ML-Benchmark prediction submission: ParamDiffusion-orog

[MNLR]

Institution ID

LSCE-IPSL

Institution Name

Laboratoire des Sciences du Climat et de l’Environnement (LSCE), Institut Pierre-Simon Laplace (IPSL)

Emulator identifier (emulator_id)

ParamDiffusion-orog

Emulator Description

Diffusion Model, using a residual approach on the ParamUNET model. 2 separate models for each variable. The model does use orography information. The predictors are standardised by channel. Precipitation has been applied a log transformation and standardised. Temperature has been standardised. No bias adjustment. Training on historical and future experiments. This is a generative model.

Hardware and Training Details

One node with a single Nvidia V100 GPU (32Gb). Training time (per experiment and variable): ~ 120 minutes (total).
Inference time: 10 minutes per year.

Stochastic/Probabilistic Output

yes

Reference URL

In preparation

Repository for Reproducibility

In preparation

Additional Notes

This model uses as background the ParamUNET model

[jgonzalezab]

Please fill out this form so we can give you access to our cluster to transfer your submission.