Institution ID
KIT
Institution Name
Karlsruhe Institute of Technology
Emulator identifier (emulator_id)
SpaGAN
Emulator Description
SpaGAN uses a UNet2D generator (HuggingFace diffusers) with channel depths [32, 64, 128, 256] and self-attention at the deepest level, paired with a convolutional discriminator, trained with a composite loss (L1 + MSE + GAN + diversity). The 15 coarse GCM predictor variables are bilinearly upsampled to 128×128 and normalized to [-1, 1], with sinusoidal day-of-year encoding applied as temporal conditioning and a stochastic noise channel appended; for precipitation, a log₁₀ transform is applied before scaling to [-1, 1]. An ensemble of 5 members is generated per time step via different noise samples, with a diversity loss encouraging spread; the discriminator updates twice per generator step with instance noise for regularization.
Hardware and Training Details
1 Node with 1 Nvidia Tesla V100 32GB GPU
Train: 2 minutes/epoch
Inference: 128 samples/second
Stochastic/Probabilistic Output
yes
Reference URL
https://doi.org/10.1038/s41612-025-01103-y but heavily modified
Repository for Reproducibility
https://github.com/jpolz/ml-benchmark-spategan
Additional Notes
no orography version (already submitted)
Institution ID
KIT
Institution Name
Karlsruhe Institute of Technology
Emulator identifier (
emulator_id)SpaGAN
Emulator Description
SpaGAN uses a UNet2D generator (HuggingFace diffusers) with channel depths [32, 64, 128, 256] and self-attention at the deepest level, paired with a convolutional discriminator, trained with a composite loss (L1 + MSE + GAN + diversity). The 15 coarse GCM predictor variables are bilinearly upsampled to 128×128 and normalized to [-1, 1], with sinusoidal day-of-year encoding applied as temporal conditioning and a stochastic noise channel appended; for precipitation, a log₁₀ transform is applied before scaling to [-1, 1]. An ensemble of 5 members is generated per time step via different noise samples, with a diversity loss encouraging spread; the discriminator updates twice per generator step with instance noise for regularization.
Hardware and Training Details
1 Node with 1 Nvidia Tesla V100 32GB GPU
Train: 2 minutes/epoch
Inference: 128 samples/second
Stochastic/Probabilistic Output
yes
Reference URL
https://doi.org/10.1038/s41612-025-01103-y but heavily modified
Repository for Reproducibility
https://github.com/jpolz/ml-benchmark-spategan
Additional Notes
no orography version (already submitted)