110th NIA CFD Seminar: Integrated Field Inversion and Machine Learning With Embedded Neural Network Training
Speaker: Jon Holland, PhD Candidate, University of Maryland – College Park
Date: Tuesday, March 19, 2019
Time: 11am-noon (EST)
Room: NIA, Rm137
Abstract: A new approach is presented towards the end of developing data-augmented models. The goal is to effectively reduce model form errors in a Reynolds Averaged Navier-Stokes setting. Since the information required for model improvement is not directly available in higher fidelity simulation or experimental data, model augmentations have to be extracted from the data using the solution of inverse problems. Existing data-driven turbulence modeling approaches either ignore the inference step – in which case, learning is applied directly on the data – or separate the inference and learning steps. In the proposed approach, the learning step is integrated into the field inversion process. This integrated approach ensures that the process generates learnable model discrepancy, and thus results in a consistent machine learned model that can be embedded in a predictive setting. Additionally, a new layered approach is proposed to train neural networks, demonstrating the promise to greatly reduce training time. In this presentation I will present two methods of integrated learning during field inversion, and will discuss recent results on both a simple model problem and RANS applications using the Stanford University Unstructured (SU2) code.
Speaker Bio: Jon Holland completed his undergraduate work at the University of Notre Dame, and is currently an Aerospace Engineering Ph.D. Candidate at the University of Maryland, College Park. With advisors Dr. James Baeder at University of Maryland and Dr. Karthik Duraisamy at University of Michigan Jon has been studying methods for augmenting physics based simulations with data.