Deep Fluids: A Generative Network for Parameterized Fluid Simulations

Abstract

This paper presents a novel generative model to synthesize fluid simulations from a set of reduced parameters. A convolutional neural network is trained on a collection of discrete, parameterizable fluid simulation velocity fields. Due to the capability of deep learning architectures to learn representative features of the data, our generative model is able to accurately approximate the training data set, while providing plausible interpolated in‐betweens. The proposed generative model is optimized for fluids by a novel loss function that guarantees divergence‐free velocity fields at all times. In addition, we demonstrate that we can handle complex parameterizations in reduced spaces, and advance simulations in time by integrating in the latent space with a second network. Our method models a wide variety of fluid behaviors, thus enabling applications such as fast construction of simulations, interpolation of fluids with different parameters, time re‐sampling, latent space simulations, and compression of fluid simulation data. Reconstructed velocity fields are generated up to 700× faster than re‐simulating the data with the underlying CPU solver, while achieving compression rates of up to 1300×.

Enabling real-time multi-messenger astrophysics discoveries with deep learning

“…In the same way that numerical relativity simulations of BBH mergers have been critical for the detection and characterization of these sources with GW observations, numerical relativity simulations of BNS and NSBH mergers are critical to get insights into the physical processes that may lead to the production of electromagnetic and astro-particle counterparts, and to better interpret MMA observations 67 . These modeling efforts do not currently benefit from DL, but recent studies have suggested the possibility to improve the efficiency and robustness of simulations, enabling the inclusion of detailed microphysics [68][69][70][71][72] , and a significance increase in the speed with which partial different equations are solved 73,74 .…”

Sketch2VF: Sketch‐based flow design with conditional generative adversarial network

“…To ensure that a user's intention is reflected in the generation process, studies have typically used a painting interaction (e.g., brushing, copy, and paste), which allows users to adjust the output appearance with user drawings while preserving the realism of the generated results . For physics simulation, machine learning approaches have been attracting worldwide attention because these systems can construct plausible results faster than resimulating (or synthesizing) new data . For example, tempoGAN introduced a super‐resolution method to generate an arbitrary high‐resolution fluid result (highly detailed and temporally coherent features) from a low‐resolution input .…”

“…22 For physics simulation, machine learning approaches have been attracting worldwide attention because these systems can construct plausible results faster than resimulating (or synthesizing) new data. 23 For example, tempoGAN introduced a super-resolution method to generate an arbitrary high-resolution fluid result (highly detailed and temporally coherent features) from a low-resolution input. 24 In this study, we utilized an interactive machine learning approach for 2D vector field generation in flow design problems.…”

Sampling and resolution characteristics in reduced order models of shallow water equations: Intrusive vs nonintrusive

“…Recently, machine learning (ML) tools, especially artificial neural networks (ANNs) have been playing a significant role in different fields of science and engineering [88][89][90]. The merits of machine learning algorithms have been demonstrated for prototypical fluid mechanics applications, including flow modeling, reconstruction, and control [91][92][93][94][95]. This is due to their superior capability to identify the linear and non-linear maps between input and output data as well as extracting underlying patterns [96][97][98][99].…”