Conclusion: directly using deep neural nets to predict the final result is not even close to a viable approach.
- A stationary (in time) target doesn't always exist.
- The numerical variation of the flow parameters (Mach, AoA, etc.) serves too weak as a conditioning factor. The model would degenerate to just output the average of what it have seen during training.
- Spatial discretization is crucial for high fidelity but is unfortunately tricky.
- Example: a conv-kernel is not consistently defined
-
- Enforcing physical constraints (e.g. applying boundary conditions) may help but is difficult in formulation.
- report
Run "convert_to_np" in utils.data with specified source and target directories to generate data in .npy format. By default, all fields are retained. This would produce output of shape (H, W, I, C) where H, W and I describes the dimensions of the mesh (I, which stands for z-axis, is actually not used here), and C is the number of fields.
