Regensburg 2025 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 3: Data-driven Materials Science: Big Data and Worksflows

MM 3.2: Talk

Monday, March 17, 2025, 10:30–10:45, H10

Assessing the role of physical constraints in machine learning potentials — •Marcel F. Langer, Sergey N. Pozdnyakov, Filippo Bigi, and Michele Ceriotti — Laboratory of Computational Science and Modeling (COSMO) and National Centre for Computational Design and Discovery of Novel Materials (MARVEL), Institute of Materials, École Polytechnique Fédérale de Lausanne, 1015 Lausanne, Switzerland

Machine learning potentials, which approximate the potential energy surface of atomistic systems to enable larger and longer simulations than first-principles methods, have advanced rapidly in recent decades. Much of this development has been driven by the increasingly sophisticated treatment of physical symmetries, in particular invariances, in the underlying machine learning models. However, the rise of so-called unconstrained models, which replace exact invariance with learned approximations, has sparked debate over this approach. Some models even choose to directly predict forces, rendering the resulting force fields non-conservative. We investigate the effectiveness of such models and evaluate the impact of disregarding physical constraints for practical simulations. In particular, we study the effects of breaking rotational symmetry in a machine-learning potential for water [1] and discuss the potential consequences of direct force predictions.

[1]: M.F. Langer, S.N. Pozdnyakov, and M. Ceriotti, Mach. Learn.: Sci. Technol. 5 04LT01 (2024)

Keywords: machine learning; equivariance; water; machine learning potential; symmetry