Simulation tools such as SCAPS (Solar Cell Capacitance Simulator) have enabled significant advances in numerical modeling [2] (Burgelman et al., 2000); however, the literature continues to report considerable discrepancies between simulated and experimental efficiency values, even under similar device architectures [4] (Karthick et al., 2020). Several studies have analyzed these discrepancies, but they focus on qualitative comparisons, such as those presented by [5] (Yadav et al.,2022), who attributed the differences between SCAPS simulations and experimental results to interface recombination and absorber quality without quantifying their statistical impact.

Unlike previous studies where discrepancy analysis relies on a few pairs of simulated and experimental data [1] (Abnavi et al., 2024), present work introduces a systematic statistical analysis based on a single PSC architecture, evaluating 191 combinations by varying the thicknesses of the active layers.

In this work, we conducted a comparative statistical analysis between experimentally reported power conversion efficiency (PCE) values [3] (Jacobsson et al., 2022) and PCE values simulated using SCAPS-1D. The device architecture FTO/TiO₂/MAPbI₃/Spiro-OMeTAD/Au was kept constant across all configurations, while the thicknesses of the ETL, perovskite, and HTL layers were varied according to the experimental dataset. A baseline simulation was manually calibrated to replicate the behavior of a reference device. Based on this architecture, the 191 structures were simulated individually. For each case, the PCE was extracted from the simulated J-V curve and compared with the corresponding experimental value.

The statistical analysis included: Pearson correlation, mean absolute error (MAE), Kolmogorov–Smirnov (K-S) test, probability density functions (PDF), cumulative distribution functions (CDF), and a Taylor diagram. The probability density functions (PDF) were estimated using the Kernel Smoothing method. The results show that, while a statistically significant correlation value, between simulated and experimental PCE, was found (r = 0.41, p < 0.001). We regard this value as unexpectedly low. On the other hand, the MAE reached 3.98%.

The comparison between density distributions revealed that simulated PCE values are concentrated in a narrower range, with a peak around 17%, whereas experimental values are more widely distributed and skewed toward lower efficiencies. The CDF analysis showed that 50% of the simulations achieve PCE values above 16.5%, while only 50% of the experimental devices exceed 13.2%.

The scatter plot showed a broad and dispersed data cloud aligned with a linear trend, suggesting the influence of non-modeled factors such as interfacial defects, recombination, or experimental conditions not represented in the simulation. These observations are consistent with the Pearson correlation coefficient and the K-S test, which confirmed there are statistically significant differences between the two distributions.

These findings highlight the complexity of accurately predicting real device efficiency using numerical simulations. They emphasize the need to refine interfacial modeling and explore new calibration strategies, especially when simulations are intended to serve as data sources for predictive modeling, including approaches based on machine learning.

These findings suggest that improving the predictive accuracy of numerical models could be achieved by indirectly incorporating synthesis-related conditions into the simulation, for example by adjusting sensitive parameters such as defect density, interfacial recombination, or charge carrier mobility. In this context, as a future work, we consider that modeling the synthesis process through machine learning techniques trained on experimental data represents a promising approach to establish correlations between fabrication conditions and device performance.