Hybrid simulation models combine physics equations with trainable components to improve simulation results and performance. Physics-enhanced neural ordinary differential equations (PeN-ODE) are a promising type of hybrid models that combine artificial neural networks (NN) with the differential equations of a dynamic system. Dynamical simulation models are often part of embedded control algorithms of cyber-physical systems (CPS); compliance with the safety and real-time requirements of such embedded environments is, however, challenging.

In this work, we propose a workflow to incorporate trained NNs in Modelica models to form hybrid simulation models that are PeN-ODEs. We thereby focus on the transformation steps from equation-based trained PeN-ODEs in Modelica towards causal solutions suited for the embedded domain -- up to and including MISRA C:2023 compliance checks and final software-in-the-loop (SiL) tests of generated production code in the modeling environment -- for which we leverage eFMI standard compliant tools (Dymola and Software Production Engineering). It is of particular interest, how the trained NNs of the hybrid model are implemented. We present two approaches: (1) generation of C code using existing Open Neural Network Exchange (ONNX) tooling and (2) pure Modelica code with the tensor-flow represented as multi-dimensional equations. Both approaches are discussed, highlighting why (2) is, in the long run, a better option given the eFMI technology space.

The ability to systematically compare and evaluate diverse control strategies is essential for the development of effective control algorithms in autonomous driving. This contribution presents the VDCWorkbench Modelica Library, a unified platform designed to support the development, testing, validation and verification of vehicle dynamics controllers and energy management strategies. The presented library is an extension of the IEEE VTS Motor Vehicle Challenge 2023 models and offers multi-physical component modeling, including a hybrid energy storage system (fuel cell & hydrogen tank and battery with aging model), as well as vehicle dynamics control for autonomous driving research projects. Two path-following approaches are featured: an open-loop lateral controller with a static inversion of a single-track model, and a closed-loop state-dependent geometric path-following controller with static control allocation. The library may also serve as the foundation for development of vehicle control methods, such as two-degree-of-freedom control approaches concepts. One example is given for this combination of a feedforward controller with residual reinforcement learning, where a learned agent improves the performance of the open loop controller. The entire library will be released as open source on GitHub in September 2025.