Data-driven models and control

Modern engineering systems are becoming increasingly complex, nonlinear, and data-rich. From robotics and autonomous vehicles, to energy and biological systems, new sensing and computation capabilities now allow us to observe dynamics at unprecedented scale. The challenge is no longer the lack of data—but how to turn it into models that support analysis, prediction, and control.

Purely black-box models—such as neural networks, recurrent architectures, Gaussian processes can capture complex behavior directly from data, but often sacrifice interpretability, structure, and reliability. At the other extreme, fully physics-based models provide insight but can be difficult to derive and adapt especially when the system is large. Gray-box models seek the best of both worlds, combining data with mathematical structure to produce models that are both expressive and usable for control.

The Koopman operator framework provides a gray-box perspective on nonlinear dynamical systems. By representing system evolution through observable functions, it enables us to obtain linear models of nonlinear systems. These models can integrate both the prior physical knowledge, and the data from the system, giving rise to hybrid physics–data Koopman models. These representations are structured, interpretable, and directly compatible with established tools for prediction, estimation, optimization, and control. 

Research Goals

I am interested in the following research questions