基于物理约束注意力增强的拉格朗日流体仿真OA

To address the fundamental conflict between data-driven methods and physical principles in neural network fluid simulation where excessive reliance on learning leads to deviations from conservation laws,while strong constraints impair com-plex flow representation,particularly causing structural distortion in long-term simulations involving boundary interactions,this paper developed a novel MFANet-PD.The model employed a dual-channel cooperative architecture:the FFE dynamically extracted fine-grained motion features using continuous convolution operators and channel attention recalibration;a physics-aware constraint path explicitly handled obstacle boundaries and projects conservation quantities,achieving synergistic optimiza-tion of data-driven and physical principles.Experiments on the Liquid3D dataset demonstrate that,compared to baseline models,MFANet-PD improved short-term prediction accuracy by 14.6％,reduced long-term instability by 18.7％,and enhanced distribution consistency by 11.1％,while exhibiting superior resistance to structural distortion.Ablation studies confirmed a 10.1％reduction in long-sequence error for the full model versus the FFE-free version.Canyon terrain tests validated its ge-neralizability,effectively handling 60° inclined rock walls and dynamic scales from 2 000 to 8 000+particles.The algorithm significantly enhances computational accuracy and physical consistency in fluid simulation through multi-scale feature fusion and physics-constrained mechanisms.