TopSearch: a Python package for topographical analysis of machine learning models and physical systems

Abstract

Machine learning (ML) is now ubiquitous in all scientific fields, but there remains a significant challenge to understanding and explaining model performance (Angelov et al., 2021;Zhang et al., 2021). Therefore, there is increasing interest in applying methods from other scientific disciplines (e.g. physics and biology) to improve the performance and explainability of machine learning algorithms (Hassabis et al., 2017;Karniadakis et al., 2021). One methodology that has proved useful to understand machine learning performance is the energy landscape framework from chemical physics (Wales, 2003).The energy landscape framework is a set of algorithms that map the topography of continuous surfaces by their stationary points. The topography is encoded as a weighted graph (Noé & Fischer, 2008) and in application to potential energy surfaces all physical properties of a system can be extracted from this graph (Swinburne & Wales, 2020). Examples of the methodology applied to potential energy surfaces explain physical phenomena for proteins (Röder et al., 2019), small molecules (Matysik et al., 2021), atomic clusters (Csányi et al., 2023 and crystalline solids (Pracht et al., 2023).