AI speeds up material discovery for advanced perovskite solar technology

AI speeds up material discovery for advanced perovskite solar technology
by Riko Seibo
Tokyo, Japan (SPX) Jul 30, 2025

A collaborative team from Peking University and its Shenzhen Graduate School has developed machine learning models that can swiftly and precisely predict critical electronic properties of halide perovskites - key materials in next-generation solar cells. Their work aims to streamline the search for optimal compounds by focusing on essential parameters such as conduction band minimum (CBM), valence band maximum (VBM), and bandgap energy.

Halide perovskites, with their ABX3 crystal structure, are promising materials due to their impressive photovoltaic performance, ease of fabrication, and low cost. These materials are highly tunable, allowing researchers to optimize electronic properties to enhance power conversion efficiency (PCE), which has now surpassed 27% in single-junction and over 30% in tandem solar cells. However, persistent challenges - such as lead toxicity and stability issues - necessitate the discovery of improved compositions with ideal band structures.

Precise knowledge of a perovskite's CBM, VBM, and bandgap is fundamental to optimizing device efficiency, as these properties dictate light absorption and charge transport capabilities. Traditional methods for analyzing these factors, like high-throughput screening and density functional theory (DFT) simulations, are reliable but resource-heavy.

To address this, the researchers employed Extreme Gradient Boosting (XGB) to build predictive models capable of estimating band structure features across both inorganic and hybrid halide perovskites. Their XGB model yielded high accuracy, achieving test set R values of 0.8298 for CBM, 0.8481 for VBM, and 0.8008 for bandgap predictions using the Heyd-Scuseria-Ernzerhof (HSE) functional. Using the Perdew-Burke-Ernzerhof (PBE) functional for a broader dataset, the model improved further with an R of 0.9316 and a mean absolute error (MAE) of just 0.102 eV.

In addition, SHAP (SHapley Additive exPlanations) analysis revealed which chemical and structural features most influence electronic energy levels, offering a roadmap for designing better-performing perovskites. This approach not only accelerates the pace of discovery but also provides eco-friendly and cost-effective alternatives to traditional methods.

Looking forward, the researchers aim to integrate the interpretability of shallow machine learning models with the depth of neural networks to further refine materials discovery. Their approach holds significant promise for developing next-generation solar technologies with improved efficiency, stability, and environmental safety.

Research Report:Machine learning for energy band prediction of halide perovskites