Prediction of unmeasured optoelectronic properties of Au/(GO:P3C4MT)/SiO2 heterojunctions via ensemble machine learning models

Abstract

This study presents a statistically validated machine learning approach for predicting unmeasured optoelectronic responses at an Au/SiO2/n-Si semiconductor interface incorporating a graphene oxide (GO) and poly(3-cyclohexyl-4-methyl-2,5-thiophene) (P3C4MT) hybrid interfacial layer. Although GO-polymer hybrid interfaces exhibit promising photoactive properties, comprehensive experimental characterization of current-light intensity relationships is often constrained by the high cost and time demands associated with high-illumination measurements. The experimental dataset used for model training was obtained by measuring current-voltage characteristics at room temperature, with voltage varied from - 2 to + 2 V in 0.02 V increments under illumination intensities ranging from 20 to 100 mW/cm(2) at 20 mW/cm(2) intervals. Five ensemble and tree-based supervised regression models (Random Forest, Gradient Boosting, LightGBM, HistGradientBoosting, and ExtraTrees) were employed using voltage and light intensity as input features to predict current responses at intermediate and extended illumination levels (10, 30, 50, 70, 90, and 110 mW/cm(2)). Model performance was evaluated using R-2, mean squared error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), with all models achieving high predictive accuracy (R-2 > 0.99). Among them, the ExtraTrees regression model consistently produced the lowest prediction errors across all illumination conditions. To ensure statistical rigor, a two-way ANOVA was conducted to examine the effects of model type and illumination intensity, revealing statistically significant main and interaction effects (p < 0.001). Post hoc Tukey HSD analysis further confirmed the statistically superior performance of the ExtraTrees model. These results demonstrate that reliable interpolation and controlled extrapolation of optoelectronic behavior can be achieved using sparse experimental data, providing a cost-effective, scalable, and statistically robust methodology for the characterization and design of hybrid semiconductor interfaces in photonic, photovoltaic, and photosensor applications.