We currently assume a constant scalar background structure. The inter-order residuals are now among the most conspicuous features when training is allowed to proceed indefinitely. Hooray, that's a good problem to have since the out-of-slit background is the least important nuisance feature remaining! In fact it's so unimportant we can almost just ignore it.
But if we wanted to, we could fit a more flexible model like a polynomial surface in x and y, or we could consider a GP model for the inter-order scattered light. The inter-order background mostly just affects the zero-point if we were trying to determine what the continuum sky background emission is---mostly of academic interest---or it would become important if we were observing extended objects that fill the slit, like a solar system planet like Jupiter. Training a GP is tricky because we would have two tasks: An out-of-slit GP model and an in-slit GP model. That's fine, we just add the losses, and backpropagate through both GPs, but in practice sounds tiresome, and a bit of overkill.
The problem with a more flexible model like a polynomial surface in x and y is that it can become too flexible and over-fit genuine structure in the sky or along a cluster of bright sky lines. So we almost want to mask the sky traces regardless. Another option could be to pre-process the background by masking all of the echelle traces. That's not a bad idea.
We currently assume a constant scalar background structure. The inter-order residuals are now among the most conspicuous features when training is allowed to proceed indefinitely. Hooray, that's a good problem to have since the out-of-slit background is the least important nuisance feature remaining! In fact it's so unimportant we can almost just ignore it.
But if we wanted to, we could fit a more flexible model like a polynomial surface in
xandy, or we could consider a GP model for the inter-order scattered light. The inter-order background mostly just affects the zero-point if we were trying to determine what the continuum sky background emission is---mostly of academic interest---or it would become important if we were observing extended objects that fill the slit, like a solar system planet like Jupiter. Training a GP is tricky because we would have two tasks: An out-of-slit GP model and an in-slit GP model. That's fine, we just add the losses, and backpropagate through both GPs, but in practice sounds tiresome, and a bit of overkill.The problem with a more flexible model like a polynomial surface in
xandyis that it can become too flexible and over-fit genuine structure in the sky or along a cluster of bright sky lines. So we almost want to mask the sky traces regardless. Another option could be to pre-process the background by masking all of the echelle traces. That's not a bad idea.