Three types of interpolation are available: nearest-neighbor, linear, or cubic.
Be careful about the choice of interpolation. In particular, when mapping volume data to the cortical surface, it is easy for "holes" to occur, depending on the extent to which valid values exist in the volume data and depending on the type of interpolation used.
In general, transformation between volume and surface spaces is lossy, in the sense that information loss and discretization errors are inevitable. One strategy is to perform analysis of the functional data fully in volume format and then transform to surface space at the very end (e.g. for visualization). A different strategy is to simply start up front with the "nativesurface" preparation (in which we have already transformed/interpolated the NSD betas to FreeSurfer's surface space) and then conduct analyses.
The conversion of surface data to volume format is a tricky procedure that involves certain assumptions. One particular method is implemented by nsd_mapdata (and is described in the NSD data paper), and this method was used in order to create volumetric versions of surface-oriented ROI labels (e.g. prf-visualrois). Other methods are possible.