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 3.1: Talk

Monday, March 17, 2025, 10:15–10:30, H10

Benchmarking DFT Functionals at Finite Temperature with ASSYST and MLIPs — •Marvin Poul and Jörg Neugebauer — Max-Planck-Institut für Nachhaltige Materialien

A key ingredient to the accuracy of Density Functional Theory (DFT) calculations is the chosen approximation to the exchange-correlation functional. Local Density Approximation (LDA) and Generalized Gradient Approximation (GGA) calculations often bracket experimental observations, but systematic exploration of the behavior of different density functionals is hindered by the high computational cost of DFT in realistic applications, especially concerning finite temperature properties. Using the ASSYST[1] method, we automatically generate unary, general purpose Atomic Cluster Expansion (ACE) Machine Learning Interatomic Potentials (MLIPs) for a range of metals using LDA, PBE and r2SCAN functionals. The key advantage of ASSYST lies in the small cells (≤ 10 atoms per cell) that it generates as training data. This allows us to relabel the data using different functionals very efficiently. We then use these potentials to calculate melting curves, thermal expansion, and formation energies of various defects (grain boundaries, surfaces, point defects) to systematically assess strengths and weaknesses of the DFT functionals. In general, we find good agreement with corresponding DFT results, showing that ASSYST can reliably create transferable potentials for metals at DFT accuracy.

[1]: https://www.researchsquare.com/article/rs-4732459/v1

Keywords: MLIP; DFT; Interatomic Potentials