[Paper Review] Machine Learning Design of Perovskite Catalytic Properties

Ryan Jacobs, Jian Liu+2 more|arXiv (Cornell University)|Nov 2, 2023

Machine Learning in Materials Science4 citations

TL;DR

This paper presents a machine learning framework that predicts key perovskite catalytic properties—oxygen surface exchange, diffusivity, and area specific resistance (ASR)—using simple elemental features, achieving higher accuracy and speed than ab initio-based models. The model enables temperature-dependent ASR predictions with calibrated uncertainty and identifies over 19 million promising, earth-abundant perovskites, including those with underexplored elements like K, Bi, Y, Ni, and Cu.

ABSTRACT

Discovering new materials that efficiently catalyze the oxygen reduction and evolution reactions is critical for facilitating the widespread adoption of solid oxide fuel cell and electrolyzer (SOFC/SOEC) technologies. Here, we develop machine learning (ML) models to predict perovskite catalytic properties critical for SOFC/SOEC applications, including oxygen surface exchange, oxygen diffusivity, and area specific resistance (ASR). The models are based on trivial-to-calculate elemental features and are more accurate and dramatically faster than the best models based on ab initio-derived features, potentially eliminating the need for ab initio calculations in descriptor-based screening. Our model of ASR enables temperature-dependent predictions, has well calibrated uncertainty estimates and online accessibility. Use of temporal cross-validation reveals our model to be effective at discovering new promising materials prior to their initial discovery, demonstrating our model can make meaningful predictions. Using the SHapley Additive ExPlanations (SHAP) approach, we provide detailed discussion of different approaches of model featurization for ML property prediction. Finally, we use our model to screen more than 19 million perovskites to develop a list of promising cheap, earth-abundant, stable, and high performing materials, and find some top materials contain mixtures of less-explored elements (e.g., K, Bi, Y, Ni, Cu) worth exploring in more detail.

Motivation & Objective

• To accelerate the discovery of efficient perovskite catalysts for solid oxide fuel cells and electrolyzers (SOFC/SOEC).
• To develop machine learning models that predict catalytic properties using only elemental features, avoiding costly ab initio calculations.
• To enable temperature-dependent ASR predictions with well-calibrated uncertainty estimates.
• To identify promising, stable, and earth-abundant perovskites through large-scale screening.
• To provide interpretability of feature importance using SHAP analysis for model featurization.

Proposed method

• The authors train supervised machine learning models on a dataset of perovskite materials with target properties: oxygen surface exchange, oxygen diffusivity, and area specific resistance (ASR).
• Features are based solely on elemental properties (e.g., atomic number, electronegativity, ionization energy), making them trivial to compute.
• The models are trained using gradient-boosted decision trees and validated via temporal cross-validation to simulate prospective discovery.
• A separate model for ASR incorporates temperature dependence and provides uncertainty estimates using Bayesian optimization.
• SHapley Additive ExPlanations (SHAP) are used to interpret feature contributions and compare different featurization strategies.
• A high-throughput screening of over 19 million perovskites is performed using the trained models to identify top candidates.

Experimental results

Research questions

• RQ1 Can machine learning models based on elemental features outperform ab initio-derived descriptors in predicting perovskite catalytic properties?
• RQ2 Can the model make reliable, temperature-dependent predictions of ASR with well-calibrated uncertainty?
• RQ3 Does temporal cross-validation demonstrate the model’s predictive power for future material discovery?
• RQ4 Which elemental features and featurization strategies are most informative for catalytic performance prediction?
• RQ5 What are the most promising, stable, and earth-abundant perovskite candidates for SOFC/SOEC applications?

Key findings

• The ML model using elemental features achieves higher accuracy and significantly faster inference than models based on ab initio-derived features.
• The ASR prediction model provides reliable, temperature-dependent outputs with well-calibrated uncertainty estimates.
• Temporal cross-validation confirms the model’s ability to identify promising materials before their experimental discovery.
• Over 19 million perovskites were screened, yielding a list of high-performing, stable, and earth-abundant candidates.
• Top-performing materials include combinations of less-explored elements such as K, Bi, Y, Ni, and Cu, suggesting new directions for experimental exploration.