Physics-Preserved Neural Differentiable Computing for Predictive Modeling of Rough-Wall Turbulence
Abstract
,Almost all boundary layer flows involved in naval systems can be categorized as rough-wallboundedturbulence due to typical high-Reyn,olds-number operating environments, and surface roughness can significantly affect hydro/aeroacoustic behaviors, as well as hydrodyn,amic and thermodynamic performance (e.g., drag, momentum transport, and turbulence mixing) of naval vehicles, propulsors, and other,marine systems. Therefore, to understand, predict, and control the effects of surface roughness on wall-bounded turbulence is of fun,damental and practical interest in the naval industry and engineering. Existing predictive methods based on traditional computationa,l fluid dynamics (CFD) techniques have substantial limitations: first, due to a lack of full understanding of the underlying flow ph,ysics, a complete set of equation forms that accurately and efficiently model turbulent flows over arbitrary complex roughness surfa,ces is not available; second, most existing computational models suffer from high computational costs, particularly when fully resol,ving the roughness, limiting their applications to toy problems and ideal conditions (e.g.,equilibrium flow, simple geometries, etc.,or real-time control and design optimizations, which require massive model queries. Therefore, the objective of this project is,to establish the next-generation AI-assisted, physicsbasedcomputational modeling and simulation framework, enabling efficient and re,liable predictions of turbulent flows over complex rough surfaces, towards real-world naval applications. To achieve this goal, we p,ropose to integrate physics knowledge, multi-resolution data, multi-fidelity models, numerical techniques, and geometric deep learni,ng within a novel neural differentiable programming framework, which can leverage GPU parallelism and is expected to be massively sc,alable for forward and inverse predictive modeling of complex physical systems. In particular, we will develop a graph-based neural,differentiable solver by preserving the known portion of governing PDEs as graph convolutional residual connections, where the unkno,ybrid neural solver. Combined with sparse learning, the learned model can be interpreted as analytical expressions for explicit mode,l-form discovery. A novel graph auto-encoding and message-passing scheme will be developed for effective dimension reduction and rec,overy of large-scale turbulence data across different scales. Moreover, the proposed method will be formulated in a scalable variati,onal Bayesian manner, allowing forward and inverse propagation of predictive uncertainties from multiple sources. We anticipate t,he following outcomes from this project: (1) fundamental theory and algorithms of the proposed neural differentiable model for spati,otemporal predictions of rough-wall turbulence;(2) a suite of software, including a neural differentiable solver for rough-wall turb,ulence, model-form discovery module, turbulence data compression and recovery module, which will be released to the public for free;, (3) discovered model forms and physics knowledge (e.g., turbulence closures, wall models, and roughness functional correlations) fo,r a variety of roughness topologies and configurations. The success of this project will result in revolutionary improvements in the, US Navy?s predictive modeling capabilities in hydro/aerodynamics, as well as advances in the knowledgebase of rough-wall flow physi,cs, allowing optimization design of future ships and submarines with high efficiency, maneuverability, and low noise in real-world o,perating scenarios.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 06, 2022
- Source ID
- N000142312071
Entities
People
- Jian-Xun Wang
Organizations
- Office of Naval Research
- United States Navy
- University of Notre Dame