Physics-informed reinforcement learning enabling ergonomic aerodynamic shape optimization
For the purpose of finding the optimal design of a diffusor with respect to highest homogeneity in the flow of an air source heat pump a physics-informed reinforcement learning approach is developed for improved optimization performance of aeroynamic shape design. The physical insight about the homogeneous flow being preserved mostly under a low net degree of turbulence allows introducing a surrogate model yielding an extra term in the reward function. Specifically, it is shown that this physics informed reinforcement learning approach for a two-dimensional Navier–Stokes flow-problem yields an increase in optimization speed.