A Multi-Pass GAN for Fluid Flow Super-Resolution

Abstract

Figure 1: A 100 3 simulation (left) is up-sampled with our multi-pass GAN by a factor of 8 to a resolution of 800 3 (right). The generated volume contains more than 500 million cells for every time step of the simulation. In the middle inset, the left box is repeated as zoom-in for both resolutions. ABSTRACTWe propose a novel method to up-sample volumetric functions with generative neural networks using several orthogonal passes. Our method decomposes generative problems on Cartesian field functions into multiple smaller sub-problems that can be learned more efficiently. Specifically, we utilize two separate generative adversarial networks: the first one up-scales slices which are parallel to the XY -plane, whereas the second one refines the whole volume along the Z −axis working on slices in the YZ -plane. In this way, we obtain full coverage for the 3D target function and can leverage spatio-temporal supervision with a set of discriminators. Additionally, we demonstrate that our method can be combined with curriculum learning and progressive growing approaches. We arrive at a first method that can up-sample volumes by a factor of eight along each dimension, i.e., increasing the number of degrees of freedom by 512. Large volumetric up-scaling factors such as this one have previously not been attainable as the required number of weights in the neural networks renders adversarial training runs ://doi.org/10.1145/nnnnnnn. nnnnnnn.prohibitively difficult. We demonstrate the generality of our trained networks with a series of comparisons to previous work, a variety of complex 3D results, and an analysis of the resulting performance.

Unsupervised deep learning for super-resolution reconstruction of turbulence

“…(2018) and Werhahn et al. (2019) applied GANs on super-resolution smoke data. In all these prior studies, researchers used a supervised deep-learning model, which required labelled low- and high-resolution data for training.…”

Dynamic Upsampling of Smoke through Dictionary-based Learning

“…Tompson et al [2017] trained a neural network to predict pressure without solving a Poisson equation, while Umetani and Bickel [2018] proposed to predict aerodynamic forces and velocity/pressure fields from an inflow direction and a 3D shape. A few data-driven approaches directly synthesized flow details for smoke and liquid animations from low-resolution simulations instead: e.g., Chu and Thuerey [2017], Werhahn et al [2019], and Xie et al [2018] created high-frequency smoke details based on neural networks, while Um et al [2018] modeled fine-detail splashes for liquid simulations from existing data. Yet, these recent data-driven upsampling approaches do not generate turbulent smoke flows that are faithful to their physical simulations using similar boundary conditions: The upsampling of a coarse motion often fails to reconstruct visually expected details such as leapfrogging in vortex ring dynamics, even if the coarse motion input is quite similar to exemplars from the training set.…”

“…To compromise between efficiency and visual realism for largescale scenes, the general concept of physics-inspired upsampling of dynamics [Kavan et al 2011] can be leveraged: Low-resolution simulations can be computed first, from which a highly detailed flow is synthesized using fast procedural models that are only loosely related with the underlying fluid dynamics, e.g., noisebased [Bridson et al 2007;Kim et al 2008a] or simplified turbulence models [Pfaff et al 2010;Schechter and Bridson 2008]. Very recently, machine learning has even been proposed as a means to upsample a coarse flow simulation [Chu and Thuerey 2017] (or even a downsampled flow simulation [Werhahn et al 2019;Xie et al 2018]) to obtain finer and more visually pleasing results inferred from a training set of actual simulations. However, while current upsampling methods can certainly add visual complexity to a coarse input, the synthesized high-resolution fluid flow often fails to exhibit the type of structures that the original physical equations are expected to give rise to: The inability to inject physically consistent small-scale vortical structures leads to visual artifacts, making the resulting flow simulations less realistic.…”

Multi-fidelity generative deep learning turbulent flows

“…Third, in this work the input of our model is an inexpensive low-fidelity simulation that provides a coarse yet fairly inaccurate prediction. This contrasts to many works in machine learning for turbulent applications where compressed [19,42] or sub-sampled [56,68] fields of the high-fidelity target are used as the input. Some auto-regressive models, such as the deep neural network (DNN) in [18], are in fact even more dependent on a high-fidelity simulation which is needed to start the time-series prediction.…”

“…The final category we discuss is direct fluid flow prediction where the machine learning model is used to predict the state variables of the fluid flow directly. This includes the use of machine learning to approximate fluid flows for graphical simulations [28,63,69], prediction of steady-state flows [16,57], prediction of oscillating/unsteady flows [3,18,47,48], and the super-resolution, compression or reproduction of various fluid systems [19,42,56,68]. While machine learning has become a popular tool to predict the behavior of fluids, we note that the majority of the test cases considered are focused on simple non-turbulent problems.…”