AI/PINNs for Micro-turbine CFD? Seeking surrogate models for 100 kHz+ rotor optimization
Hi everyone,
I’m new to the sub. I’m a postdoc in a research group where we use micro-turbines to pneumatically spin rotors for chemical applications. We are currently pushing the limits of spinning frequencies (>50 kHz, goal: 200 kHz range) using very small rotors (approx. 0.5-2 mm).
Historically, cylindrical rotor designs have been used, but we are now transitioning to spherical rotor designs and improving their spinning efficiency. The stator for these spherical rotors is sort of a "cup" with air nozzles. I am looking to perform a multi-parameter optimization of the stator (nozzle count/angles, cup depth, exhaust geometry) to improve torque efficiency and stability.
Coming from a computational chemistry background, I’m used to Machine Learning Interatomic Potentials (MLIPs) that can solve chemical problems with the accuracy of classical quantum packages (computationally expensive) but at a fraction of the computational cost. Nowadays, we can even do calculations on personal laptops with the same accuracy of calculations that used to take a few days in HPCs with multi-node setups.
I’m wondering: Is there a CFD equivalent that is mature enough for this?
Specifically:
Are there any foundation models that actually handle high-frequency, transonic micro-flows well? Or am I better off sticking to a traditional package like OpenFOAM?
My goal is to run through a massive candidate space for optimization. If you were starting this from scratch today, would you build a PINN, use a Neural Operator, or just script a massive OpenFOAM parallel run on a cluster?
Disclaimer: I’m a CFD novice but a experienced in molecular simulations and HPC calculations. I’d love to hear from anyone using AI/ML to bypass the "re-meshing nightmare" of parametric optimization.
Thanks and feel free to correct me if I am saying something or assuming something that is wrong!