CAREER: Bayesian Machine-Infused Physics-Based Data Assimilation for Digital Twinning and Uncertainty Quantification of Dynamical Systems

This Faculty Early Career Development (CAREER) award supports research that enables a new paradigm for the integration of data with machine-infused physics-based computational models to develop digital twins of dynamical systems, thereby promoting the progress of science, and advancing prosperity and welfare. The advent of the big data era necessitates robust, efficient, and scalable scientific tools that assimilate raw measurements into computational models to enhance the quality and resolution of actionable information in near real-time. Despite the recent prolific trend in the predictive modeling literature, current hybrid data-driven modeling approaches focus on physics-informed machine learning and often face shortcomings in real-world applications due to limited interpretability, limited generalizability outside training domain, and limited training data. Research completed in association with this project will address this critical gap by embedding neural-network components in the core of physics-based models to account for structural errors, simplifications, approximations, idealizations, and unknown physics in the model formulation, and train/update the integrated model with data. The resulting theoretical framework will then be applied to three use-inspired societal challenges: post-earthquake damage assessment of critical civil structures, predictive maintenance of offshore wind infrastructure, and emergency response to wildland fires. In addition to its broad potential applications across various science and engineering fields, the project will improve engineering education and broaden participation in STEM through its robust educational outreach plan that includes K-12, undergraduate, and graduate level engagement with students from underrepresented demographics. This research aims to improve computational efficiency, scalability, and practicality of dynamic models to enhance situational awareness and informed decision-making. It will achieve this goal by developing a Bayesian machine-infused physics-based data assimilation framework to address the fundamental challenge of modeling error in physics-based data assimilation and uncertainty quantification. An inherent part of any physics-based model, modeling error – or model form uncertainty – can disparage the physics-based data assimilation process and the subsequent identification, prediction, virtual sensing, diagnosis, and prognosis applications. The machine-infused physics-based model will be trained using system-level measurements through a Bayesian inference approach. The intellectual merits of the project include (1) theoretical development and implementation of the framework, (2) addressing technical challenges regarding identifiability, robustness, and computational efficiency and scalability, and (3) application to three engineering problems, namely, post-earthquake damage assessment of civil structures using a machine-infused finite element data assimilation, predictive maintenance of wind turbine drivetrains using a machine-infused multibody dynamics data assimilation, and operational prediction of wildfires using a machine-infused coupled fire-atmosphere data assimilation. This project is jointly funded by the Dynamics, Control and Systems Diagnostics (DCSD) program, and the Established Program to Stimulate Competitive Research (EPSCoR). This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.