Materials-informed long-term photovoltaic energy yield prediction: Unveiling the dynamic degradation mechanisms of anti-reflective coatings via explainable machine learning

Accurate long-term prediction of photovoltaic energy yield is critically hampered by the failure of current models to incorporate the dynamic effects of material degradation. This study introduces a novel materials-informed machine learning framework to bridge this gap, explicitly linking the evolving physical properties of anti-reflective coatings to energy yield predictions. To address this, we introduce a materials-informed data-driven framework, and create a unique dataset that couples dynamic material properties with energy yield. A gradient boosting decision tree model, integrated within our framework, demonstrates a profound increase in predictive fidelity when informed by these material features. Evaluated through a rigorous rolling-origin cross-validation, the materials-informed model achieves a mean coefficient of determination of 0.965 and reduces the normalized root mean square error by up to 57.7 % compared to a model using only meteorological data. Explainable artificial intelligence techniques quantitatively decoded distinct aging mechanisms: the performance of hydrophilic coatings is governed by rainfall-driven cleaning (threshold >12 mm), while superhydrophilic coatings depend on maintaining a low water contact angle (<17.2°). Our work establishes a novel, generalizable methodology for integrating material science insights into machine learning, providing a robust scientific basis for coating selection and intelligent maintenance to maximize the lifecycle value of photovoltaic assets.