Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 19: Poster Session
MM 19.8: Poster
Dienstag, 10. März 2026, 18:00–20:00, P5
Benchmarking Local Geometry Optimization Algorithms for Realistic Potential Energy Surfaces of Solids — •David Greten1, Konstantin Jakob1, Karsten Reuter1, and Johannes T. Margraf1,2 — 1Fritz-Haber-Institut der MPG, Berlin — 2Universität Bayreuth
Efficient and robust local structure relaxations are central to computational materials discovery. We benchmark several widely used local optimizers (based on the BFGS, FIRE and conjugate gradient approaches) for relaxing inorganic crystal structures on complex many-body potential energy surfaces obtained from a universal machine-learned interatomic potential. From relaxations of over 170,000 trial structures generated via element substitution, we quantify convergence rates, relaxation efficiency, and the stability and diversity of obtained minima. Some algorithms (e.g. BFGS with line-search and SciPy's conjugate gradient) show substantially higher convergence rates and consistently yield low energy minima. Meanwhile, the computational effort of the tested algorithms is broadly comparable. Based on these insights we discuss how to optimally combine algorithms to obtain a good trade-off between robustness and diversity of explored minima.
Keywords: Relaxation; Optimizers; Benchmark; MACE; MLIP