Fluid simulation is a complex domain that integrates deep learning with traditional methods. This study proposes a novel hybrid fluid simulator that leverages density constraints to model fluid dynamics. By integrating Position-Based Fluids (PBF) with deep learning, our method enhances computational efficiency and ensures stable, physically accurate simulations. Key innovations include: (1) integrating density constraints into neural networks; (2) using spatial hashing to accelerate feature extraction; (3) developing a hybrid simulator that combines PBF stability with neural network efficiency. Experiments show superior performance in visual quality, physical fidelity, and computational efficiency compared to existing methods.