Spatiotemporal Flow Dynamics Prediction by Using
Data-Driven Design Method and Machine Learning

Yu-Hung Chang

Friday, March 3rd, 3:00 - 4:30 PM  |  Guggenheim Building Room 442

ABSTRACT

This is an interdisciplinary work combining machine-learning techniques, statistics, and flow physics. The main purpose and prime contribution of this work are to demonstrate a paradigm of design strategy for new generation engineering via using a swirling injector as example. To develop a robust combustion system efficiently, understanding of underlying physics, coupling and conflicting of the design parameters, and data-driven analysis is significantly important to achieve an efficient design process with the optimal response. Large eddy simulation (LES) has been generally used to simulate flow physics and combustion characteristics inside rocket engines for decades; however, this technique consumes great amount of time and resources, which is impractical to apply for complex design studies. This thesis proposes a design strategy that can conduct a robust design exploration and build a highly efficient model to predict spatiotemporal flowfield with detailed mechanisms of a new design. The whole strategy connects flow physics, statistics, big data analysis, and machine learning.

The applications of Kernel-smoothed Proper Orthogonal Decomposition (KSPOD), Robust PCA (RPCA), and multi-class classification with convolutional neural network (CNN) for spatiotemporal flow mechanism prediction are proposed in this these. The primary results with application of KSPOD, which is trained by high-fidelity simulation datasets, have successfully predicted the instantaneous flow dynamics of a swirl injector by utilizing Kriging based weighting function from design matrix through design of experiment. The implemented machine learning concept enhances the model’s natural applicability in emulation provides an improvement over traditional POD. This model performs well for the analytical estimation of the performance measures such as liquid film thickness and spreading angle. Furthermore, POD and power spectrum densities from POD coefficients from prediction are alike to those from simulation. At the same time, the elapsed computation time for evaluating new design points is reduced significantly compared with other algorithms. The prediction is over 3,000 times faster than simulation.

The future work will focus on the applications of RPCA and multi-class classification with CNN in flow physics to enhance the accuracy of prediction model and design process. To further improve the design strategy with highly efficient emulation method, an enhanced data decomposition method is required. RPCA, an unsupervised learning process, is a modification of POD that works well with respect to grossly corrupted observations. Therefore, mathematically RPCA should work better than POD with turbulence simulation dataset. Multi-class classification with neural network will be used for deeper unsupervised learning with flow structure identification, which can further improve the prediction model providing response with better accuracy. This thesis will eventually conduct machine-learning concepts to probe into flow physics, speed up the design process, and improve the flowfield prediction accuracy.   

Committee members:

1. Dr. Vigor Yang (Advisor)

School of Aerospace Engineering

Georgia Institute of Technology

1. Dr. Lakshmi N Sankar 

School of Aerospace Engineering

Georgia Institute of Technology