Accelerated Uncertainty Quantification of Photovoltaic Solar Cell Using Data-Driven Surrogate Modeling: A Comparative Study

Abstract Accurate Uncertainty Quantification (UQ) is essential for the design and performance prediction of photovoltaic devices, where fabrication tolerances can significantly affect electrical characteristics. This study applies data-driven surrogate modeling to accelerate UQ in silicon solar cells, modeled using the single-diode equivalent circuit with series and shunt resistances. Four fabrication-related parameters—series resistance, ideality factor, cell area, and shunt resistance—are treated as random variables with prescribed probability distributions to capture manufacturing variability. Monte Carlo simulation serves as the benchmark for statistical assessment. Four established surrogate techniques—ordinary kriging, radial basis function networks, support vector machines, and ordinary polynomial regression—are evaluated alongside a Modified Polynomial Regression (MPR) method, previously proposed by the first author, which integrates transformation-based feature engineering and optimized parameter selection. Surrogates are trained using Latin hypercube sampling with a limited set of high-fidelity simulations and tested on large-scale Gaussian and uniform samples. While Kriging can achieve slightly higher point accuracy in some cases, its performance varies more across repeated runs. In contrast, MPR delivers competitive accuracy with greater robustness, achieving the highest overall ranking based on signal-to-noise metrics, enabling faster and more reliable UQ in solar energy applications.