Semiempirical Quantum Chemistry in the Age of ab initio Data and Differentiable Programming: II. Parameter Optimization
Year: 2026
Type: preprint
Abstract: Semiempirical quantum chemistry (SQC) represents an invaluable asset in the toolbox of molecular and materials modeling. Thanks to their explicit quantum-mechanical treatment of electrons, they allow to capture an array of local and non-local interaction mechanisms and offers access to electronic structure properties, which can be of crucial importance in order to better understand and design molecular systems and materials. In a previous publication, we presented our recent advances within the PySEQM package, enabling rapid gradient evaluations for SQC parameters and thus developing an efficient and reliable back-end for SQC parametrizations. In this work, we detail our implementation of this parametrization procedure, SQCFit, and its features. Offering a high degree of automation, SQCFit implements a suite of tailor-made tools for the optimization of SQC parameters. Specifically, we present local versus global parametrization schemes, the use of train-validation splits and early stopping to avoid overfitting, and advanced scoring/loss functions including appropriately standardized quantitative error measures as well as more qualitative measures such as the performance in ranking candidate systems according to a given molecular property. Based on a set of test applications, we highlight how SQCFit allows to obtain performant parametrizations tailor-made for the modern use cases of SQC.
Source: ChemRxiv
Authors: Martin Stöhr, Todd J. Martínez
Institutions: SLAC National Accelerator Laboratory, Stanford University
Language: English
Subfield: Atomic and Molecular Physics, and Optics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221606900