@yrrepy thank you for the extensive detective work! I did some more digging by myself. Here are my conclusions.
- The anisotropy issue
The issue was resolved, as you suggested, both by switching off the ww surface checks and also by shifting the ww mesh grid.
- The slowness/performance issue
With the anisotropy issue gone I could focus on the performance issue. And the underlying cause I think, as you suggested in the bug report, is the lack of checks of the WW at a certain amount of mean free paths. I confirmed this in two ways.
First: I set the nmfp parameter of the MCNP wwp card to 1000. MCNPs performance dropped and I basically got the performance match of analog simulations (OpenMC is still only about 80 % as fast as MCNP). Interestingly with increasing nmfp MCNP also showed some anisotropy of the results so I also bumped the WW grid and the issue went away (might report this to the MCNP forum). Figure shows OpenMC (default ww settings, shifted ww mesh) relative error on the left and MCNP (nmfp 1000, shifted ww mesh) on the right.
Second: I went ahead and added additional led and air cells to the OpenMC model to force nmfp behavior with surface crossing checks in OpenMC. And I also got to the 80 % of MCNP performance - excellent result! Figure shows OpenMC (default ww settings, shifted ww mesh, additional lead (every 0.5 cm) and additional air (every 5.5 cm) cells) relative error on the left and MCNP (everything default) on the right.
Conclusion
IMO an nmfp equivalent should be implemented in OpenMC as this will highly improve the performance of calculating radiation dose fields in air filed regions. A very common application I would assume.
One question remains - why is OpenMC (only) 80% as fast as MCNP with gamma only transport problems and should we address this?