horizontal diffusivity with backward simulations

[Angie-P]

Dear Opendrift community,
I'm doing backward releases of particles (10000) from a small area (I tried with 1, 10, 30 m of radius) using an unstructured grid reader from shyfem, and I got that all the particles travel together, or at most they divide in a small number of groups during the simulation. The reason why I put a small radius for the seeding is that the aim is to simulate the possible origin of particles that I could collect with a sampling.
I thought about the possible reasons for this behaviour and my hypothesis is that since the resolution of the grid is greater than the radius that I put for the simulation, and in the shyfem.py code opendrift seems to assign to each particle the velocity of the nearest node, the nearest node in this case is the same for all the particles, and therefore they are subject to the same velocity vector. I then thought to add some horizontal diffusivity to add a caotic and more randomic component in order to obtain more possible trajectories, but since it is a backward release I'm not sure it makes sense from a physic point of view to add a diffusivity component in this case, and I wanted to kindly ask your opinion about it.
Thank you for the attention,
Angelica

[Angie-P]

does anyone have an opinion about this topic? Any suggestion would be very helpful, thank you a lot

[knutfrode]

Hi,
Radii of 1, 10 and 30m are quite small, so you consider using larger radii also.
I am not using unstructured models myself, but yes, I believe those readers use nearest neighbour/pixel, and no spatial interpolation.
That could be updated by someone who have time and interest.

To me it makes sense to use diffusivity also for backwards simulations, with the interpretation that this is to account for the unresolved uncertainty.

[Angie-P]

Thank you for your answer!
I'd like to kindly ask you, or anyone from the Opendrift community who wants to join the conversation, how you choose the right value of diffusivity to add on the simulations, and even if it is the case to add it or not. I'm referring with this question not only to the backward simulations but also to the forward ones.
I've done some trials to test the sensitivities of the trajectories to this parameter, but then, after excluding extreme values, I still don't know which can be the more correct value to assign, and in the literature there seems not to be a clear criteria to follow to select it.
Any suggestion is very welcome, thank you again

[knutfrode]

Indeed, and unfortunately, I think there is as little consensus on "correct diffusivity" within OpenDrift community as within literature in general.
The "correct" value would also depend on the actual ocean model you are using, and the scale that it actually resolves (which is not necessarily the same as the pixel sizes).