Model overview¶

The recommended system is a stack, not a choice between unrelated models:

Stage 1 — deterministic baseline: learn a physical correction around ERA5.
Stage 2 — residual diffusion: freeze Stage 1 and sample a signed correction around its output.

Start with the complete two-stage workflow.

Path	Output	Inputs	Primary use
Two-stage baseline + diffusion	frozen baseline + sampled residual	GEO, ERA5, derived context, masks, Stage 1 field	main reconstruction system
Deterministic baseline only	one dense wind field	GEO, ERA5, derived context, masks	interpretable control and Stage 1 checkpoint
Residual diffusion on ERA5	ERA5 + sampled residual	GEO, ERA5, derived context, masks	portable generative ablation without Stage 1
Absolute conditional diffusion	absolute wind sample	GEO, optional ERA5, derived context, masks	standalone research baseline

flowchart LR
  C[GEO + ERA5 + derived context] --> S1[Stage 1 deterministic baseline]
  S1 --> B[Committed wind field]
  C --> S2[Stage 2 residual diffusion]
  B --> S2
  S2 --> O[Baseline + sampled correction]

All paths use the same PairedDataModule, physical target conversion, masks, and storm-centric metrics. This keeps comparisons focused on model behavior rather than on different data plumbing.

Main articles¶

Two-stage baseline + diffusion — the central workflow, exact handoff, channel counts, and training order.
Stage 1 deterministic baseline — residual connection to ERA5, physical Huber loss, and zero-initialized head.
Stage 2 residual diffusion — signed residual transform, frozen baseline, guidance, and ensemble diagnostics.

Supporting and ablation articles¶

Standalone conditional diffusion — generate an absolute field directly from a noise latent and conditions.
Sampling — DDPM, DDIM, schedules, clipping, and reproducibility.
Diffusion in one page — a short conceptual primer.