Institution ID
IFCA
Institution Name
Instituto de Física de Cantabria (IFCA), CSIC-Universidad de Cantabria
Emulator identifier (emulator_id)
ViT-IFCAv1-orog
Emulator Description
This emulator is based on the Vision Transformer (ViT) architecture [1]. The large-scale predictors are partitioned into patches, embedded, and passed through a series of Transformer blocks. The resulting patch representations are then upscaled to match the resolution of the local-scale predictand. As loss function it uses the CRPS-Spectral [2]. The noise is inyected in the ViT through conditional layer normalization [3]. Orography is added to the final embeddings of the Transformer block.
Input data are standardized at the grid-point level using the mean and standard deviation of the predictors corresponding to each training experiment. Models are trained for a variable number of epochs, determined via an early-stopping strategy based on a randomly split validation dataset.
Hardware and Training Details
The model was trained on two V100 (32GB) GPUs using DDP. Each epoch takes approximately 30–40 seconds.
Stochastic/Probabilistic Output
yes
Reference URL
In preparation
Repository for Reproducibility
https://github.com/jgonzalezab/ViT_CORDEX-ML-Bench
Additional Notes
No response
Institution ID
IFCA
Institution Name
Instituto de Física de Cantabria (IFCA), CSIC-Universidad de Cantabria
Emulator identifier (
emulator_id)ViT-IFCAv1-orog
Emulator Description
This emulator is based on the Vision Transformer (ViT) architecture [1]. The large-scale predictors are partitioned into patches, embedded, and passed through a series of Transformer blocks. The resulting patch representations are then upscaled to match the resolution of the local-scale predictand. As loss function it uses the CRPS-Spectral [2]. The noise is inyected in the ViT through conditional layer normalization [3]. Orography is added to the final embeddings of the Transformer block.
Input data are standardized at the grid-point level using the mean and standard deviation of the predictors corresponding to each training experiment. Models are trained for a variable number of epochs, determined via an early-stopping strategy based on a randomly split validation dataset.
Hardware and Training Details
The model was trained on two V100 (32GB) GPUs using DDP. Each epoch takes approximately 30–40 seconds.
Stochastic/Probabilistic Output
yes
Reference URL
In preparation
Repository for Reproducibility
https://github.com/jgonzalezab/ViT_CORDEX-ML-Bench
Additional Notes
No response