NSF Award
2436801
Digital twins mimic actions and processes of physical assets in real-life executions. This research project concerns with development of a framework for learning digital twins of physical systems capable of incorporating real-world data into first-principles based mathematical representations. Learning digital twin from real data is a novel capability which can help enable effective strategic planning in various domains such as space exploration, autonomous transportation, sustainable water future, smart manufacturing, critical mineral mining, alternative power generation, and healthcare to name a few. This project focuses on applications to important problems in healthcare sciences related to data-informed decision-making exploiting virtual representations of human physiology and has implications for the development and evaluation of new therapies and treatments. One compelling example application is glucose metabolism in people with Type 1 diabetes (T1D). Patients with T1D must replace insulin exogenously as determined by multiple daily measurements of the blood glucose concentration, to maintain glucose homeostasis and avoid hypo / hyper-glycemia and life-threatening diabetic ketoacidosis. As a result, the person with diabetes has to make multiple, complex decisions each day based on food composition, exercise, hormonal cycles and other behavioral factors. Personalized glucose metabolism digital twins developed through this award will be used to devise new ethical treatment modalities and evaluate safety and effectiveness of automated insulin delivery systems without risk factors. Digital twins as such can also feed essential knowledge about system safety and effectiveness to regulatory agencies through assurance cases and advance regulatory science in profound ways. Both study sites University of Houston and Arizona State University are Hispanic serving institutions and the research is integrated with educational and outreach activities to create awareness, especially among youths, and understanding of diabetes and its management, broaden participation of groups traditionally underrepresented in STEM and contribute positively to engineering education. <br/><br/>The first-principles informed data-enabled framework seeks to advance foundational techniques underpinning the development and use of digital twins and synthetic data in biomedical and healthcare domains, by combining advances across mathematical modeling, machine learning (ML) and systems’ theory with human physiology. This research will (1) develop advanced structures based on neural networks (NNs) for the recovery of an underlying physics-based model that are capable of operating in real-world conditions characterized by limited data availability, low and non-uniform sample rate and spatial and temporal noise, (2) develop novel parametrizations of black-box dynamics using NNs from a class of models with "built-in" properties of stability and robustness to perturbations, (3) integrate real-world physical twin generated data which are heterogeneous, scarce and noisy, into its virtual first-principles based mathematical representation, (4) develop a novel framework for learning unmodeled dynamics due to e.g., unaccounted for inputs, inter- and intra-individual variability. Extensive evaluation of this methodology will be conducted using publicly available datasets specially for T1D patients.<br/><br/>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.
