I have an idea for an ML emulator for the global ocean that I am hoping to develop and use with @rmholmes, @PSpence and @navidcy . I know @edoyango has some availability to help the community on ML applications like these, and I was wondering what the process is to apply for some help from ACCESS-NRI in developing something like this.
I am specifically interested in making an ocean emulator that shows skill in future projections, something that has been hard to achieve for existing emulators in this space. Statistical downscaling and parameterisations could also be some low-hanging fruit to aim for.
But the key first step would be to build a prototype emulator, with a well-defined workflow for pre-treating input (gridded) datasets, applying the ML method (and adjusting hyperparameters as necessary) and outputting key diagnostics. I think this is where ACCESS-NRI support could really help.
Please let me know what the key steps are to get some support here. I am happy to put in an application or similar if there is lots of interest.
Hi Taimoor. Have you done a bit of literature review and identified the kind of model you might like to use? I believe there are some published results which might inform the kind of direction you could go from a model architecture perspective. Have you determined which data sets you want to use?
Yes, I have done a bit of research in this space already - I know some groups have used ConvNEXT/UNet to produce emulators for the ocean, but I am interested in trying reinforcement learning algorithms. I know less about these, but I suspect RL could be better performing at long-term future forecasts. As for datasets, as a first pass I was thinking of a simple validation model dataset, like the ACCESS-CM2 historical and SSP runs.
I was hoping we could involve @edoyango or @micael in this as I don’t really have as much expertise on the ML side, and @edoyango may have some time or capacity to help.
I think as a first step it would be really useful to organise a meeting, so that we can get a better understanding of your plans and what kind of support we could provide.