I don't have a good answer for your specific question, but my desire for precise ND graduated filters has led me to think about solutions somewhat similar to your problem.
To begin with, it's instructive to understand the current state of the art for how graduated ND filters are created. As an example, here's a video from Lee Filters about how their ND grads are made. Personally, I'm appalled at how imprecise and error-prone their filters are created. The technician is literally hand-dipping the plastic filter into a hot dye bath, where the duration of submersion determines the ND factor, and the rate of slow removal from the bath determines the hard/soft transition region (while the technician talks to the interviewer, looks over her shoulder and causes the filter to dip up and down). This is something that really should be automated by a robot.
Furthermore, you can understand how the demonstrated dip technique cannot easily create filters such as reverse-ND grad strip filters, or apodization filters.
One technique that I think could create precise graduation regions of arbitrary shape would be to control a vapor deposition process (such as chemical vapor deposition (CVD) or physical vapor deposition (PVD)) by some means, such as electrostatic control, to deposit the vaporized ND material in precise locations and amounts (i.e., density). Essentially, it would be somewhat analogous to laser printing the gradient of ND material across the substrate.
Speaking directly to your idea of removing material from an ND filter to create an apodization filter, I'd be inclined to try mechanical removal rather than chemical. I'd experiment with different semispherical shapes/radii of soft buffing wheels mounted to a mill or drill press, different buffing rotational speeds and durations, and possibly different buffing/rouge compounds, to remove the deposited material from a ND filter. Almost certainly, you'd want to use glass ND filters rather than optical resin filters. The glass is much harder (i.e., resistant to scratching on the Moh scale) than resin, so if you use soft enough buffing wheels, you shouldn't get circular micro-scratches (on the substrate, at least) that act almost like a Fresnel lens. You'd need to be careful about scratch patterns in the ND material, of course. Again, I'd emphasize soft buffing wheels, and slow drill/mill speeds, to determine the ND material remove rate and pressures.
Additionally, to help minimize the neat pattern of strictly circular buffing, you could use something like a random-orbit sander motion to introduce more-or-less random motion in the buffing disc.
