The Development of a Data-Driven Model Calibration Method for Plasma Physics Applications
Abstract
The development and subsequent studies regarding statistical convergence for a data-driven calibration approach to physics-based models are presented with the ultimate goal of using the resulting method to analytically quantify anomalous behavior seen in experimental data of Hall effect thrusters. This calibration approach uses a single output signal to calibrate unknown input parameters of a computational model to a reference solution, either trusted analytical results or experimental data.
The dimension of the output signal is increased by taking a time-delay embedding, based on the Takens Embedding Theorem, and the resulting time-lag phase portrait is binned as a probability distribution function. The first Wasserstein metric is used to quantify the difference between two solutions as a single variable. This process is automated using an evolutionary algorithm function from Sandia National Laboratory’s DAKOTA algorithm. The canonical chaotic Lorenz attractor, a zero-dimensional bulk plasma model, and a two-dimensional Hall effect thruster model are used to characterize and minimize the numerical uncertainties incurred by this model calibration method and give conditions for the definition of an optimal solution. Results indicate verification of the method’s ability to uncover unknown input parameter values. In particular, the model calibration method is shown to obtain results within 1% of the reference solution for various signals that were not used during the calibration process. Additionally, a more active, online, calibration technique is developed in conjunction with this thesis to detail the first step in the development of a more robust method in future work
Citation
Greve, Christine Marie (2019). The Development of a Data-Driven Model Calibration Method for Plasma Physics Applications. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /188739.