Open-Source High-Speed Flight Surrogate Modeling Framework

Abstract

High-speed flight vehicles, which travel much faster than the speed of sound, are crucial for national defense and space exploration. However, accurately predicting their behavior under numerous, varied flight conditions is a challenge and often prohibitively expensive. The proposed approach involves creating smarter, more efficient machine learning models (also known as surrogate models or meta models) that can fuse data generated from a variety of fidelity levels -- to include engineering methods, simulation, wind tunnel, and flight test data -- to make more accurate predictions. These models are able to move the bulk of the computation from high performance computing (HPC) to single user machines (laptop, desktop, etc.). The project builds upon previous work but introduces code improvements and an informed perspective on the direction of the field. The new surrogate modeling framework is now modular and, by design, broadly applicable to many modeling problems. The new framework also has a more robust automatic hyperparameter tuning capability and abstracts away most of the pre- and post-processing tasks. The Gaussian process regression and deep neural network-based models included in the presented framework were able to model two datasets with high accuracy (R^2>0.99). The primary conclusion is that the framework is effective and has been delivered to the Air Force for integration into real-world projects. For future work, significant and immediate investment in continued research is crucial. The author recommends further testing and refining modeling methods that explicitly incorporate physical laws and are robust enough to handle simulation and test data from varying resolutions and sources, including coarse meshes, fine meshes, unstructured meshes, and limited experimental test points.