Data resolution of the data-driven models

[SauryChen]

Hi, thanks for providing the dataset and benchmark for the S2S tasks! I am wondering about the data resolution you used during retraining the data-driven models and calculating the metrics (PW, GC and FCN2). Do you directly use the 1.5 deg resolution data to retrain the models, or do you use the 0.25 deg and intepolate the results to 1.5 degrees?

[juannat7]

Hi, thanks for providing the dataset and benchmark for the S2S tasks! I am wondering about the data resolution you used during retraining the data-driven models and calculating the metrics (PW, GC and FCN2). Do you directly use the 1.5 deg resolution data to retrain the models, or do you use the 0.25 deg and intepolate the results to 1.5 degrees?

Hi, thank you for your question. We did not retrain these models, and interpolate the results to the native 1.5-degree resolution of the center's subseasonal models. If you want to get the original 0.25 degree results, you can simply switch off the post-processing steps in the script we released here: https://github.com/leap-stc/ChaosBench/blob/main/scripts/process_sota.py