Abstract
In recent years, machine learning (ML) for predicting material properties hasemerged as a quicker alternative to experimental and simulation-basedinvestigations. Standard ML approaches tend to utilize specific domainknowledge when designing feature inputs. Each ML property predictor thenrequires a set of tailored structural features - this can commonly lead todrawbacks, due to the small number of implemented features and their lack ofprediction transferability across different predictors. The latter has beenempirically observed in the case of guest uptake predictors for nanoporousmaterials, where local and global porosity features become dominant descriptorsat low and high pressures, respectively. Here, we provide a more holisticfeature representation for materials structures using tools from topologicaldata analysis and persistent homology to describe the geometry and topology ofnanoporous materials at various scales. We demonstrate an application of thesetopology-based feature representations to predict methane uptakes for zeolitestructures in the range of 1-200 bar. These predictions show a root-mean-squaredeviation decrease of up to 50% at low pressures and up to 100% across highpressures in comparison to a model based on commonly used local and globalfeatures. Similarly, the topology-based model shows an increase of 0.2-0.3 inR2 score in comparison to the commonly used porosity descriptors. Notably,unlike the standard porosity features, the topology-based features showaccuracy across multiple different pressures. Furthermore, we show featureimportance in terms of different topological features, thus elucidatinginformation about the channel and pore sizes that correlate best to adsorptionproperties. Finally, we demonstrate that ML models relying on a combination oftopological and commonly employed descriptors provide even better guest uptakeregressors.