Difference between your forward and backward trajectories

[HatYaz]

What causes trajectory divergence between forward and backward OceanDrift runs when using identical forcing parameters?

Note.
Drift simulations with identical parameters (2.5% windage and Stokes drift)

Thanks

[knutfrode]

Hi, can you provide a minimalistic example to illustrate your question?

Diffusivity and numerical errors would generally give differences between forward and backward runs.
Here is an illustration of such numerical errors: https://opendrift.github.io/gallery/example_backandforth.html

[Angie-P]

Good afternoon,
I’m joining this conversation because I encountered the same problem and would like to share my experience and ask for your advice.

Here’s what I did:

• I released 100,000 particles backward in time from a point in the Venice Lagoon (with a radius of 10).
• I recorded the final positions of the particles.
• I then released particles forward in time from these positions, exactly as shown in the backandforth.py example you mentioned.
• Finally, I counted how many particles reached the original point at the end of the forward simulation.

My setup was as follows:

• shyfem reader (unstructured grid)
• OceanDrift as the OpenDrift model
• no diffusivity
• both Euler and Runge-Kutta schemes to compare results

The outcome was that no particles reached the original point, and the closest particles ended up about 1 km away.

I hypothesized that this might be due to the explicit advection scheme: the particle is advected using the velocity of the closest node. In a forward step, the particle moves from A to B using the velocity at the node closest to A. In a backward step from B, the particle uses the velocity at the node closest to B. If these velocities differ in direction, the particle doesn’t necessarily return to A. Do you think this hypothesis makes sense?

My goal is to identify the possible origin area of eDNA particles sampled at a site. My idea was to:

1. Release particles backward in time from the sampling site.
2. Draw a polygon around their final positions.
3. Build a uniform grid inside this polygon and release particles forward in time.
4. Count which particles reach the sampling site and generate a heatmap.

However, if backward and forward trajectories don’t match at all, I suppose this approach may not be feasible.

I was wondering whether using an implicit advection scheme could, in theory, help address this issue.
If so, I’d be interested to know whether such an approach has been considered before, and if there are known reasons why it might be difficult or impractical to implement.

Any suggestions or advice would be greatly appreciated.

Thank you very much for your time,
Angelica