Tobias Thummerer
... is an advanced doctoral student at University of Augsburg. He completed both his Bachelor's and Master's degree in “Engineering and Computer Sciences” at this same university. His research focuses on Scientific Machine Learning, and he is specifically engaged with the area of Hybrid Modeling, the combination of simulation models – often from engineering – and novel machine learning approaches. His research includes the development of new methods in this field, but explicitly also the improvement of new and established approaches so that they can be applied in a demanding real-world environment.
Sessions
The Functional Mock-up Interface (FMI) is the standard for exchanging industrial simulation models in a variety of different applications. Although sensitivity analysis for continuously differentiable systems is directly supported by the standard, for systems with state discontinuities, it is only possible to determine correct sensitivities to a limited extent. In this position paper, we investigate how sensitivity analysis for discontinuous Functional Mock-up Units (FMUs), i.e. including state and time events, works in theory and which additional steps are required to obtain correct results in practice. We further investigate that these steps are unnecessarily computationally intensive from a mathematical point of view, but cannot be implemented in a more efficient way under the current restrictions of the standard. We therefore make a concrete proposal for the new layered standard sensitivity analysis (LS-SA) that remedies the current deficits of FMI in the sensitivity analysis of discontinuous systems. In this way, LS-SA opens FMI towards a variety of next-level applications — including (scientific) machine learning and optimal control — by providing fully differentiable FMUs under high computational performance.
Testing real hardware and simulation models in combination in a software- or hardware-in-the-loop set-up is challenging. One of the key factors is the high demand for accuracy in the simulation model. If classical modeling based on physical principles is not sufficient to reach the desired level of accuracy, hybrid modeling, the combination of physical simulation models and machine learning can be applied. In this publication, we train a hybrid model for a controlled electric motor within the electro-hydraulic braking system of a car under the conditions and restrictions of a real engineering application in the field. We apply state-of-the-art modeling patterns for this, and further extend them with application specific methodological optimizations. Finally, we investigate and show the quantitative and qualitative advantages of the proposed approach for this specific application, resulting in a gain in accuracy by multiple factors.