Overall, the aim here is learn how to create a machine learning emulator for access-om2 models. The hope is that an emulator will allow us to generate ensembles of simulated output based on a high resolution model data, at a much cheaper cost than running an ensemble of the model itself. The primary benefit being that it will allow us to better distinguish internal from forced variability in our model simulation results.
Create latent space for vertically integrated ocean heat content and net surface heat fluxes. Follow the pyearth tools autoencoder_example tutorial. See here for details: #7
Emulate SST from ACCESS-CM2 (using SAT and wind stress as inputs). Essentially reproduce some results from Dheeshjith et al. 2024 (https://arxiv.org/abs/2405.18585). See here for details: #1 (comment) Regrid: 1 deg om2, 1 deg global since om2 has 1/3deg near the equator and 1 deg at poles to resolve the undercurrents.
Redo Aim 1, but using ACCESS-OM2-01 ocean data and future atmosphere from Qian or Hannahs future warming runs. See here: #2 (comment)
Since emulating SST from SAT doesn't seem that challenging, we would like to try to autoregressively emulate ACCESS-OM2-1’s vertically integrated ocean heat content evolution given surface forcing (basically, emulate Huguenin et al. 2022; https://www.nature.com/articles/s41467-022-32540-5 Nat Comms.) See here: #3 (comment)