(659g) Large-Scale Exploration of Perovskites for Oxygen Evolution Via Adaptive Machine Learning

Perovskites with a ABO3 structure have shown great promise for catalyzing the thermodynamically and kinetically unfavorable oxygen evolution reaction (OER) in the (photo)-electrochemical systems. While it is very time consuming and costly to search for highly optimized catalysts by either empirical testing or the quantum-chemical calculations. In recent years, we have seen a dramatic rise of the use of machine-learning techniques for the materials design[1–3]; however, the biggest challenge of a machine-learning approach is the development of robust feature representations for the materials of interest and the accessibility to large training data sets.

In this regard, we introduce a novel electronic-structure descriptor, namely DOS entropy index, for representing the perovskite in the machine-learning models. The DOS entropy index characterizes the dissimilarities of the perovskites’ complex electronic structures according to the pairwise Kullback-Leibler (KL) divergence of the associating atomic orbital-wise projected density of states (PDOS). Based on the selection of reference systems, the DOS entropy index has two generalized forms, i.e., atomic pairwise entropy index and entropy eigen-spectrum. Starting from a limited ~20 perovskites’ DFT-calculated properties and the corresponding OER activities from the literatures, the Bayesian neural network optimized by the DOS entropy index is able to inference above 1000 double perovskites’ OER activity likelihoods and the associating prediction uncertainties. In practice, our belief towards the model prediction accuracy can be recursively improved by validating the predicted materials with high uncertainties in the actual experiments and incorporate the validation information into the training sets. This adaptive learning framework shows great value in the practical applications for the large-scale exploration of perovskites’ measured OER activities using the easily accessible computational descriptors. Furthermore, the quantified uncertainties by the Bayesian rules also allow a potential improvement of the model’s generalizability with a minimum cost of experiment validation efforts.

References

[1] Z. Li, S. Wang, W.S. Chin, L.E. Achenie, H. Xin, High-throughput screening of bimetallic catalysts enabled by machine learning, J. Mater. Chem. A Mater. Energy Sustain. (2017). doi:10.1039/C7TA01812F.

[2] Z. Li, X. Ma, H. Xin, Feature engineering of machine-learning chemisorption models for catalyst design, Catal. Today. 280 (2017) 232–238.

[3] X. Ma, Z. Li, L.E.K. Achenie, H. Xin, Machine-Learning-Augmented Chemisorption Model for CO2 Electroreduction Catalyst Screening, J. Phys. Chem. Lett. 6 (2015) 3528–3533.