Machine learning approach for MAPbI3 perovskite solar cells


Thesis Type: Postgraduate

Institution Of The Thesis: Gazi University, Fen Bilimleri Enstitüsü, FOTONİK BİLİMİ VE MÜHENDİSLİĞİ ANABİLİM DALI, Turkey

Approval Date: 2025

Thesis Language: Turkish

Student: SÜHA GÜL KARA

Supervisor: Nihan Akın Sönmez

Open Archive Collection: AVESIS Open Access Collection

Abstract:

In this study, Random Forest (RF), Multilayer Perception (MLP) and Deep Neural Network (DNN) models were used to estimate PCE, Voc, Jsc, FF parameters using multiple output regression learningThe models were trained on the training data and evaluated on the test data using three metrics - R² score, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) to assess their accuracy and reliability.Using the RF model, the R2 value was found to be 0.80-0.54, the RMSE value was 0.05-2.58, and the MAE values were in the range of 0.03-1.76 for the test and training data. For the MLP model, the R2 value for the test and training data was 0.70-0.42, the RMSE value was 0.06-3.18, and the MAE values were in the range of 0.04-2.09. For the test and training data of the DNN model, the R2 value was 0.70-0.36, the RMSE value was 0.06-3.29, and the MAE values were in the range of 0.04-2.17. Parity and residual plots were drawn to determine the performance of the model and whether it is biased or overfitting. The graphs show that these three models do not exhibit a high level of overfitting and bias. The test R2 of the RF model was found to be in the range of 0.75-0.54, RMSE value was found to be in the range of 0.06-2.58, MAE values were found to be in the range of 0.04-1.76. As a result of the analysis of the parity and residual plots, it was observed that the RF model performed better than the other models used. SHAP plots were drawn to obtain feature importance. Multiple output approach can help simplify the complex process of perovskite solar cell (PSC) design.