Organic field-effect transistors (OFETs) act as fundamental devices for constructing artificial electronic skins. As to the technology computer-aided design (TCAD) simulation, however, it remains challenging to address the high parameter dimensionality and insufficient generalization validation in physical modeling of OFETs. Therefore, this work proposes a parameter inversion and electrical performance prediction method that integrates the TCAD with machine learning. The methodology establishes the intrinsic parameter space of OFETs through dimensionality reduction and constructs a surrogate model using the random forest algorithm. When applied to simulate flexible OFETs incorporating a polystyrene interface passivation layer, the logarithmic-domain root mean square error values of transfer and output characteristics are controlled at 0.449, 0.920 and 0.506, 1.011 for the non-passivated and passivated device sets under full-data joint calibration, respectively. For electrical performance prediction, relying solely on inversion parameters of output characteristics to predict transfer characteristics, the deviations of threshold voltages and logarithmic on/off current ratios for the thinner passivation layer-based device are only 0.112 V and 0.130 dec, respectively. By contrast, the corresponding deviations increased to 0.152 V and 0.436 dec are observed for the thicker passivation layer-based device. This study indicates that combining data-driven approaches with semiconductor device simulation enables effective parameter inversion and cross-characteristic prediction under limited data conditions. The proposed strategy also provides meaningful insights for quantitative analysis and data validation in flexible electronic systems.