Numerical Simulation and Machine Learning-Driven Engineering of K ₂ GeI ₆ Perovskite Solar Cells for High-Efficiency and Sustainable Photovoltaics

Perovskite solar cells (PSCs) have garnered significant attention due to their exceptional optoelectronic properties and potential for high power conversion efficiency (PCE). This study presents a comprehensive numerical investigation of lead-free potassium germanium hexaiodide (K 2 GeI 6 )-based double perovskite absorbers using the SCAPS-1D simulation tool. A total of 84 device configurations were explored by varying combinations of electron transport layers (ETLs) and hole transport layers (HTLs). The ITO/Ws 2 /K 2 GeI 6 /CuSbS 2 /Ni structure delivered the highest PCE of ∼28.69%. Further optimization was performed by analyzing the absorber thickness, intrinsic defect density, and the thickness and doping levels of the transport layers. The absorber layer was optimized with a thickness of 0.7 μm, an acceptor density ( N A ) of 1 × 10 18 cm – 3, and a defect density ( N t ) of 1 × 10 15 cm – 3 . Band alignment factors, including conduction and valence band offsets (CBO and VBO), were also examined to understand their role in carrier extraction and recombination. Additionally, the effects of temperature, series resistance, and shunt resistance on device performance were studied. Key photovoltaic parameters─such as J – V response, quantum efficiency, recombination and generation rates, capacitance, and Mott–Schottky behavior─were thoroughly evaluated. Additionally, this study established a rigorous analytical framework with robust visualization metrics to align PCE outputs from SCAPS-1D simulations with machine learning (ML) predictions. An ensemble-based ML pipeline was systematically applied, trained, and validated on high-fidelity SCAPS-1D data to predict key photovoltaic metrics, including PCE, open-circuit voltage ( V OC ), short-circuit current density ( J SC ), and fill factor (FF). Seven algorithms─Linear Regression, Ridge, Lasso, Random Forest, Gradient Boosting, Support Vector Regression (SVR), and Multilayer Perceptron (MLP)─were evaluated. The optimized Random Forest regressor demonstrated excellent predictive accuracy ( R 2 > 0.98 for all targets) and identified absorber N t as the most critical performance driver. With absorber thickness, N t, and temperature set to (0.7, 1.0 × 10 15, 300), the Random Forest model predicted a PCE of 28.04%.