We currently output snapshots every dz=0.25, but have we tested that we have converged here? Increasing the number of snapshots would increase the size of our training set, and this might translate into improved predictions. Of course, this will saturate at some small value of dz, but I don't think we have checked where this happens.
A trivial test would be to setup an emulator that only has every other snapshot, and compare predictions with the current emulator. If we see very similar performance level, we could save quite a lot of disk space by outputing fewer snapshots...
We could also imaging defining the outputs in log(1+z) instead of dz, that would mean dz twice larger at z=5 than at z=2.
We currently output snapshots every dz=0.25, but have we tested that we have converged here? Increasing the number of snapshots would increase the size of our training set, and this might translate into improved predictions. Of course, this will saturate at some small value of dz, but I don't think we have checked where this happens.
A trivial test would be to setup an emulator that only has every other snapshot, and compare predictions with the current emulator. If we see very similar performance level, we could save quite a lot of disk space by outputing fewer snapshots...
We could also imaging defining the outputs in log(1+z) instead of dz, that would mean dz twice larger at z=5 than at z=2.