Modelpublic

cs-labs/experiment-5-v1-interpolation-model

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

Model Card

Experiment 5 v1 — Interpolation DiT (Flow Matching)

Model

3D Diffusion Transformer for atmospheric spatial interpolation using Flow Matching. Takes coarse atmospheric state and generates fine-grained local atmospheric fields.

Station

—Target: RKSI (Incheon International Airport, South Korea)
—Coordinates: 37.4692°N, 126.4505°E

Training Results (Best Epoch 7 / 12)

—Val T2 MAE: 0.9937 K
—Val T2 Bias: 0.9791 K
—Val FM Loss: 0.2880
—LR at best: 8.652e-05

Training Curve

Epoch	Val FM Loss	T2 MAE (K)	T2 Bias (K)	LR
1	1.1642	1.1759	0.5837	1.998e-04
2	0.9870	1.2539	0.9651	1.960e-04
3	0.7686	1.4457	1.3613	1.843e-04
4	0.5216	1.4016	1.3781	1.657e-04
5	0.3710	1.5597	1.5510	1.420e-04
6	0.3259	1.6226	1.6055	1.148e-04
7	0.2880	0.9937	0.9791	8.652e-05

Hyperparameters

—Architecture: 3D DiT (Diffusion Transformer)
—Batch size: 16
—Optimizer: Muon (LR=2e-4, WD=5e-3) + AdamW (LR=2e-5)
—Warmup: 1 epoch (cosine decay to 1e-6)
—Total epochs: 12 (early stopped at 7 due to NCCL crash, best=epoch 7)
—Flow steps: 1000
—Loss: Charbonnier + 0.05 Gradient + 0.05 LPIPS
—EMA decay: 0.9995
—Early stopping patience: 6

Data

—Training data: ERA5 atmospheric reanalysis (2005-2014)
—Grid: 10°x10° around RKSI at 0.25° resolution
—Variables: 7 atmospheric variables (T2m, U10, V10, SP, T850, U500, V500)

Hardware

—8x NVIDIA B200 GPUs (DDP via Accelerate)
—Training time: ~7.5 hours