Machine Learning-Based Surrogate Modeling for Aerodynamic Loads Predictions
Abras, Jennifer (HPCMP CREATE)
Intersection of Digital Engineering and High Performance Computing/High End Computing
The rapid prediction of aerodynamic loads for engineering applications is of high interest in the community. One method for obtaining these rapid predictions is through surrogate modeling. Such models enable a faster turnaround time for various engineering needs that high-fidelity computational models cannot accommodate. A Machine Learning (ML) framework to support surrogate modeling of aircraft integrated aerodynamic loads predictions is investigated in this effort. The ML framework includes core Deep Neural Network (DNN) components built to support surrogate models for both steady and unsteady surrogate aerodynamics. A vital aspect of a successful surrogate model is the prediction accuracy when non-linear flow phenomena such as flow separation and transonic effects impact the flowfield. Parameter studies of the impact of different input feature sets on the surrogate model’s predictive capabilities are conducted. A discussion on the potential benefits of incorporating Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) based DNN architectures for the surrogate model is also included.
Initial verification studies were conducted using 2D steady and dynamic stall problems. The surrogate models were trained using HPCMP CREATETM-AV Kestrel CFD simulations. Model building efforts for full 3D aircraft maneuvering are currently underway. This model will be automatically trained to operate within user-defined parameters and will also incorporate an uncertainty assessment of the trained model against user-selected parameters. Further future efforts will investigate adding model calibration using measured data sources such as wind tunnel or flight test data. The goal of the complete capability is to contribute to the general area of digital engineering and enable the development of a trusted surrogate model for rapid/real-time system assessments.
Material presented in this abstract is a product of the CREATE (Computational Research and Engineering for Acquisition Tools and Environments) element of the U.S. Department of Defense HPC Modernization Program. In addition, the authors would like to acknowledge the support of the supercomputing resources provided by the HPCMP, in particular the Army Engineer Research and Development Center (ERDC) and the Air Force Research Laboratory (AFRL).