Bayesian Parameter Estimation for Predictive Modeling of Illumination-Dependent Current-Voltage Curves

TL;DR

This paper uses Bayesian methods and neural networks to improve the reliability of parameter estimation in solar cell models by validating them against physical JV curve data under different illumination conditions.

Contribution

It introduces a validation approach for neural-network-based parameter estimation using physical JV curve modeling and emphasizes the importance of specific data inputs for accuracy.

Findings

Correct treatment of dark shunt resistance improves predictions.

Including at least one illuminated JV curve enhances parameter reliability.

Emphasizing shifted current during fitting increases accuracy at moderate illumination.

Abstract

Machine learning enables rapid estimation of material parameters in solar cells via neural-network-based surrogate models. However, the reliability of extracted parameters depends on underlying assumptions such as the choice of one-dimensional drift-diffusion model and selection of free material parameters. To validate the inferred parameters, we perform predictive modeling of light-intensity-dependent current-voltage (JV) characteristics. Well-known physical effects, including the influence of external resistance and recombination dynamics on illumination-dependent device performance, are reflected in parameter estimation and prediction workflow. We show that correct treatment of dark shunt resistance and emphasizing shifted current (J + Jsc) during fitting enhances prediction accuracy at low to intermediate illumination level. Additionally, we analyze the information content of various input JV curve combinations, demonstrating that including at least one illuminated JV, preferably not under high illumination due to series resistance effects, is critical for reliable parameter estimation and device performance prediction.