ML surrogate

The surrogate has been run on Cases A to E; its scores and figures are on each case’s Validation page: Case A, Case B, Case C, Case D and Case E. It has not been run on Cases F to M.

Eddy3D’s Wind Predictor (the public Yel-2.0 network) predicts \(\lvert\mathbf{U_{\mathrm{mean}}}\rvert/U_{\mathrm{ref}}\) at one height, 1.8 m. Each case is scaled so its measurement plane lands there (A ×180, B ×144, C ×90, D ×360, E ×0.9). Its buildings are rasterised onto the model’s 504 × 504 lattice of 2 m cells with the shipped grid code, the prediction is read at the tunnel’s own sensor coordinates, and \(U_{\mathrm{ref}}\) is converted to \(U_H\) with each case’s approach-flow file. The metric code and settings are those of the OpenFOAM rows.

FB is positive when the model is low.

  • FB is +0.23 to +1.07: the mean predicted speed is 0.79 to 0.30 of the mean measured speed.
  • Reversing the wind changes the predicted field by at most \(1 \times 10^{-4}\) \(U_{\mathrm{ref}}\) over all 254,016 cells, in every case.
Interpretation: Wind direction and what the scores grade
  • The network responds to spatial contrast in a channel and has no response to its level. On a lattice that fills the window the direction channels are constant planes; with the lattice cut at ±250 m the direction enters through the step at its edge, and the reversed wind then scores as well as the right one or better on A, B and D.
  • The scores grade the geometry given to the model; the wind direction does not enter.