NEW AND SCALABLE PARADIGMS FOR DATA-DRIVEN MODEL PREDICTIVE CONTROL

Control and automation technologies have been instrumental in ensuring that societal systems (e.g., buildings, power networks, manufacturing facilities, autonomous vehicles, materials/fuels production) are operated in a safe, reliable, and sustainable manner. Advances in sensing technologies make possible the development of more efficient control technologies, but such devices generate data in complex formats (e.g., visual and thermal images) that need significant additional processing to be compatible with current automatic control systems. The goal of this project is to develop mathematical methods that enable the use of complex-format data sources for control. These data will be used to construct mathematical models of the system to be controlled to predict the behavior of the system in a reliable manner and to quantify risks associated with inaccurate predictions. This project will also support the development of new educational materials and computational tools that will help K-12, undergraduate, and graduate students better visualize and make sense of complex data; such skills are essential for enabling data-driven science and engineering careers. This project will develop a scalable paradigm for model predictive control (MPC) that make effective use of complex data (as opposed to single-point measurements). This will be done by integrating concepts of control, topology, machine learning (ML), and Bayesian analysis. Specifically, topology will be used as a general framework that facilitates representation of data that is attached to complex spaces (point clouds, fields/manifolds, and graphs/networks) and that enables the reduction of such data into informative topological descriptors that can be used for control. These descriptors will then be used to construct data-driven, dynamical models in a low-dimensional space using ML (e.g., recurrent neural networks), models that then will be embedded in MPC formulations. To guide data collection, Bayesian MPC formulations will be developed which interpret the controller as a real-time experimental design oracle that aims to simultaneously gather information to mitigate model uncertainty (exploration) and to maximize control performance (exploitation). A key research objective is the development of fast and scalable uncertainty quantification strategies that can work with large ML/physics-based models, making it possible to study the interplay between computational tractability and performance. The effectiveness of this new MPC formulation will be demonstrated with applications in energy, manufacturing, and materials systems. 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.