Regensburg 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 34: Development of Calculation Methods

MM 34.5: Vortrag

Freitag, 21. März 2025, 11:15–11:30, H10

Bayesian Uncertainty Estimates for Spin-Component-Scaled Second-Order Møller-Plesset Perturbation Theory — •Elisabeth Keller and Johannes Margraf — Universität Bayreuth, Bayreuth, Germany

Spin-component-scaled second-order Møller-Plesset perturbation theory (SCS-MP2) improves upon MP2 by separately scaling the same-spin and opposite-spin MP2 contributions, achieving near coupled cluster (CCSD(T)) accuracy in some applications. However, the optimal scaling parameters vary by target domain, limiting the transferability of any given SCS-MP2 parameterization. To address this limitation, we employ a Bayesian multilevel linear regression model to obtain a robust parameterization for SCS-MP2, termed BSCS-MP2, predicting various energetic properties, including total, atomization, reaction, and non-covalent interaction energies. Additionally, the Bayesian model provides credible intervals to quantify the uncertainty of the BSCS-MP2 energy predictions. We evaluate how these uncertainty estimates adjust to data quality and model complexity, and assess their robustness for out-of-sample inference.

Keywords: SCS-MP2; MP2; Bayesian; Uncertainty estimate