Machine Learning Control of Jet Noise
Abstract
The purpose of this action is to add FY23 funds in the amount of $273,872.57 for a new start Grant. GRANT#13800982.--The use of episode-based machine learning is being investigated for the purpose of discovering new control laws for active open-loop control of jet noise from complex nozzles. This study takes place within a numerical simulation environment that was previously developed, using large-eddy simulation to predict the noise emissions of a supersonic rectangular twin-jet. The simulation has proven to be highly accurate, with predictions matching experimental results to within two dB over a range of polar angles. To build on this, we are proposing a research design that incorporates expert knowledge into the latest generation of model-free machine learning algorithms. The focus is on finding new control laws for reducing or redirecting noise by harnessing new actuation strategies like symmetry-breaking. Two machine learning paradigms, deep reinforcement learning and gradient-enriched genetic programming, have shown efficiency and effectiveness in previous studies of active flow control and will be considered. The implementation of an episode-based learning approach enables the reinforcement-learning or genetic-programming agent to be informed by both simulation and experimental data. If desired, and through close collaboration with our experimental partners, some of the learning data can be provided by the experiment in a hybrid simulation effort. This high-risk, high-reward research project is, to our knowledge, the first known attempt to apply machine learning control to such a complex engineering system.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 15, 2023
- Source ID
- N000142312457
Entities
People
- Oliver Schmidt
Organizations
- Office of Naval Research
- United States Navy
- University of California, San Diego