Approximations of the Iterative Stockholder Analysis scheme using exponential basis functions

Abstract

In this work, we introduce several approximations of the Iterative Stockholder Analysis (ISA) method based on exponential basis functions. These approximations are categorized into linear and non-linear models, referred to as LISA and NLIS, respectively. By particular choices of hyperparameters in the NLIS model, both LISA and the Minimal-Basis Iterative Stockholder (MBIS) method can be reproduced. Four LISA variants are constructed using systematically generated exponential basis functions derived from the NLIS model applied to atomic systems. The performance of these LISA variants and NLIS models is benchmarked on 15 small molecules, including neutral, anionic, and cationic species. To facilitate comparison, we propose several metrics designed to highlight differences between the methods. Our results demonstrate that LISA, employing Gaussian basis functions derived from the NLIS model on isolated atomic systems, achieves an optimal balance of computational accuracy, robustness, and efficiency, particularly in minimizing the objective function.