In fact the smaller the grid step is, the better the approximation will be. However, for small step values you'll generally get holes... which are not good for the approximation of the volume. You can of course try to use the 'interpolation' filling strategy to cope with that issue. But it depends on the size of the holes and how smooth is the real surface.
Anyway, assuming the clouds are regularly sampled, a rule of thumb for the grid step would be:
- compute the average "2D" surface of a cloud by simply multiplying the X and Y dimensions of its bounding-box (S = DX * DY)
- divide this surface by the number of points (S / N)
- compute the square root of the per-point surface to get a first approximation of your ideal grid step