Speakers - 2026

Title: Interpretable CatBoost Models for Rapid Prediction of Perovskite Solar Cell Efficiency

Abstract

Perovskite solar-cell performance depends on coupled decisions across composition, film properties, and fabrication and stack design, making experimental optimization costly and time-intensive. This work presents a lightweight and reproducible machine-learning pipeline to predict power conversion efficiency (PCE) using a compact tabular dataset of 34,742 perovskite device records and 23 descriptors. The input features include encoded identifiers for perovskite deposition procedures and device stack sequences (ETL, HTL, back contact, and substrate), alongside physics- and chemistry-informed variables such as band gap, thickness, 2D/3D dimensional indicators, A-site (FA/MA/Cs) and halide (I/Br/Cl) fractions, and a derived interaction term (bandgap × thickness). An efficient gradient-boosting regression model (CatBoost) is trained and evaluated using cross-validation with MAE, RMSE, and R² metrics to assess generalization performance. To ensure scientific interpretability beyond predictive accuracy, SHAP analysis is employed to identify the dominant efficiency drivers and reveal actionable trends linking processing choices and material parameters to device performance. The proposed workflow provides an accurate, low-compute, and explainable screening framework that supports accelerated data-driven optimization in materials science and nanotechnology.