STELLAR SCIENCE LTD. CO. — Department of Defense STTR Phase I: AF22A-T005

Description

The USAF has identified a need to rapidly predict the safe time between the door being opened on a weapons cavity and the release of the stored device so the store does not rise back into the cavity or exit with an undesirable pitch or yaw. Computational fluid dynamic (CFD) runs for a sample problem are provided as input. This proposed effort uses unique tools and techniques developed at Stellar Science and draws on experience from both Stellar Science and the Computer Science department at the University of New Mexico (UNM) to make novel applications of existing machine learning (ML) techniques. These surrogate models will predict the time, pitch, and yaw for various release times, allowing the prediction of safe times without computationally expensive CFD runs. The models will also directly identify the characteristics of the cavity that result in safe or unsafe releases, supporting further CFD runs and sensor placement in physical cavities. Our proposal directly addresses the two major problems that surrogate models must resolve to accurately model chaotic fluid flow. The first is the large amount of data per sample. Output from CFD runs typically contain a large amount of data, but surrogate models benefit from having less data or only the relevant data. We will address this issue using multiple techniques to remove unneeded information for surrogates, including data categorization, recursive feature elimination, proper orthogonal decomposition, and principal coordinate analysis. The second issue that must be addressed to create good quality surrogates is the number of training samples. While 100 generated runs represent a substantial time investment for CFD runs, this is a very small number of samples for some kinds of surrogates. We will address this issue in two ways. The first will be to use a variety of types of surrogates, some of which rely more heavily on the data reduction techniques described above but require fewer samples to train. The second will be to apply techniques to identify how many samples are needed to get a surrogate with the needed accuracy. This will guide the Phase I effort by indicating whether the supplied data is sufficient and support the Phase II effort by using error estimates from surrogates to determine both the number and the starting conditions for CFD runs. Our Phase I effort will result in a tool chain that can create surrogates for rapid analysis of separation events of all kinds in airflows simulated by CFD codes. A byproduct of the data reduction process will be identifying the data that most likely discriminates between successful or unsuccessful separation. The surrogates can also be used for sensitivity studies to further identify potential discriminators. The process of guiding CFD runs using surrogate error in the Phase II effort will rely on these discriminators to ensure enough data is collected to provide good results while minimizing the number of required CFD runs.