Answers are grounded in this paper and cite their sources.
Machine learning has made fitting and parameter estimation in solar cells much faster and easier to use. It therefore becomes more important to validate the inferred parameters before using them, for example, as input to feedback algorithms in high-throughput optimization. However, almost every measurement analysis extracts parameters through a model, and inferred values, including those from neural-network surrogate models, depend on the assumptions built into the forward model and the fitting objective and therefore the correctness of the parameters is almost impossible to establish for experimental data. We thus propose predictive power, the ability of a model and its inferred parameters to reproduce data that were not used for fitting, as a pragmatic and quantifiable alternative to validate the inferred parameters. We demonstrate this criterion using light-intensity-dependent current-voltage ( J V ) characteristics of an organic solar cell (T1:BTP-4F-12) by varying the treatment of dark shunt resistance and the fitting objective (raw J vs. shifted current J + J s c ). The predictions under illumination conditions beyond the fitting data depend on parameter estimation models, which indicates which physical mechanisms are captured by each model. Moreover, we show that the predictive performance is limited by the information content of the input data. A single dark J V curve alone gives broad posterior distributions of the material parameters and less accurate predictions, whereas adding one illuminated J V curve improves the predictions and narrows the posterior and prediction distribution. These results suggest that predictive power is a practical criterion for comparing parameter estimation models, complementary to goodness of fit.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: