FluidsNet: End-to-end learning for Lagrangian fluid simulation

Abstract

Over the past few decades, fluid simulation has emerged as an important tool in computer animation. However, traditional physical-based fluid simulation systems are time consuming and requires large computational resources to generate large-scale fluid flows. Intelligent systems provide us a new method of data-driven to accelerate simulations. Previous intelligent system manually crafted feature vectors by using a context-based integral method and resulted in high computational and memory requirements. Unlike the existing techniques, we do not use any manually crafted feature, instead directly operate on Lagrangian fluid simulation data and use machine learned features. This paper presents a novel end-to-end deep learning neural network that can automatically generate a model for fluid animation based-on Lagrangian fluid simulation data. This approach synthesizes velocity fields with irregular Lagrangian data structure using neural network. Every fluid particle is treated independently and identically. We use symmetric functions to capture space structure and interactions among particles and design different network structures to learn various hierarchical features of fluid. We test this method using several data sets and applications in various scenes with different sizes. Our experiments show that the model is able to infer velocity field with realistic details such as splashes. In addition, compared with exiting simulation system, this method shows significant speed-ups, especially on large scene simulations.