research associate (m,f,x) for 3years with 39,83 hours per week (TV-L E 13)

Priority Programme 2304 Carnot-Batteries: Inverse aerodynamic design of turbo components for Carnot batteries (by means of physics informed networks enhanced) by generative learning

Within the Priority Programme SPP2304 (DFG, German Research Foundation - Priority Programme “Carnot Batteries: Inverse Design from Markets to Molecules” (SPP 2403)), the project “Inverse aerodynamic design of turbo components for Carnot batteries by means of physics informed networks enhanced by generative learning” aims at developing methods from artificial intelligence (AI) to accelerate the inverse design of turbo-machinery components of Carnot batteries. In particular, we will apply generative adversarial networks (GAN) and physics informed neural networks (PINN) to the simulation of turbulent flows with vast variations in the boundary conditions. By the enormous speed-up realizable by AI-driven simulation and design procedures, we endow the inverse design of Carnot batteries with new tools that have the potential to accelerate formerly used unsteady fluid dynamics simulations by a factor of 100 or more. The ideal candidate will work in close cooperation with the research partner TU Berlin specifically on the generation of a virtual validation/training datasets and on the development of a PINNs computation platform, to be extended to encompass the relevant phenomenology of a condensing fluid mixture. The optimal form of the mathematical model, so far only analyzed for very simple physical systems, must be investigated for the construction of an appropriate set of loss functions. Also the appropriate form and composition of the unknown variable vector will have be to identified to guarantee stability and convergence of the solver. The solution of a direct as well as inverse set of problems will be attempted, whereby increasingly complex sets of data (i.e. feature-rich but sparse and/or noisy) will be adopted to train the algorithm. Also the best combination of optimizer and neural network architectures will be assessed. The coarse-grained PINNs solutions thus obtained will be continuously provided to the research partner to establish the envisaged inverse design tool-chain. We look for highly motivated, enthusiastic and highly qualified research assistants, who will carry out the proposed research engaging in numerical investigations with the possibility of obtaining a doctorate.