Accurately simulating the interaction between two fluids and their dynamic behavior under complex conditions, particularly in Loss-of-Coolant Accident (LOCA) scenarios, remains a significant challenge in fluid dynamics research. Existing models often struggle to capture rapid variations, such as spikes or discontinuities, that arise in such problems. In this study, a Physics-Informed dual-driven Kolmogorov-Arnold Networks network (PI-KAN) algorithm, which integrates physical laws with observational data within a deep learning framework is introduced. By leveraging a learnable edge activation network, the PI-KAN algorithm substantially improves simulation accuracy, successfully capturing subtle variations in complex fluid interactions. To assess the performance of the PI-KAN algorithm, two representative benchmark problems are selected for testing. The results indicate that the algorithm achieves exceptionally low error metrics across both cases. In Case 1, the MAE, RMSE, and R2 values are 0.0010, 0.0018, and 0.9994, respectively. In Case 2, the MAE, RMSE and R2 values are 0.0015, 0.0039, and 0.9988, respectively. These results demonstrate the PI-KAN algorithm’s strong ability to effectively simulate and solve the two-fluid flow process under LOCA conditions. Overall, the PI-KAN algorithm represents a significant advancement in the simulation and prediction of two-fluid dynamics under LOCA scenarios, offering a robust, accurate, and computationally efficient solution for high-precision modeling in complex fluid systems.