Multi-Level Simulation
Multi-Level Simulation
Simulations play an important role in many areas of research. Especially in Engineering disciplines it is established practice to evaluate and optimize sketches by means of simulations, enabling cost and time savings in building prototypes. The degree of abstraction is determined by the problem under consideration and can vary strongly. A factory with a production line with multiple machines can serve as an example: On a high level, the production line may be modeled abstractly with machines connected only via the exchange of input and output. On the more detailed machine level the mechanical interplay of the different machine’s components is simulated.
However, the examination of individual levels is not always adequate. In the example given above, it may be reasonable to simulate machines that are bottlenecks in detail as to predict machining times more accurately. The more abstract factory level, on the other hand, forms the context for the detailed simulation of single machines. For instance, the speed at which a machine receives upstream input influences e.g. its temperature and other metrics.
For these reasons, the project addresses so-called multi-level-simulations. They allow to couple the different levels of abstraction, in the context of our example the more detailed machine-level simulation with the more general factory model. Currently, this approach has two main obstacles. First, the mapping between different abstraction levels is a non-trivial task. Second, detailed simulations have a high and also varying demand for compute resources. The problems arising from these circumstances will be addressed by the project.
As multi-level-simulations examine different layers of abstraction, a synchronization between their states has to be performed. For example, the results of a detailed simulation have to be fed into higher levels of abstraction. Presently, there are no satisfying solutions for this complex problem as the state of the art is to manually define the state mappings, severely limiting the area of application for multi-level-simulations. The approach of this project is to employ machine learning for automatically generating the interfaces instead.
The second important problem is the computational complexity. Often, even parts of a simulation take multiple days to finish on a workstation computer. To achieve the goal of multi-level-simulations to model components of a complex system in detail and on demand, the necessary compute resources have to be provisioned elastically. To account for this, we will employ technologies from Cloud Computing in this project to dynamically scale the resources to the needs of the simulation.
The goal of this project is the development of a platform for multi-level-simulations that is able to automatically create the interfaces between different abstraction layers and which dynamically allocates compute resources, so that the user can fully utilize the capabilities of multi-level-simulations.