Merge pull request #544 from matthewhoffman/landice/esmf_interp_fix

Fix bugs in ESMF interpolation method

This merge fixes a number of issues with the 'esmf' method for interpolation in the MALI interpolation script:
* indexing was off by one, which made interpolation with unstructured source meshes garbage. For structured source meshes, it made interpolation shifted by one grid cell.
* support MPAS source fields with a vertical dimension when using the 'esmf' method
* refactor to use sparse matrix multiply, which speeds up interpolation a few hundred times
* add destination mesh area normalization support. This is necessary when using the ESMF 'conserve' method for destination cells that are only partly overlapped by the source cells.

Two issues remain for interpolating between two MPAS meshes with the 'esmf' method:

1. If the destination mesh is larger than the source mesh, those locations are filled with zeros. The ESMF 'conserve' method does not support extrapolation, and there is no obvious solution to this issue and it would need to be handled manually on a case by case basis.
2. Some fields are only defined on subdomains (e.g. temperature is only defined where ice thickness is nonzero) and the script currently has no mechanism for masking them. This results in garbage values getting interpolated in, e.g. temperature values near the margin will have values around 100 K because of interpolating realistic values around 250K with garbage values of 0K. This issue applies to the barycentric method as well. However, this situation can be worked around by performing extrapolation of the temperature field on the source mesh before doing interpolation between meshes.

This PR also includes a refactoring of create_SCRIP_file_from_planar_rectangular_grid.py that speeds it up by orders of magnitude for large meshes, as well as a minor update to define_cullMask.py

Author: matthewhoffman

landice/mesh_tools_li/define_cullMask.py

@@ -101,7 +101,7 @@
               if keepCellMask[n] == 1:
                   keepCellMaskNew[iCell] = 1
     keepCellMask = np.copy(keepCellMaskNew)  # after we've looped over all cells assign the new mask to the variable we need (either for another loop around the domain or to write out)
-    print('  Num of cells to keep: {}'.format(sum(keepCellMask)))
+    print(f'Num of cells to keep: {keepCellMask.sum()}')
 
   # Now convert the keepCellMask to the cullMask
   cullCell[:] = np.absolute(keepCellMask[:]-1)  # Flip the mask for which ones to cull
@@ -148,7 +148,7 @@
       ind = np.nonzero(((xCell-xCell[iCell])**2 + (yCell-yCell[iCell])**2)**0.5 < dist)[0]
       keepCellMask[ind] = 1
 
-  print('  Num of cells to keep:'.format(sum(keepCellMask)))
+  print(f'Num of cells to keep: {keepCellMask.sum()}')
 
   # Now convert the keepCellMask to the cullMask
   cullCell[:] = np.absolute(keepCellMask[:]-1)  # Flip the mask for which ones to cull

landice/mesh_tools_li/interpolate_to_mpasli_grid.py

@@ -32,6 +32,7 @@
 import math
 from collections import OrderedDict
 import scipy.spatial
+import scipy.sparse
 import time
 from datetime import datetime
 
@@ -64,9 +65,15 @@
     S = wfile.variables['S'][:]
     col = wfile.variables['col'][:]
     row = wfile.variables['row'][:]
+    n_a = len(wfile.dimensions['n_a'])
+    n_b = len(wfile.dimensions['n_b'])
+    dst_frac = wfile.variables['frac_b'][:]
     wfile.close()
     #----------------------------
 
+    # convert to SciPy Compressed Sparse Row (CSR) matrix format
+    weights_csr = scipy.sparse.coo_array((S, (row - 1, col - 1)), shape=(n_b, n_a)).tocsr()
+
 print('') # make a space in stdout before further output
 
 
@@ -78,15 +85,20 @@
 #----------------------------
 
 def ESMF_interp(sourceField):
-    # Interpolates from the sourceField to the destinationField using ESMF weights
+  # Interpolates from the sourceField to the destinationField using ESMF weights
+  destinationField = np.zeros(xCell.shape)  # fields on cells only
   try:
-    # Initialize new field to 0 - required
-    destinationField = np.zeros(xCell.shape)  # fields on cells only
-    sourceFieldFlat = sourceField.flatten()  # Flatten source field
-    for i in range(len(row)):
-      destinationField[row[i]-1] = destinationField[row[i]-1] + S[i] * sourceFieldFlat[col[i]]
+    # Convert the source field into the SciPy Compressed Sparse Row matrix format
+    # This needs some reshaping to get the matching dimensions
+    source_csr = scipy.sparse.csr_matrix(sourceField.flatten()[:, np.newaxis])
+    # Use SciPy CSR dot product - much faster than iterating over elements of the full matrix
+    destinationField = weights_csr.dot(source_csr).toarray().squeeze()
+    # For conserve remapping, need to normalize by destination area fraction
+    # It should be safe to do this for other methods
+    ind = np.where(dst_frac > 0.0)[0]
+    destinationField[ind] /= dst_frac[ind]
   except:
-     'error in ESMF_interp'
+    print('error in ESMF_interp')
   return destinationField
 
 #----------------------------
@@ -328,7 +340,7 @@ def interpolate_field_with_layers(MPASfieldName):
         if filetype=='cism':
            print('  Input layer {}, layer {} min/max: {} {}'.format(z, InputFieldName, InputField[z,:,:].min(), InputField[z,:,:].max()))
         elif filetype=='mpas':
-           print('  Input layer {}, layer {} min/max: {} {}'.format(z, InputFieldName, InputField[:,z].min(), InputField[z,:].max()))
+           print('  Input layer {}, layer {} min/max: {} {}'.format(z, InputFieldName, InputField[:,z].min(), InputField[:,z].max()))
         # Call the appropriate routine for actually doing the interpolation
         if args.interpType == 'b':
             print("  ...Layer {}, Interpolating this layer to MPAS grid using built-in bilinear method...".format(z))
@@ -349,7 +361,10 @@ def interpolate_field_with_layers(MPASfieldName):
                 mpas_grid_input_layers[z,:] = InputField[:,z].flatten()[nn_idx_cell]  # 2d cism fields need to be flattened. (Note the indices were flattened during init, so this just matches that operation for the field data itself.)  1d mpas fields do not, but the operation won't do anything because they are already flat.
         elif args.interpType == 'e':
             print("  ...Layer{}, Interpolating this layer to MPAS grid using ESMF-weights method...".format(z))
-            mpas_grid_input_layers[z,:] = ESMF_interp(InputField[z,:,:])
+            if filetype=='cism':
+                mpas_grid_input_layers[z,:] = ESMF_interp(InputField[z,:,:])
+            elif filetype=='mpas':
+                mpas_grid_input_layers[z,:] = ESMF_interp(InputField[:,z])
         else:
             sys.exit('ERROR: Unknown interpolation method specified')
         print('  interpolated MPAS {}, layer {} min/max {} {}: '.format(MPASfieldName, z, mpas_grid_input_layers[z,:].min(), mpas_grid_input_layers[z,:].max()))

mesh_tools/create_SCRIP_files/create_SCRIP_file_from_planar_rectangular_grid.py

@@ -125,18 +125,19 @@
 print ('Filling in corners of each cell.')
 grid_corner_lon_local = np.zeros( (nx * ny, 4) )  # It is WAYYY faster to fill in the array entry-by-entry in memory than to disk.
 grid_corner_lat_local = np.zeros( (nx * ny, 4) )
-for j in range(ny):
-  for i in range(nx):
-    iCell = j*nx + i
-
-    grid_corner_lon_local[iCell, 0] = stag_lon[j, i]
-    grid_corner_lon_local[iCell, 1] = stag_lon[j, i+1]
-    grid_corner_lon_local[iCell, 2] = stag_lon[j+1, i+1]
-    grid_corner_lon_local[iCell, 3] = stag_lon[j+1, i]
-    grid_corner_lat_local[iCell, 0] = stag_lat[j, i]
-    grid_corner_lat_local[iCell, 1] = stag_lat[j, i+1]
-    grid_corner_lat_local[iCell, 2] = stag_lat[j+1, i+1]
-    grid_corner_lat_local[iCell, 3] = stag_lat[j+1, i]
+
+jj = np.arange(ny)
+ii = np.arange(nx)
+i_ind, j_ind = np.meshgrid(ii, jj)
+cell_ind = j_ind * nx + i_ind
+grid_corner_lon_local[cell_ind, 0] = stag_lon[j_ind, i_ind]
+grid_corner_lon_local[cell_ind, 1] = stag_lon[j_ind, i_ind + 1]
+grid_corner_lon_local[cell_ind, 2] = stag_lon[j_ind + 1, i_ind + 1]
+grid_corner_lon_local[cell_ind, 3] = stag_lon[j_ind + 1, i_ind]
+grid_corner_lat_local[cell_ind, 0] = stag_lat[j_ind, i_ind]
+grid_corner_lat_local[cell_ind, 1] = stag_lat[j_ind, i_ind + 1]
+grid_corner_lat_local[cell_ind, 2] = stag_lat[j_ind + 1, i_ind + 1]
+grid_corner_lat_local[cell_ind, 3] = stag_lat[j_ind + 1, i_ind]
 
 grid_corner_lon[:] = grid_corner_lon_local[:]
 grid_corner_lat[:] = grid_corner_lat_local[:]
@@ -171,3 +172,4 @@
 
 fin.close()
 fout.close()
+print('scrip file generation complete')