Accelerating device characterization in perovskite solar cells via neural network approach

Abstract

Perovskite solar cells are promising candidates for next-generation high-efficiency photovoltaic devices, especially as top cells in tandem applications. Using a physical-based optoelectronic model, we collect big data of one hundred thousand sample size to train neural network models for efficient prediction of device performance and recombination losses. Latin hypercube sampling, Bayesian regularization, and Bayesian optimization are adopted for data preparation, model training, and optimization of the neural networks, respectively. The best neural network models achieved mean squared errors below $4\times {10}^{-4}$ on a reserved testing dataset. The computational speed of the neural network is more than one thousand times faster than traditional optoelectronic models. As a result, fast device calibration can be conducted in twenty-four seconds. The reduced computational cost allows for efficient device characterization, parametric studies, sensitivity analysis, loss analysis, and optimization. After optimizing interface recombination in our in-house fabricated devices, we observed an experimental improvement of approximately 2 % in power conversion efficiency. Additionally, we predict theoretical power conversion efficiencies of 28.9 % and 25.5 % for perovskite solar cells with band gaps of 1.56 eV and 1.63 eV, respectively.