Predicting Power Conversion Efficiency in Donor-Acceptor Pairs for Organic Solar Cells Using Machine Learning Ensemble Models

This study explores the application of machine learning (ML) models, specifically tree-based and ensemble models, for accurately predicting the power conversion efficiency (PCE) of organic solar cells (OSCs). The employed ML models were examined on 319 donor-acceptor (D-A) records containing important physical/chemical descriptors of organic solar cells. This research aims to develop suitable D-A material to achieve high-performing OSC under any given material conditions. To achieve this goal, multiple ML-based regression models, including Fine Tree (FT), Medium Tree (MT), Coarse Tree (CT), Bagged Tree (BGT), and Boosted Tree (BST), were studied on the dataset. These employed models were evaluated using standard performance metrics: mean absolute error (MAE), root mean squared error (RMSE), coefficient of correlation (R), and coefficient of determination (R²). Among the models examined, the BST ensemble outperformed the others, achieving an R² of 88.75%, a lowest MAE of 0.522, and an RMSE of 0.725 for validation; it also achieved an R² of 85.26%, a lowest MAE of 0.549, and an RMSE of 0.734 for testing. These results indicate that ensemble methods (BST) offer robust generalization and superior performance in complex PCE predictions. This work demonstrates that ML can serve as a reliable and time-efficient alternative to traditional computational methods for estimating OSC physical/chemical properties. The outcomes are expected to assist researchers in accelerating the discovery of the photovoltaic cell in the laboratory.本研究探讨了机器学习（ML）模型——尤其是基于树的模型与集成模型——在精准预测有机太阳能电池（OSCs）功率转换效率（PCE）方面的应用。所用数据集涵盖319条给体-受体（D-A）样本，包含有机太阳能电池的关键物理化学描述符。本研究旨在开发适配的D-A材料，以在任意给定材料条件下制备高性能有机太阳能电池。为达成该目标，本研究针对该数据集测试了多种基于机器学习的回归模型，包括细树（FT）、中树（MT）、粗树（CT）、装袋树（BGT）与提升树（BST）。上述模型通过标准性能指标进行评估：平均绝对误差（MAE）、均方根误差（RMSE）、相关系数（R）以及决定系数（R²）。在所测试的模型中，提升树（BST）集成模型表现最优：验证集上的决定系数R²达88.75%，最低平均绝对误差为0.522，均方根误差为0.725；测试集上的决定系数R²达85.26%，最低平均绝对误差为0.549，均方根误差为0.734。上述结果表明，集成方法（BST）在复杂的功率转换效率预测任务中具备出色的泛化能力与卓越性能。本研究证实，机器学习可作为传统计算方法的可靠且高效替代方案，用于估算有机太阳能电池的物理化学属性。本研究成果有望助力科研人员加速实验室中光伏电池的研发进程。