Tandem solar cells efficiency prediction and optimization via deep learning.

Area of Science:

• Optoelectronics and Renewable Energy

Background:

• Conventional optical design for opto-electronic devices is computationally intensive and can yield suboptimal results.
• Finite difference time domain (FDTD) and finite element methods are time-consuming and resource-heavy for complex device simulations.

Purpose of the Study:

• To develop a deep learning approach for predicting and optimizing perovskite/crystalline-silicon (c-Si) tandem solar cell performance.
• To accelerate the design process and improve the efficiency of tandem solar cells.

Main Methods:

• A deep neural network was trained on FDTD simulation data to predict short-circuit current.
• Heuristic algorithms were employed for inverse optimization of device layer thicknesses.
• The model was validated against numerical simulation data.

Main Results:

• The deep neural network achieved 98.3% accuracy with micro-second simulation times, drastically reducing computational cost.
• Optimized layer thicknesses led to a projected 14.42% increase in short-circuit current and a 28.4% increase in power conversion efficiency.
• The optimized tandem solar cells are projected to reach 15.79 mA cm⁻² short-circuit current and 23.24% power conversion efficiency.

Conclusions:

• Deep learning offers a highly accurate, ultra-fast, and resource-saving solution for investigating tandem solar cell properties.
• The proposed method successfully optimizes device structure to maximize performance, demonstrating significant improvements over benchmark cells.