The issue had been also solved in UFS SRW application. (see the PR) and I found MPAS has similar code structure of computing OA1/2/3/4 but the issue haven't been solved.
This is my 15-3km domain over Japan, Japan is near 3km resolution and 15km resolution over sea. The figure shows the OA3 result (please ignore some strange pixels, it's plotting issue), and you can see that there exist negative bias in OA3. When apply the same correction in UFS_UTIL, the negative bias gone!

Before correction.

After correction.

You can apply the patch here if you want to test it.
core_init_atmosphere.patch

Jason