A Variational Inference Gradient-Free Parameter Optimizer for Simulations in Aerospace Applications

Kurzfassung

The simulation of complex technical systems has become an important tool for engineers in the production and construction industry as computer hardware has further developed in the past decades and huge computational resources are nowadays available almost everywhere. However, the quality of the simulation relies on the fit of its parameters, which have to be adjusted according to specific properties of the material or system of interest. In many cases, there exists no direct correspondence between the measurements and the model parameters so that specific - simplified - methods have to be used to translate the given data into those parameters. Those methods are often decoupled from the simulation framework, leading to a bad alignment of the simulation result with the observations, that is also missing any uncertainty information, since measurement errors are not considered in this process. We present a novel black-box optimization framework that uses variational inference methods for the parameterfit of complex technical systems and processes in aerospace. The optimizer is built around the simulation software, which is considered as the black-box. Due to the absence of derivates we base this optimization framework on gradient-free techniques such as BOBYQA in order to minimize the difference between the simulation and given observation datapoints. Furthermore, we incorporate measurement errors with suitable error distributions via variational inference methods into the optimization process. The resulting model parameters are then given as posteriori distributions instead of single scalar values per parameter. We can therefore propagate the data uncertainties through the parameterfit into the simulation. Our method is used in aerospace applications, where a corresponding simulation framework is present. Its model parameters are fitted with our new approach and the resulting simulation outcome is then compared with real measurements.