YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Experiment 4 v1 — Forecast Model (VAE + DiT)
Overview
Weather forecasting model for RKSI (Incheon Airport, Seoul) daily high temperature prediction. Combines a 3D Causal VAE for atmospheric field compression with a HunyuanVideo-style DiT for 6-hourly forecasting via flow matching.
Models
3D Causal VAE (247.3M params)
- Input: 23-channel 304x304 atmospheric fields
- Compression: 4x spatial (304->76), 2x temporal
- Train loss: 0.804, val station T2 error: 0.135
- Calibrated scaling factor: 0.0947
Forecast DiT (560.1M params)
- Architecture: hidden=1024, 16 heads, 8 double + 16 single blocks
- Best T2 MAE at RKSI: 2.24 K (epoch 6/8)
- Optimizer: Muon (LR=4e-4) + AdamW
- Flow matching with Charbonnier loss
Training Config
- VAE: 1 epoch, batch=1, seq_len=4, LR=5e-5
- DiT: 8 epochs, batch=16 x 8 GPU, Muon LR=4e-4, warmup=2 epochs
- Hardware: 8x NVIDIA B200
- Data: 2015-2019 (43,812 triplets)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support