上海科技大学知识管理系统

Transient fluid simulation is of crucial importance for both visual effects and engineering design. Achieving the steady state of transient fluid provides valuable physical insights and information about the simulated scenario, which is desirable in many applications. However, initializing the fluid flow solver to quickly reach a dynamic statistical equilibrium state for further simulation and processing is challenging. Typically, initializing a flow field based on domain geometry is computationally expensive, leading practitioners to use simple initializations like a constant velocity field, which hampers efficient solver convergence. In this paper, we propose the first machine learning approach for efficient initial fluid field prediction that could be used to accelerate any fluid flow solver. Our method features a novel neural network model that integrates dense local and sparse global information using a U-shaped convolutional neural network with attention blocks. This architecture enhances predictability even with simple fluid domain geometries. We evaluate our learning-based initialization approach across various cases, including geometries not present in the training set, demonstrating superior convergence to the desired state. This capability is beneficial for both smoke generation in special effects and drag evaluation in engineering applications. Our results show up to 4-fold speedup compared to the widely-used constant velocity field initialization. Additionally, our method is effective for predicting both mean and turbulent field initializations, leading to wider potential future applications. Various smoke animation results and a drag evaluation example are presented to illustrate the practical applicability of our approach.