Comparison of performance of universal Machine Learning Interatomic Potentials

Z. Holman¹, M. Khanore¹, A. Klic¹, J. Hlinka¹, J. Drahokoupil¹, P. Marton¹

¹Institute of Physics of the Czech Academy of Sciences, Prague, Czech Republic

²Institute of Mechatronics and Computer Engineering, Technical University of Liberec, Czech Republic

holman@fzu.cz

Perovskites are materials with the crystal structure ABX₃ that can contain many different cations and anions. This flexibility allows their structural, electrical, and optical properties to be tuned and makes them important for a wide range of applications.

Figure 1. Structures of simulated perovskites a) BaTiO₃ b) CaTiO₃ c) PbTiO₃

Understanding and improving their properties relies on computational modelling. Density Functional Theory (DFT) is commonly used as a reference method, but it is computationally expensive. Therefore, simplified approaches such as atomistic models, shell models, effective Hamiltonians, and phase-field methods are often used to study larger systems and processes like domain evolution and phase transitions. These models need parameters obtained either from experiments or from first-principles calculations. Recently, machine-learning interatomic potentials (MLIPs) have become a useful alternative because they can reach near-DFT accuracy while allowing relatively large and longe simulations of perovskite oxides.

Figure 2. Energy difference between DFT and PET-MAD uMLIP per structure.

In this work we evaluate several variants of universal machine-learning interatomic potentials (uMLIP) [1,2,3,4]. We compare their accuracy of energy and forces predictions with DFT calculations. In this work we focused on performance of MACE, PET-MAD uMLIP.

Acknowledgements: The research was supported by Czech Science Foundation Grant No. 24-11275S.

[1] Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. Mace: Higher order       equivariant message passing neural networks for fast and accurate force fields, 2023.

[2] Dávid Péter Kovács, Ilyes Batatia, Eszter Sára Arany, and Gábor Csányi. Evaluation of the mace force field         architecture: From medicinal chemistry to materials science. The Journal of Chemical Physics, 159(4), 2023.

[3] Arslan Mazitov, Filippo Bigi, Matthias Kellner, Paolo Pegolo, Davide Tisi, Guillaume Fraux, Sergey Pozdnyakov, Philip Loche, and Michele Ceriotti. Pet-mad, a lightweight universal interatomic potential for advanced materials modeling, 2025.

[4] Cesare Malosso, Filippo Bigi, Paolo Pegolo, Joseph W. Abbott, Philip Loche, Mariana Rossi, Michele Ceriotti, and Arslan Mazitov. High-quality, high-information datasets for universal atomistic machine learning, 2026.