Regensburg 2025 – scientific programme

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

MM: Fachverband Metall- und Materialphysik

MM 18: SYMD contributed

MM 18.1: Talk

Wednesday, March 19, 2025, 10:15–10:30, H23

Relation Between Element Specific Chemistry and Basis Set Size of Machine Learned Interatomic Potentials — •Haitham Gaafer, Jan Janssen, and Joerg Neugebauer — Max Planck Institute for Sustainable Materials, Düsseldorf, Germany

Machine learned interatomic potentials (MLIP) have gained popularity in materials science for their scalability and accuracy on par with the Density Functional Theory (DFT) training data. Based on the linear scaling with the number of neighbors, the primary focus in the recent years was increasing the flexibility of MLIPs to further improve their accuracy, as demonstrated by approaches such as Neural Network potentials and the Atomic Cluster Expansion (ACE). The Bessel functions and Chebyshev polynomials gained popularity as basis sets to represent the atomic bonds and orbitals. Nonetheless, the connection between an MLIP's basis set and its capability to represent the chemical complexity of various elements is not yet well understood.

In this study, we use ACE, as implemented in the Pacemaker software package, to investigate three non-magnetic transition metals (i.e., Al, Au, and Cu). For each element, we parameterize computationally efficient ACE potentials based on a minimal basis set to achieve a given root-mean-square error (RMSE) across both training and testing datasets. We find that the complexity of the MLIP primarily depends on the scaling of the per-atom energy distribution rather than the chemical complexity of the elements. Consequently, it is primarily a numerical effect rather than a chemical effect.

Keywords: Machine Learned Interatomic Potentials; Chemical Complexity; Hyperparameter study