Try to extend forecast time on gencast

[chicha1986]

Hello
I am trying to extend forecast time for gencast.
I run the gencast on local nvidia-A100 GPUs.
for the result, it have some noise for the forecast.
for example

the image is 1080 hour forecast.
white contour is MSLP.
shading is T2m.

Is there anyone have any think about the issue?

[LeeSingChanYau]

From my experimentation you can either do more autoregressive steps, or try timesteps of longer periods like 36h or 48h.
You can freely adjust the scale of each timestep, but if you want to use the pretrained weights you have to follow the number of timesteps they use as input.

[chicha1986]

I apologize for missing some details. I didn't train GenCast myself, meaning I used Google's pretrained weights and attempted autoregression up to 1080 hours.
The program process starts at hour 0, and for the next timestep, a 12-hour forecast is made. This process loops, generating forecasts until the 1080-hour forecast is reached. In total, I use 90 steps to reach the 1080-hour forecast. This is the approach I'm using.

[IlanPrice]

Hi there

We ourselves have not properly evaluated GenCast beyond 15 day lead times so I can’t say with confidence what results I would expect you to see after 45 days of rollout. It would be hard to weigh in on whether this is some bug in how the model is being rolled out or whether it is just a feature of GenCast trajectories after many steps — the latter is certainly possible.

We do know of work by some colleagues of ours which found that 1 degree GenCast rolled out ‘stably’ for ~100 days, for example they found that global average temperature and specific humidity at 850hPa, or Geopotential at 500hPa, do approximately track the ground truth over these 90-100 day rollouts. So I would suggest perhaps looking at a few other variables, e.g. q850, t850 and z500, and see how they do (globally averaged and also spatially disaggregated as you have done above). Maybe plotting these alongside reference ground truth could help understand things.

Hope this helps!

[chicha1986]

Thank you Ilan, that's very helpful information we can use to evaluate the weather in our specific region.
We will looking other variables in there, in my review on 0.25 degree model, the noise will not able denose from Tibetan Plateau surface variable, extend to other region,and to higher pressure level, I need to discussion with my collages is there happen in nwp models,.
We also run the 1deg model, but the resolution can not give us more information in the limit area, like small island extended weather forecast.
Moving forward, we'll also be focusing on the models' performance with typhoons, cold fronts
Thanks again for your response.
chicha