Ph.D. Defense

 Xiao (Olin) Wei

(Faculty Advisor: Dimitri Mavris)

A Surrogate-Assisted Online Adaptive Reinforcement Learning and Approximate Bayesian Computation (OARL-ABC) Method for Calibration of Digital Twins

Friday, October 31

 8:00 am

Collaborative Visualization Environment (CoVE) Weber SST II

Abstract: 

Digital twin technology has become a cornerstone in modern industry and research, providing virtual replicas of physical systems enabling real-time condition monitoring, future state simulation, and design optimization capabilities over a product's entire life-cycle. The accuracy and reliability of digital twins heavily depends on the calibration process, which aligns the digital model with real-world data. As the physical twin evolves over time, the digital twin must be properly recalibrated to remain an accurate representation of the connected physical system. Due to recent trends in digital twin applications requiring more complex model forms and more demanding recalibration time-frames, improving the efficiency and accuracy of this calibration process is critical for enhancing the performance and applicability of digital twins across various domains.

This thesis seeks to contribute to the enhancement of digital twin calibration by developing a new surrogate model assisted Online Adaptive Reinforcement Learning and Approximate Bayesian Computation (OARL-ABC) method for the calibration and validation of digital twins. This proposed calibration approach performs model selection via reinforcement learning techniques and parameter calibration via conventional, tested Bayesian inference methods, utilizing a hybrid Bayesian reinforcement learning calibration framework that combines the adaptability and efficiency of reinforcement learning for optimizing complex, dynamic systems with the uncertainty quantification and updating capability of Bayesian inference methods. This combination leverages the best attributes of both techniques, integrating the principled uncertainty handling of Bayesian inference with the adaptive learning capabilities of reinforcement learning, making it particularly suitable for the calibration of highly complex digital twins.

To further enhance the efficiency of Bayesian reinforcement learning, this thesis integrates the use of surrogate modeling to provide computationally efficient approximations of more complex models for more rapid evaluation in the learning process. This surrogate-assisted OARL-ABC method aims to reduce the computational intensity of the sampling requirements in the original method by utilizing multiple surrogate model representations of a real system. Among various surrogate modeling techniques investigated, Bayesian network models were identified as the optimal choice in this context, as these models maintain the Bayesian framework's ability to manage uncertainty while significantly reducing the computational demands, accelerating the calibration process without compromising accuracy.

By incorporating Bayesian network models, the OARL-ABC method is expected to perform with increased adaptability to different model forms, increased efficiency in complex applications, and an ability to quantify inherent problem uncertainties. This ability to capture complex system behaviors, account for uncertainties, and build in continuous learning capability, all in a single loop, reduces the computational cost of calibration compared to conventional approaches for higher-order models.

Committee:

Dr. Dimitri Mavris (advisor), School of Aerospace Engineering
Dr. Lakshmi Sankar, School of Aerospace Engineering
Dr. Jonnalagadda Prasad, School of Aerospace Engineering
Dr. Olivia Pinon Fischer, School of Aerospace Engineering
Mr. Andrew Dugenske, Georgia Tech Manufacturing Institute
Dr. Emmanuel Motheau, Siemens Technology