Simulating physical systems with a spatially distributed domain requires solving intricate and computationally demanding differential equations. Utilizing the mesh-like structure inherent in these methods and integrating machine learning techniques in place of traditional equations allows for a substantial enhancement in computation time during the evaluation of the systems. This fundamental idea underlies the MeshGraphNets framework [1] developed by Google DeepMind. By formulating MeshGraphNets as NeuralODEs [2], you can fully leverage the extensive array of solvers available for ordinary differential equations. This adaptability allows for the application of previously trained models under a wide range of operating conditions. Incorporating the solver into the training process allows for a deeper understanding of the network’s grasp on the underlying system. Furthermore, optimizing the surrogate model’s evaluation is attainable by selecting an appropriate solver – implicit solvers for systems with higher stiffness or explicit methods for simpler systems result in significantly expedited evaluation times. This flexibility renders the framework applicable to industrial contexts with stringent precision requirements. We showcase an industrial application to validate the capabilities of our framework. The application consists of a model of the temperature distribution inside a vehicle cabin. Our demonstration illustrates that the network effectively learns the added intricacy of active ventilation through the air conditioning system, yielding results that meet the specified precision requirements.