Perovskite solar cells (PSCs) offer great potential in photovoltaics, they can in fact facilitate accelerated energy transition due to their low manufacturing cost and high photoconversion efficiency that is comparable to that of silicon-based conventional solar cells [1-4]. The problem of stability and lifespan continues to be a significant obstacle for this new technology, despite the remarkable recent efforts by researchers to enhance PSC performance utilising non-toxic and non-corrosive ingredients. As a result, some researchers have examined the contribution of various electron transport layers (ETLs), while others have focused on the effects of various hole transport layers (HTLs) to solve these issues [5].

This work proposes a hybrid perovskite solar cell (HPSC) architecture, based on CsSn(I_1-xBr_x)₃ and designed using a deep learning (DL) approach, by replacing the existing hole transport layer (HTL) with a tiny highly doped silicon (Si⁺⁺) layer. Electrical parameters (open-circuit voltage (Voc), short-circuit current (Jsc), fill factor (FF), and power conversion efficiency (PCE)) of this new HPSC, used to train the model, were generated using the SCAPS-1D code [6] due to its high accuracy when benchmarked to experiments. In view of the long simulation runs required, a few combinations of variables such as the bromine composition (x), metal contact work function, absorber layer, and silicon layer thicknesses, were generated and used as input parameters to build the model. Performance metrics of the model based on a Multilayer Perceptron (MLP), like the R-squared coefficient (R²), the root mean squared error (RMSE), the correlation coefficient (r), mean absolute percentage error (MAPE), and accuracy (MAE) were calculated, while the cross-validation (CV) scores were used to check the stability and the overfitting.

The best performing model had R², r, RMSE, MAPE, MAE and CV equal to 0.92, 0.96, 1.58, 0.0063, 0.36, 0.86, respectively, demonstrating the model's good predictive quality. To identify the most suitable device structure for the best performing HPSC, 21 values of each input parameter were created, for a total of 21x21x21x21 = 194,481 different device configurations. Subtracting the actual data used for training, 192,981 potential samples were obtained . The best performing HPSC was obtained by considering the highest efficiency, within the Shockley-Queisser limit. It was found to have an efficiency of roughly 26.92% for 55% bromine fractional composition, 6.2 eV metal contact work function, 1400 nm and 95 nm as absorber and silicon thicknesses, respectively. Results also indicate that the use of highly doped silicon as a hole transport layer (HTL) facilitates the extraction of photogenerated carriers and reduces the series resistance of the HPSC because of its high hole mobility of about 443 cm²/Vs, compared to most existing HTLs.