While deep learning models might not be able to simulate large-scale physical phenomena in the same way purpose-built supercomputers and their application stacks do, there is more research emerging that shows how traditional HPC simulations can be augmented, if not replaced in some parts, by neural networks.
An upcoming meeting of the American Physical Society that will focus on fluid dynamics and turbulence will shed light on how and where this happening with a number of presentations focused on how neural nets fit into CFD and other physics-driven simulation areas.
Researchers from Los Alamos National Lab compared three deep learning models, generative adversarial networks, LAT-NET, and LSTM against their own observations about homogeneous, isotropic, and stationary turbulence and found that deep learning, “which do not take into account any physics of turbulence explicitly, are impressively good overall when it comes to qualitative description of important features of turbulence.” Even still, they add that there are some shortcomings that can be addressed by making corrections to the deep learning frameworks through reinforcement of special features of turbulence that the models do not pick out on their own after training.
Lead authors from LANL in the research above break out results from work with generative adversarial networks in a second presentation on turbulent flow. They will present results from a fully-trained GAN that is able to rapidly draw random samples from the full distribution of possible inflow states without needing to solve the Navier-Stokes equations, eliminating the costly process of spinning up inflow turbulence. “This suggests a new paradigm in physics informed machine learning where the turbulence physics can be encoded in either the discriminator or generator.” The LANL team will also propose additional applications such as feature identification and subgrid scale modeling.
Connected to that research, another team from Brown University has proposed a new Navier-Stokes informed neural network that is trained to spot various aspects of fluid motion (velocity, pressure, etc.) as they occur in dye, smoke, or other settings. They design their algorithm to be agnostic to the geometry or initial boundary conditions. Their algorithm achieves “accurate predictions of the pressure and velocity fields in both 2D and 3D flows for several benchmark problems motivated by real-world applications. The findings demonstrate that this relatively simple methodology can be used in physical and biomedical problems to extract valuable quantitative information (e.g. lift and drag forces or wall shear stresses in arteries) for which direct measurement may not be possible.”
A UCLA team used several deep learning models to see where neural networks might fit into aerodynamics applications, specifically how they might be trained to identify incident gusts and rapid changes of wind direction. The four algorithms they work with “achieved satisfactory results in their own tasks, showing the possibility of employing deep learning on a broader scale for gust detection in aerodynamics.”
In a similar vein, work from researchers at New York University and TU IImenau in Germany have used deep learning to detect turbulent superstructures in specific convection settings. The team’s convolutional neural network was able to generate precise image segmentation from a relatively small training set of turbulent convection patterns.
All of the research mentioned here will be presented at the 71st Annual Meeting of the APS Division of Fluid Dynamics in Atlanta in November.