Machine Learning Guided Device-Level Design for High-Efficiency Tunnel Oxide Passivating Contact Solar Cells

Tunnel oxide passivating contact (TOPCon) solar cells (SCs) are dominant in crystalline silicon photovoltalic (PV) due to superior their passivation and contact properties. Achieving further efficiency improvements remains challenging, requiring comprehensive optimization. This study breaks traditional experience-driven limitations by combining a dual-directional framework (forward prediction and reverse optimization) with numerical simulations for n-type TOPCon SCs. A multidimensional database (13 000 datasets) containing 13 physical and four PV parameters is constructed. Data-driven analysis independently corroborates the well-documented phenomenon whereby excessive tunnel oxide thickness and elevated interface recombination induce significant efficiency degradation in TOPCon SCs. A high-accuracy forward prediction regression model is developed using ensemble learning, and its genetic algorithm-based reverse optimization identifies an optimal combination achieving 28.10% predicted efficiency. SHapley Additive exPlanations (SHAP) analysis reveals governing mechanisms: 1) Front emitter-related parameters most significantly impact efficiency; 2) Low efficiency primarily stems from poor parameter coordination (e.g., low front doping with high defect density); 3) Complex front emitter-rear contact interactions indicate a "front-heavy doping (shallow diffusion) and rear-heavy doping (deep diffusion)" design optimally balances carrier transport and Auger recombination suppression. This provides a scalable data-driven methodology, with revealed coordination principles offering essential insights for future PV design and optimization.