Surrogate and Hybrid Models for Control

Abstract

With access to fast computers and efficient machine learning tools, it is of interest to use machine learning to develop surrogate models from complex physics-based models. Next, a hybrid model is a combination model where a data driven model is built to describe the difference between an imperfect physics-based/surrogate model and experimental data. Availability of Big Data makes it possible to gradually improve on a hybrid model as more data become available. In this paper, an overview is given of relevant ideas from model approximation/data driven models for dynamic systems, and machine learning via artificial neural networks. To illustrate how the ideas can be implemented in practice, a simple introduction to package Flux for language Julia is given. Several types of surrogate models are developed for a simple, illustrative system. Finally, the development of a hybrid model is illustrated. Emphasis is put on ideas related to Digital Twins for control.