Seminars

Home Centers & Programs AI and Natural Sciences Seminars

TITLE: Physics-informed neural networks: Fitting a mathematical model to real data using artificial neural networks

ABSTRACT: A dynamical system $y’(t)=f(y)$ can be used to model the evolution of natural or engineered systems. Traditionally, the overall structure of the system $f$ is determined by researchers' insights (experience) or experimental observations (data). However, such insights might fail to capture the system's complexity and nonlinearities due to the limitations in human intuition and experimental precision. In response, data-driven scientific discovery methods have been developed by employing artificial neural networks (ANN). Specifically, ANN is trained to simultaneously fit the system $f$ and the data, called Physics-informed Neural Networks (PINN). In this presentation, we will study 1) the basic idea of the PINN 2) effectiveness of the PINN. Furthermore, we show how we can extend the concept of PINN to real-world dataset, thus understanding the structure of the system $f$.