Deep learning-driven anomaly detection and feature discovery in Ce-rich (Ni–Fe–Co–Ce)Ox catalysts for oxygen evolution reaction

Abstract

Developing high-performance oxygen evolution reaction (OER) catalysts is vital for energy conversion. However, extracting rare, exceptional materials from massive high-throughput experimental datasets remains challenging, as conventional machine learning models often misclassify these optimal data points as noise. Here, we propose a deep learning-driven anomaly detection framework that overcomes this limitation. By integrating atomic-level descriptors with convolutional neural networks (CNNs) for similarity stabilization analysis, our model employs an iterative data-cleaning mechanism to automatically isolate and evaluate high-performing outliers. We validated this approach on a high-throughput (Ni–Fe–Co–Ce)O_x catalyst dataset. The model achieved a robust R² score of 0.90 for inlier predictions while successfully capturing a specific optimal compositional window of Ce-rich compositions (0.3–0.6 at%) exhibiting remarkably low overpotentials. This framework offers a reliable, data-driven analytical tool, demonstrating the power of deep learning anomaly detection in accelerating the discovery and optimization of novel materials.