Machine learning driven performance for hole transport layer free carbon-based perovskite solar cells

Perovskite solar cells (PSCs) have rapidly advanced in power conversion efficiency from 3.8% (2009) to 26.1% (2024), driven by favorable properties such as tunable bandgaps, long carrier diffusion lengths, and low-cost, solution-based fabrication compatible with scalable coating techniques. Despite progress, large-scale deployment is hindered by stability issues, process variability, and the slow, trial-and-error optimization needed to manage layer morphology, thickness, porosity, and interfaces across different scalable coating methods. Machine learning (ML) offers a data-driven alternative to accelerate optimization and reduce waste by learning relationships between fabrication parameters and device performance. Considering cost reduction for commercialization, HTL-free architectures with carbon electrodes are attractive due to their low cost, abundance, and potential to incorporate HTL-like functionality. This work targets HTL-free carbon-based PSCs (C-PSCs) in planar (n-i-p) configuration and addresses the research question: can ML models, trained on physics-based simulations varying ETL/perovskite thicknesses and bandgaps, accurately predict device performance (Voc, Jsc, FF, PCE) and guide optimal parameter selection to streamline fabrication and improve efficiency?

The paper surveys recent ML applications to PSCs, spanning material discovery, bandgap and performance prediction, and process optimization. Prior efforts include ANN models trained on literature datasets to predict perovskite bandgaps and device outputs, decision trees and random forests applied to regular and inverted architectures for PCE and bandgap predictions, and association rule learning to analyze hysteresis and reproducibility. Other studies demonstrate ML-guided solvent selection for organic solar cells, ML-constrained process optimization for open-air PSC manufacturing, and ML-based screening of interface materials to minimize voltage losses. However, literature-derived datasets often suffer from heterogeneity due to varied fabrication protocols and external conditions, complicating model generalization. The present study addresses these issues by generating a coherent, simulation-derived dataset (SCAPS-1D) specific to HTL-free C-PSCs with controlled parameter ranges and sufficient volume (700 points) to train and evaluate multiple ML models under consistent assumptions.

The authors implemented a five-step approach: (1) identify key C-PSC parameters (inputs: ETL thickness, ETL bandgap, perovskite thickness, perovskite bandgap) and set parameter ranges; (2) generate a labeled dataset via SCAPS-1D simulations; (3) preprocess to remove duplicate and null entries, yielding 700 unique data points; (4) evaluate multiple ML models; (5) select the best-performing model for further analysis and optimization guidance. Device modeling: A planar heterojunction HTL-free C-PSC was simulated with the stack FTO/ETL/Interface Defect Layer (IDL)/CH3NH3PbI3 (perovskite)/Carbon (back electrode). Simulations were performed at 300 K under AM 1.5G illumination. The carbon work function was set to 5.0 eV. An IDL at the ETL/perovskite interface was included to account for recombination effects due to thickness and bandgap variations. SCAPS-1D v3.3.2021 was used, solving carrier continuity and Poisson equations with layer-specific physical parameters (per referenced table in the article). Dataset generation: Input parameter ranges were selected based on reported HTL-free C-PSC performance behaviors. Perovskite thickness: 350–500 nm; perovskite bandgap: 1.5–1.7 eV. ETL thickness: 155–165 nm; ETL bandgap: 3.2–3.6 eV. Outputs (targets): Voc, Jsc, fill factor (FF), and PCE. Initial simulations produced ~790 entries; after removing clones and sparsity-related duplicates, 700 entries remained. Data were split into training (560) and testing (140). ML models and training: Linear Regression (LR), Random Forest (RF), K-Nearest Neighbors (KNN), and Artificial Neural Network (ANN) models were benchmarked. Performance metrics included RMSE and R-squared (R2). The final ANN architecture used three hidden layers with 128, 64, and 32 neurons, respectively; input and output layers each had 4 neurons (corresponding to four inputs and four outputs). Hyperparameters and architecture were tuned to improve prediction accuracy. Model interpretability analyses included Pearson correlation matrices and feature importance/parameter contribution assessments.

- Among the evaluated models, the ANN achieved the best overall performance with RMSE = 0.028 and R2 = 0.954 on test data for the multi-output prediction task (Voc, Jsc, FF, PCE). Other models: LR (RMSE 0.28, R2 0.882), RF (RMSE 0.145, R2 0.917), KNN (RMSE 0.890, R2 0.873). - Prediction quality varied by output: Voc exhibited lower accuracy (RMSE ≈ 0.0047, R2 ≈ 0.7037) relative to Jsc, FF, and PCE, which showed tighter agreement with ground truth. - Correlation insights (Pearson): strong negative correlation between FF and Voc; strong negative correlation between Jsc and perovskite bandgap; cell efficiency showed pronounced dependence on perovskite parameters. ETL thickness and ETL bandgap had comparatively weaker linear correlations with performance metrics. - Feature importance (ANN-based analysis): perovskite bandgap had the largest influence on predictions, followed by perovskite thickness; ETL thickness and ETL bandgap contributed less. - Parametric trends from 4D analyses: higher Voc, Jsc, and PCE were associated with lower ETL thickness and higher perovskite thickness, particularly around a perovskite bandgap near 1.56 eV. FF improved with higher perovskite bandgap and thicker layers, consistent with reduced recombination and longer electron diffusion lengths at higher bandgap, though higher bandgaps tended to reduce overall PCE. - The model identified viable ETL/perovskite parameter combinations that enhance device metrics, enabling reverse-design guidance for HTL-free C-PSC optimization.

The study demonstrates that an ANN trained on a consistent physics-based dataset can accurately learn the nonlinear relationships linking ETL/perovskite thicknesses and bandgaps to HTL-free C-PSC performance metrics. This directly addresses the need to replace time-intensive, trial-and-error optimization across scalable fabrication methods by providing rapid predictions and actionable parameter guidance. The strong model performance (R2 ~0.95 overall) suggests ML can effectively capture key device physics embedded in the SCAPS simulations, while the feature importance and correlation analyses offer interpretable links between material parameters and metrics such as FF and PCE. The relatively lower predictive accuracy for Voc highlights aspects—such as interface quality and complex recombination pathways—not fully captured by the selected features or linear correlations, pointing to opportunities to expand input descriptors. Compared with literature-based models trained on heterogeneous datasets, the present approach benefits from consistency and volume, improving generalizability within the defined parameter space. Practically, the findings can guide parameter selection (e.g., perovskite bandgap near 1.56 eV, thicker perovskite, thinner ETL) to improve C-PSC performance and accelerate process development prior to experimental validation.

The authors developed an ANN-based ML framework trained on 700 SCAPS-1D simulations of HTL-free carbon-based planar PSCs to predict Voc, Jsc, FF, and PCE from ETL and perovskite thicknesses and bandgaps. ANN outperformed LR, RF, and KNN, achieving RMSE ~0.028 and R2 ~0.954 overall. Interpretability analyses revealed the dominant influence of perovskite bandgap and thickness, weaker dependence on ETL parameters, and specific parametric regimes (e.g., perovskite bandgap ~1.56 eV) conducive to improved performance. The model provides rapid, reliable predictions to guide device optimization, reducing reliance on iterative experimentation and supporting commercialization of HTL-free C-PSCs. Future work should integrate additional physical descriptors (e.g., defect densities, interface properties), broaden parameter ranges and materials, and incorporate experimental datasets to enhance robustness, especially for metrics like Voc, thereby refining model accuracy and expanding applicability to real-world fabrication.

- Dataset derived from simulations (SCAPS-1D) rather than experiments; real-world variability (processing conditions, environmental factors, interfaces, defects) may not be fully captured, limiting generalizability. - Lower predictive accuracy for Voc indicates important physical factors (e.g., interface recombination, contact energetics) not fully represented by the chosen inputs. - Parameter space is constrained (specific ranges for ETL/perovskite thickness and bandgap); extrapolation beyond these ranges is uncertain, and some algorithms (e.g., RF) inherently struggle to predict outside training bounds. - Heterogeneity in experimental protocols across literature complicates direct comparison; standardized testing and inclusion of empirical data are needed to validate and calibrate the model. - The carbon electrode and IDL assumptions (e.g., carbon work function, interface defect layer) are idealized; deviations in practice could affect outcomes.