PI: Srinivas Rangarajan
Co-PI(s): Mayuresh Kothare
University: Lehigh University
Most chemical and biological processes are dynamical systems. This means that their state variables (i.e. variables that together characterize what state the system is in) are continuously changing, often underlined by highly nonlinear correlated behavior that many not be easily captured by physics-based models. Modern plants in the energy and chemical industry have advanced data acquisition technologies, enabled in many cases by solutions offered by OSISoft LLC, the industrial participants of this project. These technologies allow for collecting, storing, and analyzing data from thousands of sensors every second (or faster). Our ultimate goal in this context is to leverage such data to design, optimize, and control new energy and chemical systems, i.e. in the absence of accurate first-principles models.
In this project, we will begin addressing this larger goal by developing algorithms that will allow us to extract the underlying ordinary differential equations from time-varying data. Such an algorithm will then allow us to take time-varying plant data and build data-driven dynamic equations that accurately captures the overall process. We specifically intend to build on the state-of-the-art algorithms from the applied mathematics community on inferring equations from data that have been successfully applied in the fluid mechanics domain by incorporating a number of new features including the concept of infusing chemical engineering domain knowledge as constraints while training the data-driven model. This project will focus on developing this method and testing it on industrial datasets with software supplied by OSISoft.