AlvaroMS90
commented
2 months ago Thanks for your enquiry. Our article clearly indicates that the assumption of 2D flows implies a certain limitation in terms of reconstruction capacity, since the spurious errors derived from that hypothesis would be accumulated in the pressure gradient as part of the physics-constrained component of the loss function. However, in certain configuration, a 2D flow consideration is perfectly justified, specially if the velocity component in the third dimension becomes negligible (or equivalent its gradient).
Reynolds number should not be the driving parameter to consider a 2D flow. Of course, low Re imply laminar flows, which are more keen to be simplified to 2D flows than turbulent flows at high Re.
To use the code for 3D components, the only necessary steps would be to include the third dimension, both in the MLP stage in order to account for that dimension in the POD process, and in the PINN architecture and extend the loss function to account for the components in the third dimension.
chenling669
I noticed in your paper that both 3D and 2D PIV data can be used, yet the examples provided are exclusively two-dimensional. I am curious about the absence of three-dimensional examples. Is this because the Reynolds numbers are relatively low, allowing the velocity in the third direction to be negligible and thus averaged out? If I were to use your code for my research, how should I address this issue? Could you provide some guidance on this matter?