Geometric Parameter Estimations of Perovskite Solar Cells Based on Optical Simulations

TL;DR

The problem addresses non-destructive estimation of perovskite solar cell geometry from optical measurements. The authors generate synthetic EQE data via the transfer matrix method across defined thickness ranges, render EQE as images, and train a CNN to invert EQE to a seven-layer thickness vector. Initial models underperformed, but through Bayesian optimization, mixed pooling, and input/sampling refinements, they reduced the RMSE to $22.54$ nm, with several layers predicted at sub-$2$ nm accuracy; transparent perovskites were used to circumvent light-sensitivity issues. This work presents a scalable, non-invasive characterization workflow that could reduce measurement costs and accelerate PSC design and optimization, while highlighting areas for future improvements such as feature scaling and ensemble approaches.

Abstract

This paper presents a non-invasive approach to estimate the layer thicknesses of perovskite solar cells. The thicknesses are predicted by a convolutional neural network that leverages the external quantum efficiency of a perovskite solar cell. The network is trained in thickness ranges where the optical properties are constant, and these ranges set the constraints for the network's application. Due to light sensitivity issues with opaque perovskites, the convolutional neural network showed better performance with transparent perovskites. To optimize the performance and reduce the root mean square error, we tried different sampling methods, image specifications, and Bayesian optimization for hyperparameter tuning. While sampling methods showed marginal improvement, implementing Bayesian optimization demonstrated high accuracy. Other minor optimization attempts include experimenting with input specifications and pre-processing approaches. The results confirm the feasibility, efficiency, and effectiveness of a convolution neural network for predicting perovskite solar cells' layer thicknesses based on controlled experiments.

Figures (8)

• Figure 1: True vs Prediction of CNN Trained using 1000 images from Halton Sampling. Layers ITO, SnO2, and Perovskite are shown to have good predictions. Spiro-OMeTAD and Au layers, which are the farthest from the incident light, exhibit poor prediction as most light is lost before these last two layers.
• Figure 2: Example of Sobol sampling to sample the required layer thicknesses that are used for generating the training dataset. Here, we sampled for transparent PSCs.
• Figure 3: Example of a training image used for the transparent perovskite's CNN. Blue is forward EQE and red is reverse EQE. Note that the axes are shown here as it was used for earlier trials. The axes are removed for later trials to reduce redundant noise in the image.
• Figure 4: Example of a 2 subplot training image used for the transparent perovskite's CNN. Blue is forward EQE and red is reverse EQE. However, this plot format produced worse results.
• Figure 5: Our first architecture used to prevalidate that the CNN is capable of converging and predicting layer thicknesses