Abstract
Complex machine learning (ML) models applied within computational chemistry and materials science tend to be seen as black boxes, yielding property predictions given some input features. While the purpose of ML methods is often to circumvent computationally expensive first-principles calculations, the fact that the inner workings of the models are not understood conceals chemical insight and knowledge regarding the underlying data and physical correlations within it. Knowing what a model is learning from the data and how outputs are formed is also useful in facilitating the justification and wider adoption of ML solutions. Here, we present an important contribution in this direction by exploring and explaining the hydrogen adsorption properties of defective nitrogen-doped carbon nanotubes (NCNTs) through density functional theory simulations and machine learning-based data analysis. As the main highlight, we demonstrate the application of a recent game-theoretic approach to deconvolute and interrogate the trained ML models, revealing how various structural, chemical, and electronic features contribute toward the hydrogen affinities of roughly 6500 different NCNT adsorption sites. The employed method of Shapley additive explanations (SHAP) attributes locally accurate importances to the investigated features, unraveling high spin polarization, narrow highest occupied molecular orbital–lowest unoccupied molecular orbital (HOMO–LUMO) gap, small dopant–adsorption site separation, and diverse angle and coordination effects as particularly impactful for increasing hydrogen adsorption strengths. The SHAP method is shown capable of promoting a deep understanding of complex feature–activity relationships, facilitating research efforts such as rational catalyst design for energy conversion applications.
Highlights
Rational materials development is pivotal in facilitating a sustainable deployment of modern energy conversion technologies empowering envisioned concepts such as the hydrogen economy.[1]
The high importance of the adsorption site spin polarization coincides with previous notions of catalytically activating electron transfer and spin modulation effects in heteroatom-doped carbon nanomaterials.[23,26,67−70] While such effects have been mainly investigated and identified for the ORR, our work reveals that similar electronic effects are likely to influence the HER as inferred using the adsorption energy as an activity descriptor
Tackling large data sets using machine learning (ML) models is attractive, and analyzing the output using the Shapley additive explanations (SHAP) methodology produces reliable feature importances that help in narrowing down which attributes of the investigated catalyst candidates are most important with respect to activity descriptors
Summary
Rational materials development is pivotal in facilitating a sustainable deployment of modern energy conversion technologies empowering envisioned concepts such as the hydrogen economy.[1]. While computational chemistry and density functional theory (DFT) have mitigated this problem by providing complementary tools to assess nanoscale structural and electronic properties of, e.g., catalyst materials,[3] the uncertain correspondence between models and complex experimental systems presents a considerable challenge.[2,4] As experimental materials are seldom uniform or monodisperse in nature, computations should ideally consider an ensemble of probable model systems to determine which configurations and moieties might contribute most to the macroscopic observables.[5] Clearly, as the number of degrees of freedom grows rapidly, high-throughput tools for screening and analyzing individual configurations and features are highly desired, a hurdle optimally tackled by machine learning (ML)-based approaches.[6]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
