Learning Optimal Aerodynamic Designs
Transcript of Learning Optimal Aerodynamic Designs
Learning Optimal Aerodynamic Designs
The University of Texas at Austin Oden Institute for
Computational Engineering & SciencesOmar Ghattas, Karen Willcox,
Anirban Chaudhuri, Tom O’Leary-Roseberry
University of Michigan MDO Lab, Dept of Aerospace Engineering
Joaquim Martins, Xiaosong Du
ARPA-E DIFFERENTIATE Program
CFD-based aerodynamic design optimization
● CFD-based design optimization is a powerful technology that has revolutionized the efficient design of aerodynamic systems, including aircraft, ground vehicles, marine vessels, and energy generation via wind, water, and gas turbines
● CFD-based design optimization is challenging due to the need for: ○ Days of computing time to do a single 3D design optimization ○ HPC resources○ Sophisticated algorithms for adjoints, differentiable and robust
shape parameterization and mesh generation, PDE-constrained optimization, robust flow solvers
Our solution: Deep learning of CFD design optimization
● Use deep neural network to learn aerodynamic design optimization
● Specifically, learn the map from design requirements to optimal aerodynamic design with high accuracy (>95%)
● The neural network is executed at interactive speeds (millisecond)● Fast decision-making for tradespace exploration: Conceptual
design phase greatly accelerated● Supplement expert knowledge: No need to set up and run an
expensive CFD optimization every time you change performance requirements
Train DNN with MACH-Aero CFD design optimization framework
Developed at University of Michigan MDO Lab (Joaquim Martins, PI)
MACH-Aero framework
Example: design optimization of CRM wing benchmark
Wave drag is eliminated; total drag reduced by 8.5%
AeroLearn: A data-parsimonious deep neural network trained by MACH-Aero
● Train a DNN to learn the map… ○ From design requirements (lift & moment bounds, geometric constraints such as
thickness, environmental parameters such as Mach & Re, across flight envelope) ...○ … To optimal shape
● This results in a neural network with high-dimensional inputs & outputs (100s to 1000s)● Training a predictive network with black-box ML techniques would require 104--106 CFD
optimizations -- completely intractable! ● Our solution: AeroLearn
○ Sensitivity-based projections to find low-dimensional manifolds for inputs & outputs○ Multifidelity neural networks by enriching expensive high-fidelity data with cheaper
low-fidelity data● Result: Neural network that has high generalization accuracy (99.8% for drag, 99% for
geometry) for few training data (100s), executing at interactive speeds (milliseconds)
AeroLearn vs. CFD-based design optimization runtime:2D airfoil case with 4 inputs and 20 outputs
Mesh MACH-Aero design optimization time AeroLearn design optimization time
medium 1555 seconds 0.0005 seconds
fine 7732 seconds 0.0005 seconds
Design requirement inputs
Optimal design variable outputs
Generalization error vs. number of training data
99% accuracy in optimal shape, 99.8% accuracy in objective function (drag) with just ~100 training data
Rapid exploration of design requirement space
Summary and invitation to partner with us● CFD-based aerodynamic design optimization is a powerful technology, but is
often inaccessible to many who could benefit from it, due to the need for HPC resources, sophisticated algorithms and solvers, and advanced expertise.
● We have developed AeroLearn -- parsimonious deep neural networks that learn aerodynamic design by training on MACH-Aero optimizations
● AeroLearn networks are capable of very high accuracy in predicting the optimal shape and objective function with a limited number of training data
● Once trained, AeroLearn networks can explore the design requirement space to generate optimal designs at interactive (< millisecond) speeds
● We are looking for partners to commercialize this technology● Please stop by our booth and chat with us, or else email:
● Omar Ghattas ([email protected])