Trustworthy Physics-Informed AI for Aerospace

Kurzfassung

In situations where computing capacity is limited and real-time capability is crucial, as is often the case in DLR's engineering applications, traditional physical approaches are insufficient and surrogate models are needed. However, with data-driven neural networks natural laws like conservation of mass, energy and momentum are not or only poorly considered. They also lack information about uncertainties. In artificial intelligence systems these can occur, for example, in the model parameters, the input data or within the model itself. Neglecting uncertainties could lead to unreliable decisions. Therefore, a deep understanding of them is essential to assess the prediction quality, especially for applications with high safety requirements. The new DLR project TIARA aims to develop advanced artificial intelligence models that incorporate both uncertainty quantification techniques and physical principles. This approach ensures that the safety aspect is considered from the outset. Furthermore, these innovative models combine the precision of traditional physical approaches with the efficiency of AI-based surrogate models. Incorporating uncertainty quantification into AI models will ultimately significantly increase the safety and reliability of simulations and the decisions based on them. It is planned to use the developed AI models in edge computing and control engineering in order to obtain a high-performance and robust controller for controlling safety-critical, autonomous systems. In aerodynamic design, the new methods are intended to serve as fast and robust reduced-order models in aviation, which can then be used for optimization chains, for example, taking uncertainties into account.