search

bendudson / hermes-3

Multifluid drift reduced fluid model
GNU General Public License v3.0
37 stars 18 forks source link

PI controller for 2D/3D simulations #166

Open mikekryjak opened 1 year ago

mikekryjak commented 1 year ago

For a valid comparison between models and simulations, you often tune the recycling fraction (or pump albedo) to a given upstream separatrix density. With my current ST40 cases it can take 12 hours to know where one particular recycling fraction setting will take it and it can take many steps to get to the desired number (let's say about 5 on average). This gets very challenging for high densities where the recycling fraction becomes a lot more sensitive. In practice it means that one simulation takes several days and a lot of user input to get right.

I initially thought that the controller would be too tricky to tune especially considering the nonlinearity of the effect of the recycling fraction on the particle balance, but @bshanahan mentioned that something like this was implemented in Hermes-2.

@bendudson @hahahasan do you have any experiences from Hermes-2 on this? I see two main challenges: it takes a long time for a change in recycling fraction to have an effect, and the impact of the fraction will be highly nonlinear since the domain particle count goes to infinity as the fraction approaches 100%.

mikekryjak commented 10 months ago

This is pretty complicated. One idea is to control for 1/recycle_frac to do something towards linearising the actuator. I've had a discussion with Gijs Derks who works on the code DIV1D and on MAST-U detachment control, and he suggested that we should first determine the system response by prescribing a sinusoidal oscillation in the control actuator and observing the results. This would allow us to figure out how to linearise the problem (especially if we do small order oscillations). It may also show that we need to include some sort of delay in the scheme to account for the fact that any change in recycle fraction takes quite some time to filter through the simulation.

This challenge is likely why SOLPS-ITER never implemented a controller on the recycle fraction or pump albedo and only has one for the puff or core source. I have also confirmed that Hermes-3 only had a controller on the core source.