Agentic workflows developed for software development have changed our software development methods. Most of the trivial coding tasks are solved by AI in a fraction of time of what the software engineer used to take. Instead of coding, the focus of a software engineer is now on specifying the work to be done, choosing the right architecture, orchestrating the agents doing the work, and verifying the results.
Modelica – from its earlier definition – has always included a duality between the language and the user interface artifacts. And the language is no different for an LLM than a programming language like C, C++, Python, Julia, Rust, Wolfram, etc. Therefore, it took only a few months before a similar change happened for modeling and simulation as for software engineering. We do not model the same way as we did a year ago.
In this tutorial, we will first discuss the recent and upcoming evolution of the role of the simulation engineer towards a rather System Engineer type of work, assisted by agents. The skills of today’s simulation engineer however remain critical for successful simulations. We will demonstrate this by hands-on examples. In a second step, we will leverage Wolfram System Modeler and their newest agent skills to develop a detailed Modelica in an agentic manner, pointing at best practices and pitfalls.
At the end of this workshop, you will be equipped to develop your own model in a much faster and efficient way, as AI-embracing simulation engineer.

This session is chaired by:

This session is chaired by:

In this work, we compare a traditional design-ofexperiments (DoE) workflow, defined up-front and then run, against an agentic AI workflow that revises its plan as the data arrives. The use case is a chassis optimization on the standard compact example in the Modelon VehicleDynamics Library. Both workflows produce a recommendation. The DoE arrives at a clean Pareto front and a stiffness change. The agentic path is longer: along the way it flags a controller saturation in the DoE winner, traces a center-ofgravity asymmetry to a load sensitivity the original scope did not cover, and finds a kinematic ceiling outside the four DoE parameters. It ends up changing a different parameter, and that parameter turns out to do more than the static KPIs asked for: dynamic verification shows the bushing change roughly halves the fishhook return lag unloaded. The architecture that makes this possible combines validated physics, Modelon Impact, the Model Context Protocol that connects them to an LLM, and a knowledge base layered by lifetime. We close with what we learned about when each approach is the right tool.

This paper introduces State-Invariant Hybrid Models (SIHM), a subset of hybrid physics-machine-learning models in which the state vector remains defined by the first-principles Modelica model, while learned components contribute missing physics locally. We qualitatively contrast global residual learning at system level with local residual learning at component level, arguing that local learning better matches the acausal modeling paradigm of Modelica. A tool-agnostic workflow is proposed in which a replaceable dummy component marks the suspected location of missing physics, while compiler-derived structural information identifies the connected states and variables. The workflow is then illustrated in the Wolfram ecosystem on two models to demonstrate the approach's modularity. The results indicate that local SIHM components recover targeted missing effects while preserving the state selection of the original model.