Samples inside a mesh should follow the deformation of the mesh volume.
Idea: Use a grid to make sample deformation fast!
If the grid relaxation/blurring can be fast enough this would decouple the sample count from performance, allowing higher sample counts than what would be feasible with mesh deform. This is because the transfer (step 3) is very cheap, compared to Mesh Deform where a full set of weights has to be computed for each sample. The grid resolution determines performance but not visual detail.
OpenVDB provides methods for "extending" fields across a fog or SDF volume based on some boundary conditions (API methods). This is probably the best method for generating a deformation vector for interior voxels.
True deformation is an unknown function
where
Grid deformation is defined at voxel centers
This introduces an error