There is an increasing need for hybrid models consisting of physical and machine-learning or surrogate model parts. XRG has developed the SMArtInt+ Library, which enables the easy generation and integration of neural networks from different sources and types. This tutorial provides a hands-on introduction to the use of this library, which also includes a Python toolbox. The main use case is the creation and integration of a transient surrogate model for a thermodynamic application.
The tutorial focuses on the generation and utilization of training data, as well as the definition and training of neural network architectures, with the objective of developing high-performance surrogate models. Participants will be introduced to different concepts and approaches for creating data-driven surrogate models and will gain an understanding of the underlying methodologies and best practices.
A central part of the tutorial is the integration of the trained neural networks and surrogate models into Modelica models using the SMArtInt+ Interface-Blocks, allowing seamless substitution of physical components with learned surrogates. Additionally, several concepts of hybrid modeling that combine physical and data-driven approaches are introduced, with a strong focus on the practical integration of trained neural networks and surrogate models within the Modelica environment.
Finally, model performance is evaluated through comparisons with the original physical model.