CFD-solver-agnostic deep reinforcement learning framework for computational fluid dynamics on HPC platforms

Turbulence modelling in CFD is limited by the tradeoff between accuracy and cost. We propose OT PINNs, Physics Informed Neural Networks with an Optimal Transport based loss, to improve training stability and accuracy under noisy data. With SINDy for interpretability, our method rivals DNS on benchmark flows while cutting computational costs.

This project implements a PINN using TensorFlow to solve a 2D steady-state convection–diffusion PDE on a unit square by minimizing the PDE residual and enforcing Dirichlet boundary conditions. It demonstrates domain sampling, differentiation, constrained training, and inference on unseen test points without requiring labeled solution data.

Automated calibration of RANS turbulence models for hypersonic flows using SciML. Achieved 61.3% RMSE reduction at Mach 14.

2D CFD simulation of NACA0012 at 5° AoA using SST k-ω model. Mesh convergence included.

Real meteorological data (temperature, pressure, precipitation) were obtained from İSKİ. Eddy diffusivity, Monin-Obukhov length, and turbulence intensity were calculated from existing data and added to the dataset. Using 289,000 data points and 27 features, RF, SVM, LSTM, and CNN models were developed. LSTM achieved 98%, CNN 91% accuracy.