Physics-consistent machine learning: A neural ODE perspective

In this talk, we will discuss how Neural ODEs—a family of continuous-depth neural networks —can be leveraged to learn nonlinear dynamics with desired properties, such as, stability, contractivity, passivity, or energy conservation from data. Particularly, we wilL explain how to embed free parametrization from the outset to guarantee these desired properties by design. Moreover, I will introduce our recently developed linear system identification toolbox namely SIMBa that leverages machine learning tools to identify stable linear models from data. Finally, we will demonstrate the efficacy of our methods through several experiments.