<p>Accurately diagnosing microscopic defect properties from macroscopic J-V characteristics in thin-film photovoltaics remains a critical barrier to advancing solar cell efficiency. While experimental techniques like DLTS and admittance spectroscopy directly probe defect states, their widespread adoption is limited by complexity and cost. This study presents an innovative AI-driven framework, using CIGS solar cells as a representative case to demonstrate a universally applicable approach for correlating routine J-V measurements with underlying defect configurations. A comprehensive, physics-based dataset was generated through high-throughput SCAPS-1D simulations. Systematic variation of six key defect parameters (bulk and interface defects) yielded 4,096 synthetic J-V curves, each linked to specific defect states. This addresses the critical challenge of data scarcity, as obtaining such relevant and diverse experimental data is often practically impossible. A Mixture Density Network with multi-head attention was developed to probabilistically predict defect properties from J-V features, addressing the inherent ambiguity in inverse defect-property mapping. Exceptional accuracy (<i>R</i> ≈ 1) was achieved for critical defects like CIGS bulk vacancies (V_Se​, V_Cu​) and interface traps, while uncertainty was quantified for less discernible defects. The methodology’s novelty extends beyond CIGS, evidenced by advanced non-Gaussian statistical analysis revealing defect-specific performance clusters (e.g., V_OC​ bimodality tied to defect thresholds) and the introduction of innovative visualization tools (e.g., continuous coverage matrices). To foster reproducibility, the dataset and protocols are made publicly available. This work bridges device physics and machine learning to extract microscopic insights from routine solar cell electrical measurements, establishing a transformative paradigm.</p>