Improving RANS for 3D Flows Using Machine Learning and Model Interpretation
Abstract
Computational fluid dynamics codes based on the Reynolds-averaged Navier Stokes equations(RANS codes) will remain the workhorse for analysis, design, and optimization of complex flow systems including naval vehicles and energy systems for the foreseeable future. The great majority of naval systems are dominated by turbulent flows and RANS codes have poor predictive accuracy when the turbulence is out of equilibrium, especially when flows are strongly3D or include regions of flow separation. The goals of this research are to develop and apply new techniques based on machine-learning to accurately predict the Reynolds stresses in complex flows using only knowledge of the mean flow field available in a standard RANS solver. The working hypothesis is that an accurate Reynolds stress distribution will produce an accurate mean-flow prediction. Deep Neural Networks (DNNs) will be developed toindependently predict the turbulent kinetic energy and the Reynolds stress anisotropy, the latter using the Tensor Basis Neural Net (TBNN) approach, which guarantees appropriate invariance properties of the Reynolds stress tensor. The DNN~s will be trained using high-fidelity, eddyresolvingsimulations for building block flows containing isolated regions of strongly nonequilibrium flow. The training simulations will be validated against full field mean velocity measurements made with closely matched boundary conditions. The DNN~s will be implemented within the iteration loop of a standard RANS code, then the approach will be validated by comparison to experimental data for both simple and complex flows. The second aspect of the research involves development and application of new techniques for interpreting machine-learned turbulence models with two specific goals. The first goal is to useinterpretation techniques to determine when a model is being used outside of its trained domain. Ideally the RANS code would revert to a standard model when it encountered input parameters that would force the model to extrapolate. Secondly, model interpretation will be used to develop reduced order models replicating the predictive performance of the full TBNN. Successful completion of this research should provide robust models that run efficiently and canidentify when they are being used inappropriately.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 24, 2019
- Source ID
- N000141912075
Entities
People
- John K. Eaton
Organizations
- Office of Naval Research
- Stanford University
- United States Navy