Enhancing Tandem Solar Cell's efficiency through convolutional neural network-based optimization of metasurfaces

Abstract

Tandem Solar cells are composed of multiple layers with varying bandgap materials to facilitate the absorption of a wider range of solar energy. However, their performance is hampered by interface losses leading to current mismatching. To overcome these limitations, tandem solar cells often require the use of mirrors and lenses to concentrate and trap light to achieve optimal efficiency. Recently, the use of nanoscale 2D meta-materials has emerged as a promising alternative to traditional lenses. However, the optical optimization of these meta-materials can be both time-consuming, resources, and space-intensive, posing a challenge to their implementation. This study explores the use of deep learning to predict the optimal optical design for the top cell in tandem solar cells to maximize power conversion efficiency. Computational techniques are used to analyze the optical responses of metasurfaces. 2D-Convolutional Neural Networks (CNN) are used to train a dataset of 10,578 TiO2/CH3NH3PBr3/ZnO metasurfaces. Nine different CNN models were used with different architectures to identify the best hyperparameters that give the low mean square error. CNN displayed fast high prediction accuracy taking an average of 0.3 ± 0.05 seconds per prediction. The Deep SHapley Additive Explanations (SHAP) algorithm was used to gain insights into CNN's predictions and understand the behavior of complex metasurfaces integrated into a typical reference tandem solar cell architecture. The proposed metasurfaces can significantly enhance the efficiency of tandem solar cells. The active layer comprises of near 90% absorption of the solar spectrum. The average absorption of the top cell increased in the UV-vis region (650-800nm) up to 93.4%.

Keywords

• Metasurfaces design
• Tandem Solar Cells
• Deep learning
• Convolutional neural networks
• Numerical modeling
• Simulation