GraphNetMF

For a material flow simulation to be suitable for VC, it must both represent reality as accurately as possible and ensure compliance with real-time requirements. Particularly in a hardware-in-the-loop (HiL) test configuration (according to VDI/VDE 3693), time-deterministic behavior of the simulation is required. Physical simulators, which are capable of representing reality most accurately, are generally not able to meet real-time requirements due to their very high computational complexity. Previous approaches to realizing a real-time-capable simulator, for example, based on a kinematic model or a flow model, have involved trade-offs in simulation accuracy. Consequently, it has so far generally not been possible to provide a highly accurate simulation while simultaneously meeting real-time requirements. Learnable simulators based on AI models show great potential for the simulation of physical systems. Existing approaches enable a significant acceleration of computation time while maintaining high simulation accuracy. To transfer this potential to material flow simulation, an appropriate learnable simulator is required.

The objective of the research project is to develop a suitable learnable simulator for material flow simulation. This simulator is intended to transfer the potential of learnable simulators to the domain of material flow simulation. GNNs will be used as the AI models, as they have already shown promising results in the simulation of physical systems. The aim is to achieve a three-dimensional simulation of the movement of large quantities of discrete goods on a conveyor belt. Phenomena such as accumulation at barriers, tipping, and jamming of items are to be represented realistically. To validate the simulation, an experimental test rig will also be developed, enabling real material flow to be recorded experimentally and compared with the simulation results. Finally, to ensure real-time capability of the learnable simulator, integration into a real-time system is planned.