Perovskite solar cells (PSCs) have emerged as a promising next-generation photovoltaic technology, exhibiting rapid efficiency growth from 3.8% in 2009 to 26.1% in 2024. Their unique characteristics, such as flexibility, tunable bandgap, and low-cost fabrication, attract significant interest. While solution processing offers advantages over silicon's high-temperature requirements, large-scale PSC manufacturing faces challenges related to stability and consistent, automated processes. Scalable technologies like slot-die and spray coating show promise, but optimization often relies on time-consuming trial-and-error methods. This research explores machine learning (ML) as a data-driven approach to accelerate optimization, reducing material waste and development time. The increasing use of ML in PSC research, ranging from material identification to PCE prediction, underscores its potential benefits. This work focuses on HTL-free C-PSCs due to their lower production cost and the use of abundant, cost-effective carbon as a counter electrode. The goal is to develop an ML model that predicts C-PSC performance based on ETL and perovskite layer properties, facilitating efficient optimization and potentially accelerating commercialization.

The existing literature demonstrates a growing interest in integrating machine learning (ML) techniques into the development and optimization of perovskite solar cells (PSCs). Several studies have used ML models to predict various parameters, including perovskite bandgaps and overall cell performance, often relying on datasets from previously published papers or experimental data. However, these studies face limitations due to variations in experimental conditions and a lack of standardized testing protocols. The availability and quality of datasets significantly impact the accuracy and reliability of ML predictions. Some studies leverage SCAPS simulations to generate datasets, but there is still a need for larger and more comprehensive datasets to improve the accuracy of ML-based predictions. Prior studies also explore different ML techniques such as ANN, Random Forest, and others, with varying degrees of success depending on the dataset and specific application. The study by Liu et al. demonstrated a machine-learning framework for optimizing PSC manufacturing and achieving a power conversion efficiency (PCE) of 18.5%. Salah et al. provided a comprehensive analysis of various ML model integrations for PSC fabrication. The current research intends to build upon these prior efforts by generating a significantly larger and more controlled dataset using SCAPS simulations, enabling more robust and accurate prediction of HTL-free C-PSC performance.

This study employs a five-step methodology to create an ANN-based ML model for predicting the performance of HTL-free C-PSCs. The process begins with parameter identification, focusing on ETL thickness, ETL bandgap, perovskite thickness, and perovskite bandgap. These parameters define the range for dataset creation using SCAPS-1D simulation software. A planar heterojunction HTL-free C-PSC structure was used in SCAPS-1D, consisting of FTO/ETL/IDL/perovskite (CH3NH3PbI3)/Carbon. The SCAPS-1D simulation utilizes the continuity and Poisson's equations to optimize the model, considering physical parameters for each layer. The dataset generation involved varying perovskite thickness (350-500 nm) and bandgap (1.5-1.7 eV), and ETL thickness (155-165 nm) and bandgap (3.2-3.6 eV). The generated dataset initially contained approximately 790 data points, which underwent cleaning to remove duplicates and null values, resulting in a final dataset of 700 data points. This dataset was then split into training and testing sets (560 and 140 data points respectively). Various ML algorithms were tested, including Linear Regression (LR), Random Forest (RF), K-Nearest Neighbors (KNN), and Artificial Neural Networks (ANN). The models' performance was evaluated using RMSE and R-squared values. The ANN model demonstrated the best performance, exhibiting an RMSE of 0.028 and an R-squared value of 0.954. Further analysis was performed exclusively using the ANN model. Correlation analysis using the Pearson coefficient was conducted to identify relationships between input parameters and C-PSC performance.

The ANN-based ML model demonstrated high accuracy in predicting C-PSC performance parameters with a low root mean square error (RMSE) of 0.028 and a high R-squared value of 0.954. This significantly outperformed other ML models tested (LR, RF, KNN). The correlation matrix analysis revealed that the perovskite layer's properties (thickness and bandgap) had the strongest influence on the C-PSC's performance parameters, including Voc, Jsc, FF, and PCE. Specifically, a perovskite bandgap around 1.56 eV showed the most promising results. While the ETL properties played a less significant role, the model's predictions identified optimal combinations of perovskite and ETL layer parameters for maximizing device performance. The 4D scatter plots generated from the ANN model predictions visually represented the interplay between the different parameters and their impact on the overall performance. The model demonstrated its capacity to swiftly and accurately predict device performance, offering significant time and cost savings compared to traditional trial-and-error optimization methods. The analysis also showed that perovskite bandgap has the most influence on the output, followed by the perovskite thickness and then ETL layer characteristics.

The findings of this study demonstrate the efficacy of using an ANN-based machine learning model to optimize the design and fabrication of HTL-free C-PSCs. The high accuracy of the model, as evidenced by the low RMSE and high R-squared values, suggests that it can effectively predict C-PSC performance based on variations in ETL and perovskite layer thicknesses and bandgaps. The identification of optimal parameter combinations for maximizing performance is a significant contribution. This approach significantly reduces the time and cost associated with experimental trials and allows for a more efficient exploration of the design space. The interpretability of the model, as shown in the correlation analysis and 4D scatter plots, provides valuable insights into the relationship between materials properties and device performance. The observation that perovskite properties have a stronger influence on the output than ETL properties offers guidance for material selection and optimization strategies. While the model is based on simulation data, it provides a crucial stepping-stone for guiding experimental efforts and accelerating the development of high-performance HTL-free C-PSCs. Future work should focus on validating the model with experimental data and expanding the dataset to include other relevant parameters, further enhancing its predictive capabilities.

This research successfully developed an ANN-based machine learning model for predicting the performance of HTL-free carbon-based perovskite solar cells. The model achieved high accuracy, significantly reducing the time and resources needed for optimization compared to traditional methods. The model's insights into the relationships between layer properties and cell performance offer valuable guidance for experimental design. Future research should integrate experimental data to validate and refine the model, potentially incorporating additional parameters and exploring other ML algorithms for further improvements.

The model's predictions are currently based solely on SCAPS-1D simulations. While this provides a controlled environment for generating a large dataset, the simulation may not fully capture all complexities of real-world fabrication processes. Experimental validation is crucial for confirming the model's accuracy and generalizability. The dataset, while extensive, might still not encompass the full range of possible material combinations and fabrication conditions. Expanding the dataset to include more variations would improve the model’s robustness and applicability.