poster
MM 9.44: Poster
Montag, 17. März 2025, 18:30–20:30, P1
Fitting Tensorial Properties with MACE: a Study of Li2Ti5O12 Electric Field Gradient Tensors — Elena Gelzinyte1, Karsten Reuter1, Christian Carbogno1, and •Johannes T. Margraf2 — 1Fritz-Haber-Institut der MPG, Berlin — 2University of Bayreuth
Machine learning interatomic potentials, which serve as surrogate models for predicting a structure’s energy and forces, have significantly accelerated atomistic simulations. Equivalent approaches have been applied to predict other structural or atomic properties, such as charges, dipole moments, and polarisabilities. One such framework is MACE, a higher-order equivariant neural network [1]. Due to the way its internal features are constructed, the output part of the model may be readily modified to suit the symmetry of the target property. In this presentation, we discuss the required modifications for fitting atomic tensorial quantities and the resulting model’s applicability, limitations, and advantages. For illustration, we focus on the prediction of electric field gradient tensors (a per-atom traceless symmetric tensor) using a Li2Ti5O12 data set [2]. We consider the improvement in fitting the tensorial properties directly, rather than derived scalar properties, and compare the modified MACE’s results with those of λ-SOAP [3], discussed in [2].
[1] I. Batatia et al., NeurIPS 35, 11423 (2022).
[2] A.F. Harper et al., DOI: 10.26434/chemrxiv-2024-j0kp2 (2024).
[3] A. Grisafi et al., Phys. Rev. Lett. 120, 036002 (2018).
Keywords: Equivariant Neural Network; Tensorial Property prediction; Machine learning interatomic potential; MACE