Machine Learning-Assisted Discovery of Advanced Materials for Next-Generation Photovoltaic Devices

Abstract

The discovery of next-generation photovoltaic materials is essential for advancing sustainable energy technologies, yet conventional experimental and computational approaches face significant limitations in navigating the vast chemical and configurational spaces of multi-component systems. This review provides a comprehensive synthesis of machine learning-assisted materials discovery for emerging photovoltaic technologies, including metal halide perovskites, lead-free double perovskites, two-dimensional photovoltaics, and organic solar cells. We critically evaluate feature engineering methodologies, comparing composition-based vectorization, structural fingerprints, and graph-based atomic representations, alongside classical machine learning algorithms and advanced deep learning architectures such as graph neural networks, angle-aware equivariant networks, and hybrid frameworks. The review examines high-throughput screening pipelines that have successfully reduced candidate spaces from hundreds of thousands to experimentally viable compounds, identifies key predictive descriptors governing photovoltaic performance, and explores generative artificial intelligence models enabling inverse design of novel materials with target optoelectronic properties. We assess open data infrastructures, particularly the Perovskite Database Project, and the integration of large language models for automated literature mining, which collectively establish FAIR data ecosystems essential for data-driven discovery. Critical challenges addressed include publication bias, data degeneracy, quantum mechanical fidelity limitations, and structural discrepancies in training datasets. Finally, we discuss autonomous self-driving laboratories that close the experimental loop through Bayesian optimization and active learning, compressing discovery timelines from decades to weeks. This review establishes a framework for accelerating photovoltaic materials discovery through the synergistic integration of machine learning, high-throughput computation, and automated experimentation.