As a subfield of artificial intelligence (AI), machine learning (ML) has emerged as a versatile tool in accelerating catalytic materials discovery because of its ability to find complex patterns in high-dimensional and heterogeneous data. While the intricacy of cutting-edge ML models, such as deep learning, is what makes them powerful, it also makes decision-making challenging to explain. Recent advances in explainable AI technologies, which aim to make the inner workings of ML models comprehensible to human, have considerably increased our capacity to derive insights from data. An interpretable ML strategy based on the theory-infused neural network (TinNet) algorithm has recently been developed. The framework can be used to give physically meaningful and reliable explanations of model predictions without losing prediction performance. Moreover, local interpretability for a given catalyst in comparison to a standard material can aid in determining the causal relationships with electronic structure factors, opens a new avenue for ML-driven materials discovery. In this work, we utilized the oxygen reduction reaction (ORR) as an example to show how the TinNet framework can be used to screen for optimal sites and derive physical insights into site reactivity or the catalyst genome from data. One advantage of the TinNet approach compared with other ML techniques is the physical interpretation of model predictions. This is enabled by the regression module built with graph neural networks that extract a feature vector corresponding to the d-band moments of the site atoms and interaction parameters of individual adsorbate orbitals with the metal sp- and d-states. To provide deeper physical insights into adsorption energies at specific structures through these learned features, we adopted a local post-training feature importance analysis, SHAP. Exploiting these local explanations and metrics, we provide deeper insights into the physical factors that control the surface reactivity and it will ultimately help in developing new design strategies for optimizing active sites beyond the d-band center metrics.

Interpretable deep learning reveals the catalyst genome for oxygen reduction

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.