LonghaoWang commited on
Commit
09af8e7
·
verified ·
1 Parent(s): 57e06a2

Sync from GitHub 67168d0

Browse files
README.md CHANGED
@@ -17,14 +17,14 @@ tags:
17
 
18
  <div align="center">
19
 
20
- <img src="https://github.com/GISWLH/WeatherAI/blob/main/docs\assets\banner.png" alt="WeatherAI: 20 peer-reviewed weather and climate AI models in one readable PyTorch library" width="100%">
21
 
22
  **English** · [简体中文](https://github.com/GISWLH/WeatherAI/blob/main/README_zh.md) · [日本語](https://github.com/GISWLH/WeatherAI/blob/main/README_ja.md)
23
 
24
  [![License: MIT](https://img.shields.io/badge/License-MIT-0EA5E9.svg)](https://github.com/GISWLH/WeatherAI/blob/main/LICENSE)
25
  [![Python 3.9+](https://img.shields.io/badge/python-3.9%2B-1D4ED8.svg)](https://github.com/GISWLH/WeatherAI/blob/main/pyproject.toml)
26
  [![PyTorch](https://img.shields.io/badge/PyTorch-native-EE4C2C.svg?logo=pytorch&logoColor=white)](https://pytorch.org)
27
- [![JAX](https://img.shields.io/badge/JAX-supported-A855F7.svg)](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\neuralgcm.md)
28
  [![Models](https://img.shields.io/badge/models-20-14B8A6.svg)](#model-zoo)
29
  [![HF Space](https://img.shields.io/badge/%F0%9F%A4%97%20Space-GPU%20tests-FFD21E.svg)](https://huggingface.co/spaces/LonghaoWang/weatherai-graphcast-smoke)
30
  [![HF Hub](https://img.shields.io/badge/%F0%9F%A4%97%20Hub-model%20card-FFD21E.svg)](https://huggingface.co/LonghaoWang/WeatherAI)
@@ -38,10 +38,11 @@ Load the official weights, compare with the original, fine-tune, or start from a
38
 
39
  ## ✨ Highlights
40
 
41
- **19 Oct 2026 research update:** [PBC probabilistic S2S calibration, GenFocal two-stage
42
- lite downscaling, and gray-swan translocation diagnostics](https://github.com/GISWLH/WeatherAI/blob/main/docs\research_20261019.md)
43
- are importable by name and integrated with WeatherPre. All three passed the
44
- [synthetic ZeroGPU smoke](https://github.com/GISWLH/WeatherAI/blob/main/docs\results\research_20261019_zerogpu.json).
 
45
  PBC and the diagnostic wrapper extend the existing 20-model zoo; they are not
46
  two additional pretrained weather forecasters.
47
 
@@ -85,55 +86,55 @@ Each **docs** link opens a one-page summary: usage, what was verified, and what
85
 
86
  | Model | Import | Notes |
87
  |-------|--------|-------|
88
- | **Pangu-Weather** | `from weatherai.models import Pangu, Pangu_lite` | 3-D Earth-specific transformer · *Nature* 2023 · [paper](https://doi.org/10.1038/s41586-023-06185-3) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\pangu.md) · 🪶 |
89
- | **FuXi** | `from weatherai.models import FuXi` | Cube embedding + Swin-V2 U-Transformer · *npj Clim. Atmos. Sci.* 2023 · [paper](https://doi.org/10.1038/s41612-023-00512-1) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi.md) |
90
- | **FengWu** | `from weatherai.models import FengWu, FengWu_lite` | Multi-modal encode–fuse–decode · *Commun. Earth Environ.* 2025 · [paper](https://doi.org/10.1038/s43247-025-02502-y) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fengwu.md) · 🪶 |
91
- | **GraphCast** | `from weatherai.models import GraphCast, GraphCast_lite` | Grid ↔ icosahedral-mesh GNN, stage-wise parity vs JAX and PhysicsNeMo · *Science* 2023 · [paper](https://doi.org/10.1126/science.adi2336) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\graphcast.md) · 🎯 ⚡ 🪶 |
92
- | **Aurora** | `from weatherai.models import Aurora, Aurora_lite, Aurora_small` | Perceiver + Swin-3D foundation model · *Nature* 2025 · [paper](https://doi.org/10.1038/s41586-025-09005-y) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\aurora.md) · 🔓 🎯 ⚡ 🪶 |
93
- | **NeuralGCM** | `from weatherai.models.neuralgcm import NeuralGCMLearnedComponents, train` | Differentiable dynamical core + learned physics; training in JAX · *Nature* 2024 · [paper](https://doi.org/10.1038/s41586-024-07744-y) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\neuralgcm.md) · 🔓 🎯 ⚡ 🪶 |
94
- | **Aardvark Weather** | `from weatherai.models.aardvark import AardvarkE2E, load_official_encoder` | End-to-end forecasting from raw observations · *Nature* 2025 · [paper](https://doi.org/10.1038/s41586-025-08897-0) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\aardvark.md) · 🔓 🎯 ⚡ 🪶 |
95
 
96
  ### 🎲 Ensemble & generative
97
 
98
  | Model | Import | Notes |
99
  |-------|--------|-------|
100
- | **GenCast** | `from weatherai.models.gencast import GenCast_lite, GenCastDenoiser_from_official` | Diffusion ensemble: native denoiser + DPM-Solver++ sampler · *Nature* 2025 · [paper](https://doi.org/10.1038/s41586-024-08252-9) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\gencast.md) · 🔓 🎯 ⚡ 🪶 |
101
- | **FuXi-ENS** | `from weatherai.models.fuxi_ens import FuXiENS, FuXiENS_lite, load_official` | 2.5 B-parameter latent-noise ensemble, rebuilt from ONNX · *Sci. Adv.* 2025 · [paper](https://doi.org/10.1126/sciadv.adu2854) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi_ens.md) · 🔓 🎯 ⚡ 🪶 |
102
- | **ArchesWeather / Gen** | `from weatherai.models.arches import ArchesWeather_lite, ArchesWeatherGen_lite, convert` | 3-D attention forecaster + flow-matching generator · *Sci. Adv.* 2026 · [paper](https://doi.org/10.1126/sciadv.adx2372) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\arches.md) · 🔓 🎯 ⚡ 🪶 |
103
- | **NeuralGCM precipitation** | `from weatherai.models.neuralgcm import precip` | Stochastic precipitation checkpoints + trainable loss (JAX) · *Sci. Adv.* 2026 · [paper](https://doi.org/10.1126/sciadv.adv6891) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\neuralgcm_precip.md) · 🔓 ⚡ |
104
 
105
  ### 🌊 Sub-seasonal to seasonal, climate & ocean
106
 
107
  | Model | Import | Notes |
108
  |-------|--------|-------|
109
- | **CAS-Canglong** 🆕 | `from weatherai.models.canglong import Canglong, Canglong_lite, load_official` | 3-D Earth-specific transformer for 1–3-month global SST · *Adv. Sci.* · [paper](https://doi.org/10.1002/advs.78271) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\canglong.md) · 🔓 🎯 ⚡ 🪶 |
110
- | **FuXi-S2S** | `from weatherai.models.fuxi_s2s.convert import load_official` | Sub-seasonal Swin-V2 ensemble (1.04 B), rebuilt from ONNX · *Nat. Commun.* 2024 · [paper](https://doi.org/10.1038/s41467-024-50714-1) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi_s2s.md) · 🔓 🎯 ⚡ 🪶 |
111
- | **UniCM** | `from weatherai.models.unicm import UniCM, UniCMConfig` | Coupled climate-mode transformer (ENSO and other modes) · *Nat. Mach. Intell.* 2026 · [paper](https://doi.org/10.1038/s42256-026-01245-5) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\unicm.md) · 🎯 ⚡ 🪶 |
112
- | **ORCA-DL** | `from weatherai.models.orca_dl import ORCADL, load_official` | Global 3-D ocean forecaster, monthly steps · *Sci. Adv.* 2025 · [paper](https://doi.org/10.1126/sciadv.adu2488) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\orca_dl.md) · 🔓 🎯 ⚡ |
113
- | **ACE2-ERA5** | `from weatherai.models.ace2 import ACE2, load_official` | Spherical Fourier neural operator climate emulator with conservation · *npj Clim. Atmos. Sci.* 2025 · [paper](https://doi.org/10.1038/s41612-025-01090-0) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\ace2.md) · 🔓 🎯 ⚡ |
114
- | **GenFocal** | `from weatherai.models.genfocal import GenFocalNet, reflow_loss, sample_flow` | Generative debiasing flow (lite core) · *Nat. Mach. Intell.* 2026 · [paper](https://doi.org/10.1038/s42256-026-01308-7) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\genfocal.md) · 🎯 ⚡ 🪶 |
115
 
116
  ### 🌀 Regional & tropical cyclones
117
 
118
  | Model | Import | Notes |
119
  |-------|--------|-------|
120
- | **StormCast** | `from weatherai.models.stormcast import StormCast, StormCast_lite, load_official` | 3 km regional (HRRR) regression U-Net + EDM diffusion · *Sci. Adv.* 2026 · [paper](https://doi.org/10.1126/sciadv.adv0423) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\stormcast.md) · 🔓 🎯 ⚡ 🪶 |
121
- | **TropiCycloneNet** | `from weatherai.models.tropicyclonenet import TCNM, load_official` | Multimodal tropical-cyclone track + intensity · *Nat. Commun.* 2025 · [paper](https://doi.org/10.1038/s41467-025-61087-4) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\tropicyclonenet.md) · 🔓 🎯 ⚡ |
122
- | **WeatherNext Cyclones** | `from weatherai.models.weathernext_cyclones import WeatherNextCyclonesNet_from_official, WeatherNextCyclonesNative_lite` | Functional generative network for cyclones (Mini) · *Nature* 2026 · [paper](https://doi.org/10.1038/s41586-026-10953-2) · [docs](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\weathernext_cyclones.md) · 🔓 🎯 ⚡ 🪶 |
123
 
124
- More: [all model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · [verification log](https://github.com/GISWLH/WeatherAI/blob/main/docs\model_status.md) · [roadmap & planned models](https://github.com/GISWLH/WeatherAI/blob/main/docs\roadmap.md)
125
 
126
  ## 🔬 What "verified" means
127
 
128
  For each port, the unchanged official code (PyTorch, JAX or onnxruntime) and the native module receive the same weights
129
  and inputs, and their outputs are compared; the scripts are in [`scripts/`](https://github.com/GISWLH/WeatherAI/tree/main/scripts) and the numbers in
130
- [`docs/results/`](https://github.com/GISWLH/WeatherAI/tree/main/docs\results). GPU runs use [`scripts/gpu_smoke.py`](https://github.com/GISWLH/WeatherAI/blob/main/scripts\gpu_smoke.py) on the Hugging Face Space.
131
  These are implementation checks, not forecast-skill claims; each model page lists what was **not** verified.
132
 
133
  ## 🤗 Hugging Face
134
 
135
  - **Space** — [LonghaoWang/weatherai-graphcast-smoke](https://huggingface.co/spaces/LonghaoWang/weatherai-graphcast-smoke): ZeroGPU smoke tests for every model.
136
- - **Hub** — [LonghaoWang/WeatherAI](https://huggingface.co/LonghaoWang/WeatherAI): model card and model pages, synced from this repository by [`.github/workflows/sync_hf.yml`](https://github.com/GISWLH/WeatherAI/blob/main/.github\workflows\sync_hf.yml).
137
  - Official weights stay with their authors and keep their licences (Apache-2.0, MIT, CC-BY, CC-BY-NC(-ND), …): WeatherAI downloads them on demand and never redistributes them.
138
 
139
  ## 🧪 Tests
@@ -169,16 +170,16 @@ script against the official code, unit tests and a page in `docs/models/`. If We
169
  depends on data, training and weights, and nothing here replaces an operational forecasting system.
170
 
171
  <!-- readme-visual-refresh:artwork-source -->
172
- [Editable banner artwork · Excalidraw](https://github.com/GISWLH/WeatherAI/blob/main/assets\readme\banner.excalidraw)
173
  <!-- /readme-visual-refresh:artwork-source -->
174
 
175
 
176
  <!-- readme-visual-refresh:execution -->
177
  ## Demonstration captured on EHPC · 17 October 2026
178
 
179
- ![New visualization of the repository's existing parity-check JSON. This displays a recorded historical check; it is not a new training or model-validation run.](https://github.com/GISWLH/WeatherAI/blob/main/assets\readme\execution-result.png)
180
 
181
  New visualization of the repository's existing parity-check JSON. This displays a recorded historical check; it is not a new training or model-validation run.
182
 
183
- [Execution record and checksums](https://github.com/GISWLH/WeatherAI/blob/main/assets\readme\execution-record.json).
184
  <!-- /readme-visual-refresh:execution -->
 
17
 
18
  <div align="center">
19
 
20
+ <img src="assets/banner.png" alt="WeatherAI: 20 peer-reviewed weather and climate AI models in one readable PyTorch library" width="100%">
21
 
22
  **English** · [简体中文](https://github.com/GISWLH/WeatherAI/blob/main/README_zh.md) · [日本語](https://github.com/GISWLH/WeatherAI/blob/main/README_ja.md)
23
 
24
  [![License: MIT](https://img.shields.io/badge/License-MIT-0EA5E9.svg)](https://github.com/GISWLH/WeatherAI/blob/main/LICENSE)
25
  [![Python 3.9+](https://img.shields.io/badge/python-3.9%2B-1D4ED8.svg)](https://github.com/GISWLH/WeatherAI/blob/main/pyproject.toml)
26
  [![PyTorch](https://img.shields.io/badge/PyTorch-native-EE4C2C.svg?logo=pytorch&logoColor=white)](https://pytorch.org)
27
+ [![JAX](https://img.shields.io/badge/JAX-supported-A855F7.svg)](models/neuralgcm.md)
28
  [![Models](https://img.shields.io/badge/models-20-14B8A6.svg)](#model-zoo)
29
  [![HF Space](https://img.shields.io/badge/%F0%9F%A4%97%20Space-GPU%20tests-FFD21E.svg)](https://huggingface.co/spaces/LonghaoWang/weatherai-graphcast-smoke)
30
  [![HF Hub](https://img.shields.io/badge/%F0%9F%A4%97%20Hub-model%20card-FFD21E.svg)](https://huggingface.co/LonghaoWang/WeatherAI)
 
38
 
39
  ## ✨ Highlights
40
 
41
+ **19 Oct 2026 research update:** [PBC probabilistic S2S calibration, GenFocal official pretrained
42
+ downscaling, and gray-swan translocation diagnostics](https://github.com/GISWLH/WeatherAI/blob/main/docs/research_20261019.md)
43
+ use pinned official code and published weights by default, with explicit separate lite variants.
44
+ [Actual execution reports](https://github.com/GISWLH/WeatherAI/blob/main/docs/research_20261019.md#actual-validation) distinguish pretrained
45
+ GPU inference from the earlier lite-only synthetic smoke. WeatherPre defaults to the official entries.
46
  PBC and the diagnostic wrapper extend the existing 20-model zoo; they are not
47
  two additional pretrained weather forecasters.
48
 
 
86
 
87
  | Model | Import | Notes |
88
  |-------|--------|-------|
89
+ | **Pangu-Weather** | `from weatherai.models import Pangu, Pangu_lite` | 3-D Earth-specific transformer · *Nature* 2023 · [paper](https://doi.org/10.1038/s41586-023-06185-3) · [docs](models/pangu.md) · 🪶 |
90
+ | **FuXi** | `from weatherai.models import FuXi` | Cube embedding + Swin-V2 U-Transformer · *npj Clim. Atmos. Sci.* 2023 · [paper](https://doi.org/10.1038/s41612-023-00512-1) · [docs](models/fuxi.md) |
91
+ | **FengWu** | `from weatherai.models import FengWu, FengWu_lite` | Multi-modal encode–fuse–decode · *Commun. Earth Environ.* 2025 · [paper](https://doi.org/10.1038/s43247-025-02502-y) · [docs](models/fengwu.md) · 🪶 |
92
+ | **GraphCast** | `from weatherai.models import GraphCast, GraphCast_lite` | Grid ↔ icosahedral-mesh GNN, stage-wise parity vs JAX and PhysicsNeMo · *Science* 2023 · [paper](https://doi.org/10.1126/science.adi2336) · [docs](models/graphcast.md) · 🎯 ⚡ 🪶 |
93
+ | **Aurora** | `from weatherai.models import Aurora, Aurora_lite, Aurora_small` | Perceiver + Swin-3D foundation model · *Nature* 2025 · [paper](https://doi.org/10.1038/s41586-025-09005-y) · [docs](models/aurora.md) · 🔓 🎯 ⚡ 🪶 |
94
+ | **NeuralGCM** | `from weatherai.models.neuralgcm import NeuralGCMLearnedComponents, train` | Differentiable dynamical core + learned physics; training in JAX · *Nature* 2024 · [paper](https://doi.org/10.1038/s41586-024-07744-y) · [docs](models/neuralgcm.md) · 🔓 🎯 ⚡ 🪶 |
95
+ | **Aardvark Weather** | `from weatherai.models.aardvark import AardvarkE2E, load_official_encoder` | End-to-end forecasting from raw observations · *Nature* 2025 · [paper](https://doi.org/10.1038/s41586-025-08897-0) · [docs](models/aardvark.md) · 🔓 🎯 ⚡ 🪶 |
96
 
97
  ### 🎲 Ensemble & generative
98
 
99
  | Model | Import | Notes |
100
  |-------|--------|-------|
101
+ | **GenCast** | `from weatherai.models.gencast import GenCast_lite, GenCastDenoiser_from_official` | Diffusion ensemble: native denoiser + DPM-Solver++ sampler · *Nature* 2025 · [paper](https://doi.org/10.1038/s41586-024-08252-9) · [docs](models/gencast.md) · 🔓 🎯 ⚡ 🪶 |
102
+ | **FuXi-ENS** | `from weatherai.models.fuxi_ens import FuXiENS, FuXiENS_lite, load_official` | 2.5 B-parameter latent-noise ensemble, rebuilt from ONNX · *Sci. Adv.* 2025 · [paper](https://doi.org/10.1126/sciadv.adu2854) · [docs](models/fuxi_ens.md) · 🔓 🎯 ⚡ 🪶 |
103
+ | **ArchesWeather / Gen** | `from weatherai.models.arches import ArchesWeather_lite, ArchesWeatherGen_lite, convert` | 3-D attention forecaster + flow-matching generator · *Sci. Adv.* 2026 · [paper](https://doi.org/10.1126/sciadv.adx2372) · [docs](models/arches.md) · 🔓 🎯 ⚡ 🪶 |
104
+ | **NeuralGCM precipitation** | `from weatherai.models.neuralgcm import precip` | Stochastic precipitation checkpoints + trainable loss (JAX) · *Sci. Adv.* 2026 · [paper](https://doi.org/10.1126/sciadv.adv6891) · [docs](models/neuralgcm_precip.md) · 🔓 ⚡ |
105
 
106
  ### 🌊 Sub-seasonal to seasonal, climate & ocean
107
 
108
  | Model | Import | Notes |
109
  |-------|--------|-------|
110
+ | **CAS-Canglong** 🆕 | `from weatherai.models.canglong import Canglong, Canglong_lite, load_official` | 3-D Earth-specific transformer for 1–3-month global SST · *Adv. Sci.* · [paper](https://doi.org/10.1002/advs.78271) · [docs](models/canglong.md) · 🔓 🎯 ⚡ 🪶 |
111
+ | **FuXi-S2S** | `from weatherai.models.fuxi_s2s.convert import load_official` | Sub-seasonal Swin-V2 ensemble (1.04 B), rebuilt from ONNX · *Nat. Commun.* 2024 · [paper](https://doi.org/10.1038/s41467-024-50714-1) · [docs](models/fuxi_s2s.md) · 🔓 🎯 ⚡ 🪶 |
112
+ | **UniCM** | `from weatherai.models.unicm import UniCM, UniCMConfig` | Coupled climate-mode transformer (ENSO and other modes) · *Nat. Mach. Intell.* 2026 · [paper](https://doi.org/10.1038/s42256-026-01245-5) · [docs](models/unicm.md) · 🎯 ⚡ 🪶 |
113
+ | **ORCA-DL** | `from weatherai.models.orca_dl import ORCADL, load_official` | Global 3-D ocean forecaster, monthly steps · *Sci. Adv.* 2025 · [paper](https://doi.org/10.1126/sciadv.adu2488) · [docs](models/orca_dl.md) · 🔓 🎯 ⚡ |
114
+ | **ACE2-ERA5** | `from weatherai.models.ace2 import ACE2, load_official` | Spherical Fourier neural operator climate emulator with conservation · *npj Clim. Atmos. Sci.* 2025 · [paper](https://doi.org/10.1038/s41612-025-01090-0) · [docs](models/ace2.md) · 🔓 🎯 ⚡ |
115
+ | **GenFocal** | `from weatherai import GenFocal, GenFocal_lite` | Official pretrained ReFlow + SR, separate lite · *Nat. Mach. Intell.* 2026 · [paper](https://doi.org/10.1038/s42256-026-01308-7) · [docs](models/genfocal.md) · 🎯 ⚡ 🪶 |
116
 
117
  ### 🌀 Regional & tropical cyclones
118
 
119
  | Model | Import | Notes |
120
  |-------|--------|-------|
121
+ | **StormCast** | `from weatherai.models.stormcast import StormCast, StormCast_lite, load_official` | 3 km regional (HRRR) regression U-Net + EDM diffusion · *Sci. Adv.* 2026 · [paper](https://doi.org/10.1126/sciadv.adv0423) · [docs](models/stormcast.md) · 🔓 🎯 ⚡ 🪶 |
122
+ | **TropiCycloneNet** | `from weatherai.models.tropicyclonenet import TCNM, load_official` | Multimodal tropical-cyclone track + intensity · *Nat. Commun.* 2025 · [paper](https://doi.org/10.1038/s41467-025-61087-4) · [docs](models/tropicyclonenet.md) · 🔓 🎯 ⚡ |
123
+ | **WeatherNext Cyclones** | `from weatherai.models.weathernext_cyclones import WeatherNextCyclonesNet_from_official, WeatherNextCyclonesNative_lite` | Functional generative network for cyclones (Mini) · *Nature* 2026 · [paper](https://doi.org/10.1038/s41586-026-10953-2) · [docs](models/weathernext_cyclones.md) · 🔓 🎯 ⚡ 🪶 |
124
 
125
+ More: [all model pages](models/README.md) · [verification log](https://github.com/GISWLH/WeatherAI/blob/main/docs/model_status.md) · [roadmap & planned models](https://github.com/GISWLH/WeatherAI/blob/main/docs/roadmap.md)
126
 
127
  ## 🔬 What "verified" means
128
 
129
  For each port, the unchanged official code (PyTorch, JAX or onnxruntime) and the native module receive the same weights
130
  and inputs, and their outputs are compared; the scripts are in [`scripts/`](https://github.com/GISWLH/WeatherAI/tree/main/scripts) and the numbers in
131
+ [`docs/results/`](https://github.com/GISWLH/WeatherAI/tree/main/docs/results). GPU runs use [`scripts/gpu_smoke.py`](https://github.com/GISWLH/WeatherAI/blob/main/scripts/gpu_smoke.py) on the Hugging Face Space.
132
  These are implementation checks, not forecast-skill claims; each model page lists what was **not** verified.
133
 
134
  ## 🤗 Hugging Face
135
 
136
  - **Space** — [LonghaoWang/weatherai-graphcast-smoke](https://huggingface.co/spaces/LonghaoWang/weatherai-graphcast-smoke): ZeroGPU smoke tests for every model.
137
+ - **Hub** — [LonghaoWang/WeatherAI](https://huggingface.co/LonghaoWang/WeatherAI): model card and model pages, synced from this repository by [`.github/workflows/sync_hf.yml`](https://github.com/GISWLH/WeatherAI/blob/main/.github/workflows/sync_hf.yml).
138
  - Official weights stay with their authors and keep their licences (Apache-2.0, MIT, CC-BY, CC-BY-NC(-ND), …): WeatherAI downloads them on demand and never redistributes them.
139
 
140
  ## 🧪 Tests
 
170
  depends on data, training and weights, and nothing here replaces an operational forecasting system.
171
 
172
  <!-- readme-visual-refresh:artwork-source -->
173
+ [Editable banner artwork · Excalidraw](https://github.com/GISWLH/WeatherAI/blob/main/assets/readme/banner.excalidraw)
174
  <!-- /readme-visual-refresh:artwork-source -->
175
 
176
 
177
  <!-- readme-visual-refresh:execution -->
178
  ## Demonstration captured on EHPC · 17 October 2026
179
 
180
+ ![New visualization of the repository's existing parity-check JSON. This displays a recorded historical check; it is not a new training or model-validation run.](https://github.com/GISWLH/WeatherAI/blob/main/assets/readme/execution-result.png)
181
 
182
  New visualization of the repository's existing parity-check JSON. This displays a recorded historical check; it is not a new training or model-validation run.
183
 
184
+ [Execution record and checksums](https://github.com/GISWLH/WeatherAI/blob/main/assets/readme/execution-record.json).
185
  <!-- /readme-visual-refresh:execution -->
models/README.md CHANGED
@@ -1,34 +1,34 @@
1
  # Model pages
2
 
3
- [← Model zoo](..\README.md) · [Verification log](https://github.com/GISWLH/WeatherAI/blob/main/docs\model_status.md) · [Roadmap](https://github.com/GISWLH/WeatherAI/blob/main/docs\roadmap.md)
4
 
5
  One page per model: how to use it, what was verified against the official code, and what was not. The same pages are
6
  published on the Hugging Face Hub ([LonghaoWang/WeatherAI](https://huggingface.co/LonghaoWang/WeatherAI)).
7
 
8
  | Group | Model | Page | Paper |
9
  |-------|-------|------|-------|
10
- | 🎯 Probabilistic S2S calibration | PBC | [pbc.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\pbc.md) | *Nature* 2026, [doi:10.1038/s41586-026-11126-x](https://doi.org/10.1038/s41586-026-11126-x) |
11
- | 🔬 Extreme-event diagnostics | Gray-swan translocation (existing backbones) | [translocation.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\translocation.md) | *Science Advances* 2026, [doi:10.1126/sciadv.ady9406](https://doi.org/10.1126/sciadv.ady9406) |
12
- | 🌍 Medium-range | Pangu-Weather | [pangu.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\pangu.md) | *Nature* 2023, [doi:10.1038/s41586-023-06185-3](https://doi.org/10.1038/s41586-023-06185-3) |
13
- | 🌍 Medium-range | FuXi | [fuxi.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi.md) | *npj Climate and Atmospheric Science* 2023, [doi:10.1038/s41612-023-00512-1](https://doi.org/10.1038/s41612-023-00512-1) |
14
- | 🌍 Medium-range | FengWu | [fengwu.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fengwu.md) | *Communications Earth & Environment* 2025, [doi:10.1038/s43247-025-02502-y](https://doi.org/10.1038/s43247-025-02502-y) |
15
- | 🌍 Medium-range | GraphCast | [graphcast.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\graphcast.md) | *Science* 2023, [doi:10.1126/science.adi2336](https://doi.org/10.1126/science.adi2336) |
16
- | 🌍 Medium-range | Aurora | [aurora.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\aurora.md) | *Nature* 2025, [doi:10.1038/s41586-025-09005-y](https://doi.org/10.1038/s41586-025-09005-y) |
17
- | 🌍 Medium-range | NeuralGCM | [neuralgcm.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\neuralgcm.md) | *Nature* 2024, [doi:10.1038/s41586-024-07744-y](https://doi.org/10.1038/s41586-024-07744-y) |
18
- | 🌍 Medium-range | Aardvark Weather | [aardvark.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\aardvark.md) | *Nature* 2025, [doi:10.1038/s41586-025-08897-0](https://doi.org/10.1038/s41586-025-08897-0) |
19
- | 🎲 Ensemble | GenCast | [gencast.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\gencast.md) | *Nature* 2025, [doi:10.1038/s41586-024-08252-9](https://doi.org/10.1038/s41586-024-08252-9) |
20
- | 🎲 Ensemble | FuXi-ENS | [fuxi_ens.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi_ens.md) | *Science Advances* 2025, [doi:10.1126/sciadv.adu2854](https://doi.org/10.1126/sciadv.adu2854) |
21
- | 🎲 Ensemble | ArchesWeather / ArchesWeatherGen | [arches.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\arches.md) | *Science Advances* 2026, [doi:10.1126/sciadv.adx2372](https://doi.org/10.1126/sciadv.adx2372) |
22
- | 🎲 Ensemble | NeuralGCM precipitation | [neuralgcm_precip.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\neuralgcm_precip.md) | *Science Advances* 2026, [doi:10.1126/sciadv.adv6891](https://doi.org/10.1126/sciadv.adv6891) |
23
- | 🌊 S2S / climate / ocean | CAS-Canglong 🆕 | [canglong.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\canglong.md) | *Advanced Science*, [doi:10.1002/advs.78271](https://doi.org/10.1002/advs.78271) |
24
- | 🌊 S2S / climate / ocean | FuXi-S2S | [fuxi_s2s.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi_s2s.md) | *Nature Communications* 2024, [doi:10.1038/s41467-024-50714-1](https://doi.org/10.1038/s41467-024-50714-1) |
25
- | 🌊 S2S / climate / ocean | UniCM | [unicm.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\unicm.md) | *Nature Machine Intelligence* 2026, [doi:10.1038/s42256-026-01245-5](https://doi.org/10.1038/s42256-026-01245-5) |
26
- | 🌊 S2S / climate / ocean | ORCA-DL | [orca_dl.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\orca_dl.md) | *Science Advances* 2025, [doi:10.1126/sciadv.adu2488](https://doi.org/10.1126/sciadv.adu2488) |
27
- | 🌊 S2S / climate / ocean | ACE2-ERA5 | [ace2.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\ace2.md) | *npj Climate and Atmospheric Science* 2025, [doi:10.1038/s41612-025-01090-0](https://doi.org/10.1038/s41612-025-01090-0) |
28
- | 🌊 S2S / climate / ocean | GenFocal | [genfocal.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\genfocal.md) | *Nature Machine Intelligence* 2026, [doi:10.1038/s42256-026-01308-7](https://doi.org/10.1038/s42256-026-01308-7) |
29
- | 🌀 Regional / cyclones | StormCast | [stormcast.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\stormcast.md) | *Science Advances* 2026, [doi:10.1126/sciadv.adv0423](https://doi.org/10.1126/sciadv.adv0423) |
30
- | 🌀 Regional / cyclones | TropiCycloneNet | [tropicyclonenet.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\tropicyclonenet.md) | *Nature Communications* 2025, [doi:10.1038/s41467-025-61087-4](https://doi.org/10.1038/s41467-025-61087-4) |
31
- | 🌀 Regional / cyclones | WeatherNext Cyclones | [weathernext_cyclones.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\weathernext_cyclones.md) | *Nature* 2026, [doi:10.1038/s41586-026-10953-2](https://doi.org/10.1038/s41586-026-10953-2) |
32
 
33
  Journal versions were checked against Crossref when they were added (CAS-Canglong: checked 2026-10-18, *Advanced Science*,
34
- published 15 October 2026, article e78271). Planned and blocked models (FuXi-DA, NowcastNet): [roadmap](https://github.com/GISWLH/WeatherAI/blob/main/docs\roadmap.md).
 
1
  # Model pages
2
 
3
+ [← Model zoo](../README.md) · [Verification log](https://github.com/GISWLH/WeatherAI/blob/main/docs/model_status.md) · [Roadmap](https://github.com/GISWLH/WeatherAI/blob/main/docs/roadmap.md)
4
 
5
  One page per model: how to use it, what was verified against the official code, and what was not. The same pages are
6
  published on the Hugging Face Hub ([LonghaoWang/WeatherAI](https://huggingface.co/LonghaoWang/WeatherAI)).
7
 
8
  | Group | Model | Page | Paper |
9
  |-------|-------|------|-------|
10
+ | 🎯 Probabilistic S2S calibration | PBC | [pbc.md](pbc.md) | *Nature* 2026, [doi:10.1038/s41586-026-11126-x](https://doi.org/10.1038/s41586-026-11126-x) |
11
+ | 🔬 Extreme-event diagnostics | Gray-swan translocation (existing backbones) | [translocation.md](translocation.md) | *Science Advances* 2026, [doi:10.1126/sciadv.ady9406](https://doi.org/10.1126/sciadv.ady9406) |
12
+ | 🌍 Medium-range | Pangu-Weather | [pangu.md](pangu.md) | *Nature* 2023, [doi:10.1038/s41586-023-06185-3](https://doi.org/10.1038/s41586-023-06185-3) |
13
+ | 🌍 Medium-range | FuXi | [fuxi.md](fuxi.md) | *npj Climate and Atmospheric Science* 2023, [doi:10.1038/s41612-023-00512-1](https://doi.org/10.1038/s41612-023-00512-1) |
14
+ | 🌍 Medium-range | FengWu | [fengwu.md](fengwu.md) | *Communications Earth & Environment* 2025, [doi:10.1038/s43247-025-02502-y](https://doi.org/10.1038/s43247-025-02502-y) |
15
+ | 🌍 Medium-range | GraphCast | [graphcast.md](graphcast.md) | *Science* 2023, [doi:10.1126/science.adi2336](https://doi.org/10.1126/science.adi2336) |
16
+ | 🌍 Medium-range | Aurora | [aurora.md](aurora.md) | *Nature* 2025, [doi:10.1038/s41586-025-09005-y](https://doi.org/10.1038/s41586-025-09005-y) |
17
+ | 🌍 Medium-range | NeuralGCM | [neuralgcm.md](neuralgcm.md) | *Nature* 2024, [doi:10.1038/s41586-024-07744-y](https://doi.org/10.1038/s41586-024-07744-y) |
18
+ | 🌍 Medium-range | Aardvark Weather | [aardvark.md](aardvark.md) | *Nature* 2025, [doi:10.1038/s41586-025-08897-0](https://doi.org/10.1038/s41586-025-08897-0) |
19
+ | 🎲 Ensemble | GenCast | [gencast.md](gencast.md) | *Nature* 2025, [doi:10.1038/s41586-024-08252-9](https://doi.org/10.1038/s41586-024-08252-9) |
20
+ | 🎲 Ensemble | FuXi-ENS | [fuxi_ens.md](fuxi_ens.md) | *Science Advances* 2025, [doi:10.1126/sciadv.adu2854](https://doi.org/10.1126/sciadv.adu2854) |
21
+ | 🎲 Ensemble | ArchesWeather / ArchesWeatherGen | [arches.md](arches.md) | *Science Advances* 2026, [doi:10.1126/sciadv.adx2372](https://doi.org/10.1126/sciadv.adx2372) |
22
+ | 🎲 Ensemble | NeuralGCM precipitation | [neuralgcm_precip.md](neuralgcm_precip.md) | *Science Advances* 2026, [doi:10.1126/sciadv.adv6891](https://doi.org/10.1126/sciadv.adv6891) |
23
+ | 🌊 S2S / climate / ocean | CAS-Canglong 🆕 | [canglong.md](canglong.md) | *Advanced Science*, [doi:10.1002/advs.78271](https://doi.org/10.1002/advs.78271) |
24
+ | 🌊 S2S / climate / ocean | FuXi-S2S | [fuxi_s2s.md](fuxi_s2s.md) | *Nature Communications* 2024, [doi:10.1038/s41467-024-50714-1](https://doi.org/10.1038/s41467-024-50714-1) |
25
+ | 🌊 S2S / climate / ocean | UniCM | [unicm.md](unicm.md) | *Nature Machine Intelligence* 2026, [doi:10.1038/s42256-026-01245-5](https://doi.org/10.1038/s42256-026-01245-5) |
26
+ | 🌊 S2S / climate / ocean | ORCA-DL | [orca_dl.md](orca_dl.md) | *Science Advances* 2025, [doi:10.1126/sciadv.adu2488](https://doi.org/10.1126/sciadv.adu2488) |
27
+ | 🌊 S2S / climate / ocean | ACE2-ERA5 | [ace2.md](ace2.md) | *npj Climate and Atmospheric Science* 2025, [doi:10.1038/s41612-025-01090-0](https://doi.org/10.1038/s41612-025-01090-0) |
28
+ | 🌊 S2S / climate / ocean | GenFocal | [genfocal.md](genfocal.md) | *Nature Machine Intelligence* 2026, [doi:10.1038/s42256-026-01308-7](https://doi.org/10.1038/s42256-026-01308-7) |
29
+ | 🌀 Regional / cyclones | StormCast | [stormcast.md](stormcast.md) | *Science Advances* 2026, [doi:10.1126/sciadv.adv0423](https://doi.org/10.1126/sciadv.adv0423) |
30
+ | 🌀 Regional / cyclones | TropiCycloneNet | [tropicyclonenet.md](tropicyclonenet.md) | *Nature Communications* 2025, [doi:10.1038/s41467-025-61087-4](https://doi.org/10.1038/s41467-025-61087-4) |
31
+ | 🌀 Regional / cyclones | WeatherNext Cyclones | [weathernext_cyclones.md](weathernext_cyclones.md) | *Nature* 2026, [doi:10.1038/s41586-026-10953-2](https://doi.org/10.1038/s41586-026-10953-2) |
32
 
33
  Journal versions were checked against Crossref when they were added (CAS-Canglong: checked 2026-10-18, *Advanced Science*,
34
+ published 15 October 2026, article e78271). Planned and blocked models (FuXi-DA, NowcastNet): [roadmap](https://github.com/GISWLH/WeatherAI/blob/main/docs/roadmap.md).
models/aardvark.md CHANGED
@@ -1,6 +1,6 @@
1
  # Aardvark Weather — native PyTorch encoder + processor + decoder
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature* 2025, [doi:10.1038/s41586-025-08897-0](https://doi.org/10.1038/s41586-025-08897-0)
4
 
5
  Aardvark Weather (*Nature* 2025; official code CC0) = observation **encoder** → **processor** → station **decoder**.
6
  `weatherai.models.aardvark` now contains explicit ports of all three (parameter names identical to the official ones):
@@ -33,7 +33,7 @@ station, forecast, init_state = AardvarkE2E(enc, [proc], dec).eval()(task, retur
33
  encoder initial state max |Δ| = 0.0 (atol 1e-5, rtol 1e-4); station decoder (tas, lt_1) max |Δ| = 0; E2E (encoder → 1 processor → decoder)
34
  station output max |Δ| = 2.9e-6, gridded forecast max relative error 8e-5 (values up to 1.2e5); processor on random input max |Δ| ≈ 2e-4
35
  (unchanged earlier check). Unit tests: set-conv vs naive loop, NaN handling, cylindrical conv shapes, parameter names, a training step on a tiny
36
- whole system. HF ZeroGPU: see [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\model_status.md).
37
  **Not implemented / not verified:** the official data loaders and training scripts (they depend on multi-TB local memmaps; the repo itself says they
38
  cannot be executed), the FiLM / attention options of the official U-Net (unused by the released weights), `e2e_finetuned` checkpoints (not loaded/tested),
39
  decoder checkpoints other than `tas/lt_1`, processors `forecast_2..10`, multi-step E2E, other station variables (`ws`), skill vs. observations.
 
1
  # Aardvark Weather — native PyTorch encoder + processor + decoder
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature* 2025, [doi:10.1038/s41586-025-08897-0](https://doi.org/10.1038/s41586-025-08897-0)
4
 
5
  Aardvark Weather (*Nature* 2025; official code CC0) = observation **encoder** → **processor** → station **decoder**.
6
  `weatherai.models.aardvark` now contains explicit ports of all three (parameter names identical to the official ones):
 
33
  encoder initial state max |Δ| = 0.0 (atol 1e-5, rtol 1e-4); station decoder (tas, lt_1) max |Δ| = 0; E2E (encoder → 1 processor → decoder)
34
  station output max |Δ| = 2.9e-6, gridded forecast max relative error 8e-5 (values up to 1.2e5); processor on random input max |Δ| ≈ 2e-4
35
  (unchanged earlier check). Unit tests: set-conv vs naive loop, NaN handling, cylindrical conv shapes, parameter names, a training step on a tiny
36
+ whole system. HF ZeroGPU: see [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/model_status.md).
37
  **Not implemented / not verified:** the official data loaders and training scripts (they depend on multi-TB local memmaps; the repo itself says they
38
  cannot be executed), the FiLM / attention options of the official U-Net (unused by the released weights), `e2e_finetuned` checkpoints (not loaded/tested),
39
  decoder checkpoints other than `tas/lt_1`, processors `forecast_2..10`, multi-step E2E, other station variables (`ws`), skill vs. observations.
models/ace2.md CHANGED
@@ -1,6 +1,6 @@
1
  # ACE2-ERA5 — native PyTorch spherical Fourier neural operator (official weights load; matches `fme` rollouts)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *npj Climate and Atmospheric Science* 2025, [doi:10.1038/s41612-025-01090-0](https://doi.org/10.1038/s41612-025-01090-0)
4
 
5
  `weatherai/models/ace2` re-implements Ai2's ACE2-ERA5 (180×360 Gauss–Legendre grid, 44 inputs → 50 outputs, 6 h step) without depending on `fme` or `torch-harmonics`:
6
  `sht.py` (real SHT/iSHT on Gauss–Legendre and equiangular grids), `sfno.py` (SFNO: 8 blocks, embed 384, instance norm, per-degree complex channel mixing = Driscoll–Healy convolution, inner linear skip + outer identity skip, big skip), `model.py` (the stepper around the network) and `convert.py`.
 
1
  # ACE2-ERA5 — native PyTorch spherical Fourier neural operator (official weights load; matches `fme` rollouts)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *npj Climate and Atmospheric Science* 2025, [doi:10.1038/s41612-025-01090-0](https://doi.org/10.1038/s41612-025-01090-0)
4
 
5
  `weatherai/models/ace2` re-implements Ai2's ACE2-ERA5 (180×360 Gauss–Legendre grid, 44 inputs → 50 outputs, 6 h step) without depending on `fme` or `torch-harmonics`:
6
  `sht.py` (real SHT/iSHT on Gauss–Legendre and equiangular grids), `sfno.py` (SFNO: 8 blocks, embed 384, instance norm, per-degree complex channel mixing = Driscoll–Healy convolution, inner linear skip + outer identity skip, big skip), `model.py` (the stepper around the network) and `convert.py`.
models/arches.md CHANGED
@@ -1,6 +1,6 @@
1
  # ArchesWeatherGen — native PyTorch 3D-attention forecaster + flow-matching generator (official weights load; parity with geoarches)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Science Advances* 2026, [doi:10.1126/sciadv.adx2372](https://doi.org/10.1126/sciadv.adx2372)
4
 
5
  `weatherai/models/arches` re-implements INRIA's ArchesWeather (deterministic, 24 h step, 1.5° = 121×240 grid, 6 surface + 6 levels×13 pressure-level variables) and ArchesWeatherGen
6
  (4 deterministic members embedded + a conditional flow-matching network that samples the residual). No `geoarches` dependency at runtime. Weights: HF [`gcouairon/ArchesWeather`](https://huggingface.co/gcouairon/ArchesWeather) (BSD; not committed) and normalisation stats from the geoarches repo (BSD-3, fetched from GitHub raw, not committed).
 
1
  # ArchesWeatherGen — native PyTorch 3D-attention forecaster + flow-matching generator (official weights load; parity with geoarches)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Science Advances* 2026, [doi:10.1126/sciadv.adx2372](https://doi.org/10.1126/sciadv.adx2372)
4
 
5
  `weatherai/models/arches` re-implements INRIA's ArchesWeather (deterministic, 24 h step, 1.5° = 121×240 grid, 6 surface + 6 levels×13 pressure-level variables) and ArchesWeatherGen
6
  (4 deterministic members embedded + a conditional flow-matching network that samples the residual). No `geoarches` dependency at runtime. Weights: HF [`gcouairon/ArchesWeather`](https://huggingface.co/gcouairon/ArchesWeather) (BSD; not committed) and normalisation stats from the geoarches repo (BSD-3, fetched from GitHub raw, not committed).
models/aurora.md CHANGED
@@ -1,6 +1,6 @@
1
  # Aurora — native PyTorch implementation
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature* 2025, [doi:10.1038/s41586-025-09005-y](https://doi.org/10.1038/s41586-025-09005-y)
4
 
5
  `weatherai.models.aurora` re-implements Aurora (*Nature* 2025; MIT,
6
  [microsoft/aurora](https://github.com/microsoft/aurora)) as explicit `nn.Module`s with readable `forward()`s,
@@ -41,7 +41,7 @@ small = Aurora_small(pretrained=True).eval() # or Aurora_small(checkpoint_path=
41
  `strict=True` for 3 configs (window (2,2,2)/(2,3,5)/(2,6,12), uneven depths, v3 lead-time embedding), grids
42
  16×32 … 32×64 incl. the odd-latitude crop (H%4==1) and patch grids that need merge-padding / split-cropping
43
  (25×44); the official `aurora-0.25-small-pretrained.ckpt` strict-loads into the native `Aurora_small` and
44
- reproduces the official output on 33×64 inputs. HF ZeroGPU results: see [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\model_status.md).
45
  Unit tests: `tests/models/aurora` (shapes, odd grids, merge/split + window helpers, param-name layout,
46
  25-step training demo, grads, parity).
47
  **Not verified / not included:** skill on real ERA5/HRES data (random inputs only), multi-step rollout;
 
1
  # Aurora — native PyTorch implementation
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature* 2025, [doi:10.1038/s41586-025-09005-y](https://doi.org/10.1038/s41586-025-09005-y)
4
 
5
  `weatherai.models.aurora` re-implements Aurora (*Nature* 2025; MIT,
6
  [microsoft/aurora](https://github.com/microsoft/aurora)) as explicit `nn.Module`s with readable `forward()`s,
 
41
  `strict=True` for 3 configs (window (2,2,2)/(2,3,5)/(2,6,12), uneven depths, v3 lead-time embedding), grids
42
  16×32 … 32×64 incl. the odd-latitude crop (H%4==1) and patch grids that need merge-padding / split-cropping
43
  (25×44); the official `aurora-0.25-small-pretrained.ckpt` strict-loads into the native `Aurora_small` and
44
+ reproduces the official output on 33×64 inputs. HF ZeroGPU results: see [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/model_status.md).
45
  Unit tests: `tests/models/aurora` (shapes, odd grids, merge/split + window helpers, param-name layout,
46
  25-step training demo, grads, parity).
47
  **Not verified / not included:** skill on real ERA5/HRES data (random inputs only), multi-step rollout;
models/canglong.md CHANGED
@@ -1,6 +1,6 @@
1
  # CAS-Canglong — native PyTorch, bit-exact vs the official code (S2S global SST)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Advanced Science*, "Improved Skills for Predicting Sub-Seasonal to Seasonal Global Sea Surface Temperature", [doi:10.1002/advs.78271](https://doi.org/10.1002/advs.78271) (preprint: arXiv:2409.05369)
4
 
5
  `weatherai/models/canglong` re-implements CAS-Canglong, a 3-D Earth-specific transformer (Pangu-style, U-shaped with a skip
6
  concatenation) that predicts global sea surface temperature 1–3 months ahead from 16 monthly frames of 8 ocean/atmosphere
@@ -43,7 +43,7 @@ model = load_official("canglong3_0005_600ep_base.tar") # figshare checkpoint (
43
  4. Loss: lead weights (1, 1.5, 2) multiply *inside* the square, and the correlation penalty box (named `nino` in the
44
  notebook, rows 288:336 × cols 480:640) covers 18°N–6°N, 120°E–160°E on the 90°N→90°S, 0→359.75°E grid.
45
  * **Released checkpoint on released data (GPU, fp32; `scripts/canglong_official_checkpoint.py`,
46
- [`docs/results/canglong_official_checkpoint.json`](https://github.com/GISWLH/WeatherAI/blob/main/docs\results\canglong_official_checkpoint.json), 2026-10-18):** the figshare
47
  `canglong3_0005_600ep_base.tar`, `normalized_data.pt` and `ocean_mask.npy` (MD5s equal to the figshare record) on an
48
  NVIDIA RTX 4090. `load_official` and the unchanged notebook cell strict-load the same weights; on the last 16-month window
49
  of the released data the two outputs are **identical (max |Δ| = 0.0)**, finite, 0.8 s per full-size forward. Against the
@@ -51,7 +51,7 @@ model = load_official("canglong3_0005_600ep_base.tar") # figshare checkpoint (
51
  (persistence 1.19 / 2.39 / 3.59 K) — one window, a sanity check rather than a skill estimate.
52
  * **GPU smoke:** `scripts/gpu_smoke.py canglong` (lite parity + lite training step + one full-size forward) is wired into the
53
  HF Space; the same smoke on CPU (2026-10-10): **PASS** — lite vs official reference max |Δ| 0.0, lite gradients finite,
54
- full-size forward 54 s. Log: [model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\model_status.md).
55
  * **Live forecasts from the same model family:** [canglong.earth](https://canglong.earth) and the Hugging Face Space
56
  [LonghaoWang/CAS-Canglong](https://huggingface.co/spaces/LonghaoWang/CAS-Canglong) (16-month SST and ENSO outlook).
57
  The operational 16-month model is a later Swin-transformer variant described in the
 
1
  # CAS-Canglong — native PyTorch, bit-exact vs the official code (S2S global SST)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Advanced Science*, "Improved Skills for Predicting Sub-Seasonal to Seasonal Global Sea Surface Temperature", [doi:10.1002/advs.78271](https://doi.org/10.1002/advs.78271) (preprint: arXiv:2409.05369)
4
 
5
  `weatherai/models/canglong` re-implements CAS-Canglong, a 3-D Earth-specific transformer (Pangu-style, U-shaped with a skip
6
  concatenation) that predicts global sea surface temperature 1–3 months ahead from 16 monthly frames of 8 ocean/atmosphere
 
43
  4. Loss: lead weights (1, 1.5, 2) multiply *inside* the square, and the correlation penalty box (named `nino` in the
44
  notebook, rows 288:336 × cols 480:640) covers 18°N–6°N, 120°E–160°E on the 90°N→90°S, 0→359.75°E grid.
45
  * **Released checkpoint on released data (GPU, fp32; `scripts/canglong_official_checkpoint.py`,
46
+ [`docs/results/canglong_official_checkpoint.json`](https://github.com/GISWLH/WeatherAI/blob/main/docs/results/canglong_official_checkpoint.json), 2026-10-18):** the figshare
47
  `canglong3_0005_600ep_base.tar`, `normalized_data.pt` and `ocean_mask.npy` (MD5s equal to the figshare record) on an
48
  NVIDIA RTX 4090. `load_official` and the unchanged notebook cell strict-load the same weights; on the last 16-month window
49
  of the released data the two outputs are **identical (max |Δ| = 0.0)**, finite, 0.8 s per full-size forward. Against the
 
51
  (persistence 1.19 / 2.39 / 3.59 K) — one window, a sanity check rather than a skill estimate.
52
  * **GPU smoke:** `scripts/gpu_smoke.py canglong` (lite parity + lite training step + one full-size forward) is wired into the
53
  HF Space; the same smoke on CPU (2026-10-10): **PASS** — lite vs official reference max |Δ| 0.0, lite gradients finite,
54
+ full-size forward 54 s. Log: [model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/model_status.md).
55
  * **Live forecasts from the same model family:** [canglong.earth](https://canglong.earth) and the Hugging Face Space
56
  [LonghaoWang/CAS-Canglong](https://huggingface.co/spaces/LonghaoWang/CAS-Canglong) (16-month SST and ENSO outlook).
57
  The operational 16-month model is a later Swin-transformer variant described in the
models/fengwu.md CHANGED
@@ -1,10 +1,10 @@
1
  # FengWu — multi-modal encode–fuse–decode (PyTorch)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Communications Earth & Environment* 2025, [doi:10.1038/s43247-025-02502-y](https://doi.org/10.1038/s43247-025-02502-y)
4
 
5
  `FengWu` follows the 2023 preprint architecture (separate encoders per variable group, a cross-modal fuser, decoders, optional
6
  uncertainty head); `FengWu_lite` is a CPU-friendly default (64 × 128, 69 channels). Architecture notes and status:
7
- [fengwu_specs.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\fengwu_specs.md), [fengwu_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\fengwu_status.md). No official weights are published.
8
 
9
  ```python
10
  import torch
 
1
  # FengWu — multi-modal encode–fuse–decode (PyTorch)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Communications Earth & Environment* 2025, [doi:10.1038/s43247-025-02502-y](https://doi.org/10.1038/s43247-025-02502-y)
4
 
5
  `FengWu` follows the 2023 preprint architecture (separate encoders per variable group, a cross-modal fuser, decoders, optional
6
  uncertainty head); `FengWu_lite` is a CPU-friendly default (64 × 128, 69 channels). Architecture notes and status:
7
+ [fengwu_specs.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/fengwu_specs.md), [fengwu_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/fengwu_status.md). No official weights are published.
8
 
9
  ```python
10
  import torch
models/fuxi.md CHANGED
@@ -1,11 +1,11 @@
1
  # FuXi — cube embedding + Swin-V2 U-Transformer (PyTorch)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *npj Climate and Atmospheric Science* 2023, [doi:10.1038/s41612-023-00512-1](https://doi.org/10.1038/s41612-023-00512-1)
4
 
5
  `weatherai.models.fuxi` (adapted from [WeatherLearn](https://github.com/lizhuoq/WeatherLearn), MIT): space-time cube embedding of
6
  two input steps, a Swin-V2 U-Transformer and a fully connected head. Architecture only — the official FuXi weights are
7
  released as ONNX and are not loaded here. The ensemble and sub-seasonal members of the family are separate, verified ports:
8
- [FuXi-ENS](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi_ens.md) and [FuXi-S2S](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\fuxi_s2s.md).
9
 
10
  ```python
11
  from weatherai.models import FuXi
 
1
  # FuXi — cube embedding + Swin-V2 U-Transformer (PyTorch)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *npj Climate and Atmospheric Science* 2023, [doi:10.1038/s41612-023-00512-1](https://doi.org/10.1038/s41612-023-00512-1)
4
 
5
  `weatherai.models.fuxi` (adapted from [WeatherLearn](https://github.com/lizhuoq/WeatherLearn), MIT): space-time cube embedding of
6
  two input steps, a Swin-V2 U-Transformer and a fully connected head. Architecture only — the official FuXi weights are
7
  released as ONNX and are not loaded here. The ensemble and sub-seasonal members of the family are separate, verified ports:
8
+ [FuXi-ENS](fuxi_ens.md) and [FuXi-S2S](fuxi_s2s.md).
9
 
10
  ```python
11
  from weatherai.models import FuXi
models/fuxi_ens.md CHANGED
@@ -1,6 +1,6 @@
1
  # FuXi-ENS — native PyTorch (reverse-engineered from the official ONNX; numerically checked vs onnxruntime)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Science Advances* 2025, [doi:10.1126/sciadv.adu2854](https://doi.org/10.1126/sciadv.adu2854)
4
 
5
  FuXi-ENS (*Science Advances* 2025) is a latent-noise ensemble forecaster. The only official artefact is `fuxi_ens.onnx`
6
  (Zenodo [10.5281/zenodo.15124541](https://zenodo.org/records/15124541), **CC-BY-NC-4.0 — non-commercial; weights are never committed**:
 
1
  # FuXi-ENS — native PyTorch (reverse-engineered from the official ONNX; numerically checked vs onnxruntime)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Science Advances* 2025, [doi:10.1126/sciadv.adu2854](https://doi.org/10.1126/sciadv.adu2854)
4
 
5
  FuXi-ENS (*Science Advances* 2025) is a latent-noise ensemble forecaster. The only official artefact is `fuxi_ens.onnx`
6
  (Zenodo [10.5281/zenodo.15124541](https://zenodo.org/records/15124541), **CC-BY-NC-4.0 — non-commercial; weights are never committed**:
models/fuxi_s2s.md CHANGED
@@ -1,6 +1,6 @@
1
  # FuXi-S2S — native PyTorch from the ONNX, onnxruntime parity (weights CC-BY-NC-ND, never committed)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature Communications* 2024, [doi:10.1038/s41467-024-50714-1](https://doi.org/10.1038/s41467-024-50714-1)
4
 
5
  `weatherai/models/fuxi_s2s` re-implements FuXi-S2S (*Nature Communications* 2024), reverse-engineered from the released ONNX graph (no PyTorch source is published): input two 76-channel 121×240 frames + step index, 2×2 patch embedding (1536), step-embedding MLP, Swin-V2 U-Transformer (2 stages × 6 blocks, 10×10 windows, cosine attention, continuous relative-position bias, gated-GELU MLP), a variational latent head (`mean`, log-variance, low-rank `cov`) sampled with two noise tensors, a 2-stage 12-block decoder and a pixel-shuffle head; precipitation is log-transformed, and the output is fed back autoregressively (1.039 B parameters). Weights: Zenodo 15718402 `model-1.0.tar` (2.08 GB, **CC-BY-NC-ND-4.0**): downloaded by the user (`download()` / `scripts/fetch_weights.py fuxi_s2s`), **never committed or redistributed**; only scalar fingerprints (`tests/models/fuxi_s2s/data/cpu_summary.json`) are stored.
6
 
 
1
  # FuXi-S2S — native PyTorch from the ONNX, onnxruntime parity (weights CC-BY-NC-ND, never committed)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature Communications* 2024, [doi:10.1038/s41467-024-50714-1](https://doi.org/10.1038/s41467-024-50714-1)
4
 
5
  `weatherai/models/fuxi_s2s` re-implements FuXi-S2S (*Nature Communications* 2024), reverse-engineered from the released ONNX graph (no PyTorch source is published): input two 76-channel 121×240 frames + step index, 2×2 patch embedding (1536), step-embedding MLP, Swin-V2 U-Transformer (2 stages × 6 blocks, 10×10 windows, cosine attention, continuous relative-position bias, gated-GELU MLP), a variational latent head (`mean`, log-variance, low-rank `cov`) sampled with two noise tensors, a 2-stage 12-block decoder and a pixel-shuffle head; precipitation is log-transformed, and the output is fed back autoregressively (1.039 B parameters). Weights: Zenodo 15718402 `model-1.0.tar` (2.08 GB, **CC-BY-NC-ND-4.0**): downloaded by the user (`download()` / `scripts/fetch_weights.py fuxi_s2s`), **never committed or redistributed**; only scalar fingerprints (`tests/models/fuxi_s2s/data/cpu_summary.json`) are stored.
6
 
models/gencast.md CHANGED
@@ -1,6 +1,6 @@
1
  # GenCast — native PyTorch denoiser (official architecture, official Mini weights load) + EDM/DPM-Solver++ sampler
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature* 2025, [doi:10.1038/s41586-024-08252-9](https://doi.org/10.1038/s41586-024-08252-9)
4
 
5
  GenCast (*Nature* 2025) is a conditional EDM diffusion model whose denoiser is a GraphCast-style
6
  grid→mesh→grid network with a 16-layer k-hop mesh transformer; official code is JAX (`google-deepmind/weathernext`).
@@ -44,4 +44,4 @@ raw = net(x, noise_levels) # (B, Ng, 84) r
44
  **Not verified / not ported:** the *network* only was compared against the official network — **not** the full model (no end-to-end official
45
  ensemble forecast comparison; the official `InputsAndResiduals` normalisation, NaN-cleaning of SST and the autoregressive rollout are not ported); the sampling noise
46
  is iid Gaussian on the grid, official is spherical-harmonic white noise (needs `dinosaur`); the 0.25° and Operational checkpoints were not run
47
- (same code path, larger graphs); the lite config is not a trained model. Details: [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\model_status.md).
 
1
  # GenCast — native PyTorch denoiser (official architecture, official Mini weights load) + EDM/DPM-Solver++ sampler
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature* 2025, [doi:10.1038/s41586-024-08252-9](https://doi.org/10.1038/s41586-024-08252-9)
4
 
5
  GenCast (*Nature* 2025) is a conditional EDM diffusion model whose denoiser is a GraphCast-style
6
  grid→mesh→grid network with a 16-layer k-hop mesh transformer; official code is JAX (`google-deepmind/weathernext`).
 
44
  **Not verified / not ported:** the *network* only was compared against the official network — **not** the full model (no end-to-end official
45
  ensemble forecast comparison; the official `InputsAndResiduals` normalisation, NaN-cleaning of SST and the autoregressive rollout are not ported); the sampling noise
46
  is iid Gaussian on the grid, official is spherical-harmonic white noise (needs `dinosaur`); the 0.25° and Operational checkpoints were not run
47
+ (same code path, larger graphs); the lite config is not a trained model. Details: [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/model_status.md).
models/genfocal.md CHANGED
@@ -1,65 +1,51 @@
1
- # GenFocal — climate debiasing and generative regional downscaling
2
 
3
- ## Update: two-stage interface, 19 October 2026
4
 
5
- `from weatherai import GenFocal, GenFocal_lite` now exposes a composition of
6
- the ReFlow debiasing network and a conditional EDM super-resolution network.
7
- The native SR adds noise-level preconditioning, weighted denoising loss and
8
- a single-segment Heun sampler. Both native U-ViTs remain approximations;
9
- they are **not compatible with official Orbax checkpoints**. Default lite
10
- weights are random. This is an architectural/testing implementation, not a
11
- trained regional climate product.
12
 
13
  ```python
14
- from weatherai import GenFocal_lite
15
- model = GenFocal_lite(channels=4).eval()
16
- # x, cond: normalized B,T,lon,lat,C tensors, with channel:mean/channel:std.
17
- result = model.downscale(x, cond, output_size=(8, 32, 16),
18
- flow_steps=8, diffusion_steps=16)
 
 
19
  ```
20
 
21
- The native interpolation uses bilinear space and replicated time; official
22
- daily/hourly preprocessing, positional embeddings, filtered resizes, local
23
- regional padding, classifier-free guidance, multi-device overlap consensus
24
- and full official network weight parity are not claimed. Physical output
25
- requires an explicit physical baseline plus residual climatology mean/std.
26
-
27
- `GenFocalOfficial` in `weatherai.models.genfocal` builds the **actual upstream**
28
- ReFlow or EDM U-ViT from official JSON/YAML configs and optionally restores
29
- Orbax state/EMA. It is lazy and separate from the lite PyTorch API. Inspected
30
- upstream: `google-research/swirl-dynamics@0beb5d20a5b31969e12069ba7d86fc0537e08528`.
31
- The wrapper follows upstream construction/restoration code, but full checkpoint
32
- inference has not been executed here. Use upstream's recommended Linux/Python
33
- 3.12 environment; the existing Space uses Python 3.10, below upstream's >=3.11.
34
-
35
- The previous archive-size obstacle has an alternative: the public GCS prefix
36
- `gs://genfocal/super_resolution/checkpoints/conus_7d_t2m_w10m_q1000_msl/`
37
- contains 12 actual files, **2.022 GiB** (2.171 GB), excluding directory markers.
38
- `python scripts/fetch_genfocal.py` inventories it; `--download --out PATH`
39
- downloads that one subtree with size/checksum checks. The verified manifest is
40
- [recorded here](https://github.com/GISWLH/WeatherAI/blob/main/docs\results\genfocal_conus_checkpoint_manifest.json). No old
41
- content was deleted and no weights have been loaded in the native models.
42
-
43
- The complete public archives remain [Zenodo 21285802](https://zenodo.org/records/21285802):
44
- 8,737,960,532-byte SR and 39,143,363,979-byte debiasing tarballs (CC-BY-4.0).
45
- [Code archive](https://zenodo.org/records/21268077), Apache-2.0.
46
- See [three-paper integration and test scope](https://github.com/GISWLH/WeatherAI/blob/main/docs\research_20261019.md).
47
-
48
- ## Historical flow-core verification (2 October 2026)
49
-
50
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature Machine Intelligence* 2026, [doi:10.1038/s42256-026-01308-7](https://doi.org/10.1038/s42256-026-01308-7)
51
-
52
- `weatherai/models/genfocal` ports the debiasing-stage maths of GenFocal (*Nature Machine Intelligence* 2026): conditional rectified-flow (ReFlow) loss `mean(w·((x₁−x₀)−v(x_t,t,cond))²)` with `x_t = t·x₁ + (1−t)·x₀`, the RK4 sampler (forward and reversed flow) with output de-normalisation, plus a **lite** trainable 3-D U-ViT flow network (`GenFocalNet`: FiLM residual blocks, axial spatial/temporal attention with QK-norm, periodic longitude, channel-mean/std conditioning, zero-init output). The network is a structural stand-in: it is **not** weight-compatible with the official Flax `UNet3d`. The super-resolution stage (EDM denoiser) is not ported. Code: GitHub google-research/swirl-dynamics (Apache-2.0); checkpoints Zenodo 21285802 (CC-BY-4.0), not downloaded.
53
-
54
- ```python
55
- from weatherai.models.genfocal import GenFocalNet, GenFocalNetConfig, reflow_loss, sample_flow
56
- net = GenFocalNet(GenFocalNetConfig.lite(4)) # inputs (batch, time, lon, lat, channels)
57
- loss = reflow_loss(net, x0, x1, {"channel:mean": cm, "channel:std": cs}) # x0 = climate model, x1 = ERA5 (normalised)
58
- y = sample_flow(net, x0, {"channel:mean": cm, "channel:std": cs}, num_steps=8) # RK4, official schedule
59
- ```
60
-
61
- * **Verified (CPU, fp64 native vs fp32 official; oracle = unchanged `ConditionalReFlowModel.loss_fn` and `inference_utils.sampling_from_batch` on a tiny Flax network, stored reference; `scripts/genfocal_make_ref.py`, `scripts/genfocal_parity.py`, `docs/results/genfocal_parity.json`):** loss rel. error ≤5.8e-7 (uniform / logistic time draw, with and without weighted norm); 4- and 8-step RK4 samples, forward and reversed, rel. L2 ≤7.5e-8 (max abs ≤1.3e-6); 6 unit tests (also periodic-longitude equivariance and a training-loss decrease of the lite net).
62
- * **Official-code quirks reproduced:** `tspan = arange(0, 1, 1/N)` has N points, so the sampler takes N−1 steps and returns the state at t = 1−1/N, not t = 1 (`to_end=True` integrates to 1); `lognormal_sampler` is `logistic(std·u + mean)`, not log-normal.
63
- * **HF ZeroGPU (2026-10-02), random weights:** **PASS** — lite net (3.3 M params, 8×64×32×10, 64-128-128 ch) train step + backward 0.23 s, 8-step RK4 sample 0.26 s, peak 1.0 GB; lite output vs CPU max abs 2.5e-3 (not isolated; default cuDNN conv precision suspected).
64
- * **Blocker for the real model:** Zenodo 21285802 holds `super_resolution_checkpoints.tar.gz` (8.74 GB) and `debiasing_checkpoints.tar.gz` (39.1 GB), Orbax/OCDBT (tensorstore) checkpoints of the Flax `UNet3d` / `UNet` (e.g. `conus_7d_t2m_w10m_q1000_msl/300000`, 491 MB blobs); the debiasing config is a 6-level 128-channel 3-D U-ViT (`resize_to_shape` (256, 128), 8 snapshots, 10 channels, 128 heads). Not done: the 39 GB archive exceeds the box disk (~25 GB free) and the Space's storage; porting the official `UNet3d` / `FilteredResize` / `InterpConvMerge` / axial attention to read those weights; ERA5/LENS2 inputs (Google Cloud buckets). No real-data debiasing or downscaling was run.
65
- * **NOT verified / not done:** the official network (no weight load, no parity), the super-resolution stage, real-data skill, the evaluation scripts.
 
1
+ # GenFocal — official pretrained climate debiasing and downscaling
2
 
3
+ [Paper](https://doi.org/10.1038/s42256-026-01308-7) · [Author code](https://github.com/google-research/swirl-dynamics/tree/0beb5d20a5b31969e12069ba7d86fc0537e08528/swirl_dynamics/projects/genfocal) · [Published checkpoints](https://zenodo.org/records/21285802)
4
 
5
+ `GenFocal` uses the author's unchanged JAX/Flax implementation and published
6
+ parameters. `GenFocal_lite` is a separate, untrained PyTorch approximation.
7
+ The main entry never falls back to lite or initializes random replacement weights.
 
 
 
 
8
 
9
  ```python
10
+ from weatherai import GenFocal, GenFocal_lite
11
+ sr = GenFocal.from_pretrained("super_resolution")
12
+ flow = GenFocal.from_pretrained("debiasing")
13
+ # Author-normalized arrays and conditioning; this is a network evaluation:
14
+ y = sr(x, sigma, cond)
15
+ # Optional explicitly untrained experimentation:
16
+ lite = GenFocal_lite(channels=4)
17
  ```
18
 
19
+ Use Linux/Python 3.12 and the pinned environment in
20
+ [hf_official/requirements.txt](https://github.com/GISWLH/WeatherAI/blob/main/hf_official/requirements.txt).
21
+ Set `WEATHERAI_CACHE` to a large local directory outside synchronized folders.
22
+ Author source is fetched at commit `0beb5d20a5b31969e12069ba7d86fc0537e08528`
23
+ and checked for modifications. Published OCDBT tensors are fetched directly
24
+ from the public `genfocal` GCS bucket via HTTP range reads, with shape, dtype
25
+ and SHA-256 records. Optimizer state and duplicate checkpoints are excluded.
26
+ Flax module names containing literal dots are restored from Orbax key metadata.
27
+
28
+ | Stage | Released checkpoint (step 300000) | Parameter tree |
29
+ |---|---|---|
30
+ | ReFlow debiasing | `debiasing/checkpoints/reflow_uvit` | raw `params`, as in author's sampling code |
31
+ | CONUS super-resolution | `super_resolution/checkpoints/conus_7d_t2m_w10m_q1000_msl` | `ema_state.ema`, as in author's inference |
32
+
33
+ Correction to the earlier documentation: the **released debiasing checkpoint**
34
+ uses channels **768, 768, 768, 1024, 1280, 1536**, not the 128-channel example
35
+ configuration. Its published run config is loaded from the checkpoint directory.
36
+ The raw debiasing tree contains 2,656,553,626 parameters (10.63 GB).
37
+ SR uses the author's `conus_sample.yaml` (128/256/384/512 channels, 84 two-hourly
38
+ steps, four output variables). This distinction is essential for weight compatibility.
39
+
40
+ This entry loads/evaluates each official network. Full climate trajectories also
41
+ require the author's input normalization, ReFlow integration, diffusion sampling,
42
+ regional baseline and residual statistics, overlap handling and physical-unit
43
+ reconstruction. A single denoiser/vector-field call is not a completed forecast.
44
+ Both complete pretrained networks were executed on HF ZeroGPU: ReFlow 146.13 s
45
+ (synthetic normalized input), CONUS SR 52.99 s (real ERA5 conditioning, one denoiser
46
+ call). All outputs were finite. See [current validation evidence and limitations](https://github.com/GISWLH/WeatherAI/blob/main/docs/research_20261019.md).
47
+
48
+ Official code is Apache-2.0; released checkpoints are CC-BY-4.0. Original weights
49
+ remain with the authors and are not redistributed in Git. Previous lite loss/RK4
50
+ parity results remain in `docs/results/genfocal_parity.json`; they do not establish
51
+ full pretrained model parity or climate skill.
 
 
 
 
 
 
 
 
 
 
 
 
models/graphcast.md CHANGED
@@ -1,10 +1,10 @@
1
  # GraphCast — grid ↔ icosahedral-mesh GNN (PyTorch)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Science* 2023, [doi:10.1126/science.adi2336](https://doi.org/10.1126/science.adi2336)
4
 
5
  Independent re-implementation of the encode (Grid2Mesh) – process (multi-mesh message passing) – decode (Mesh2Grid) design.
6
  Stage-wise parity against DeepMind's JAX code and NVIDIA PhysicsNeMo at mesh levels 0–1 (graphs, Grid2Mesh, processor
7
- N ≤ 16, Mesh2Grid): [graphcast_parity.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\graphcast_parity.md). Architecture notes: [graphcast_specs.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\graphcast_specs.md).
8
  No official checkpoint is loaded and there is no whole-model parity claim.
9
 
10
  ```python
 
1
  # GraphCast — grid ↔ icosahedral-mesh GNN (PyTorch)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Science* 2023, [doi:10.1126/science.adi2336](https://doi.org/10.1126/science.adi2336)
4
 
5
  Independent re-implementation of the encode (Grid2Mesh) – process (multi-mesh message passing) – decode (Mesh2Grid) design.
6
  Stage-wise parity against DeepMind's JAX code and NVIDIA PhysicsNeMo at mesh levels 0–1 (graphs, Grid2Mesh, processor
7
+ N ≤ 16, Mesh2Grid): [graphcast_parity.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/graphcast_parity.md). Architecture notes: [graphcast_specs.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/graphcast_specs.md).
8
  No official checkpoint is loaded and there is no whole-model parity claim.
9
 
10
  ```python
models/neuralgcm.md CHANGED
@@ -1,6 +1,6 @@
1
  # NeuralGCM — native learned components + torch spherical harmonics (JAX wrapper kept as oracle)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature* 2024, [doi:10.1038/s41586-024-07744-y](https://doi.org/10.1038/s41586-024-07744-y)
4
 
5
  NeuralGCM (*Nature* 2024) = a differentiable spectral dynamical core (`dinosaur`: moist primitive equations on
6
  sigma levels, IMEX-RK stepping) + learned Haiku components. **This is only partly ported:**
@@ -34,7 +34,7 @@ physics 1.3e-5 (|y|≤20), surface towers ≤5.8e-6, surface blend ≤1e-5, vert
34
  SH layer on random fields: ≤3.1e-5 for transforms, vorticity/divergence and the u,v round trip (|values| up to ~100), exact (0.0) for ∇²,
35
  ∂λ; float64 basis orthonormal to 1e-10. **Not verified / not claimed:** any forecast produced by native code (there is none),
36
  whole-model parity, the dycore, feature builders, other checkpoints (1.4°/0.7°/stochastic). (JAX-on-GPU and the JAX training workflow: see the next subsection.) Details and results:
37
- [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs\model_status.md).
38
 
39
  ## NeuralGCM — training / modification workflow (JAX, official `neuralgcm` + `dinosaur` + Haiku; `weatherai/models/neuralgcm/train.py`)
40
 
 
1
  # NeuralGCM — native learned components + torch spherical harmonics (JAX wrapper kept as oracle)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature* 2024, [doi:10.1038/s41586-024-07744-y](https://doi.org/10.1038/s41586-024-07744-y)
4
 
5
  NeuralGCM (*Nature* 2024) = a differentiable spectral dynamical core (`dinosaur`: moist primitive equations on
6
  sigma levels, IMEX-RK stepping) + learned Haiku components. **This is only partly ported:**
 
34
  SH layer on random fields: ≤3.1e-5 for transforms, vorticity/divergence and the u,v round trip (|values| up to ~100), exact (0.0) for ∇²,
35
  ∂λ; float64 basis orthonormal to 1e-10. **Not verified / not claimed:** any forecast produced by native code (there is none),
36
  whole-model parity, the dycore, feature builders, other checkpoints (1.4°/0.7°/stochastic). (JAX-on-GPU and the JAX training workflow: see the next subsection.) Details and results:
37
+ [docs/model_status.md](https://github.com/GISWLH/WeatherAI/blob/main/docs/model_status.md).
38
 
39
  ## NeuralGCM — training / modification workflow (JAX, official `neuralgcm` + `dinosaur` + Haiku; `weatherai/models/neuralgcm/train.py`)
40
 
models/neuralgcm_precip.md CHANGED
@@ -1,6 +1,6 @@
1
  # NeuralGCM precipitation — official checkpoints via official code, trainable loss
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Science Advances* 2026, [doi:10.1126/sciadv.adv6891](https://doi.org/10.1126/sciadv.adv6891)
4
 
5
  `weatherai/models/neuralgcm/precip.py` wraps the two checkpoints of the *Science Advances* 2026 paper: `stochastic_precip_2_8_deg.pkl` (**predicts** precipitation; evaporation is diagnosed from the column water budget) and `stochastic_evap_2_8_deg.pkl` (predicts evaporation; precipitation is **diagnosed** from the budget). Zenodo [17109230](https://zenodo.org/records/17109230), CC-BY-4.0, downloaded to `/workspace/ckpt/ngcm_precip` and never committed. Both are stochastic models at 2.8° (128×64, 32 levels) with 1 h steps.
6
 
 
1
  # NeuralGCM precipitation — official checkpoints via official code, trainable loss
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Science Advances* 2026, [doi:10.1126/sciadv.adv6891](https://doi.org/10.1126/sciadv.adv6891)
4
 
5
  `weatherai/models/neuralgcm/precip.py` wraps the two checkpoints of the *Science Advances* 2026 paper: `stochastic_precip_2_8_deg.pkl` (**predicts** precipitation; evaporation is diagnosed from the column water budget) and `stochastic_evap_2_8_deg.pkl` (predicts evaporation; precipitation is **diagnosed** from the budget). Zenodo [17109230](https://zenodo.org/records/17109230), CC-BY-4.0, downloaded to `/workspace/ckpt/ngcm_precip` and never committed. Both are stochastic models at 2.8° (128×64, 32 levels) with 1 h steps.
6
 
models/orca_dl.md CHANGED
@@ -1,6 +1,6 @@
1
  # ORCA-DL — native PyTorch Swin ocean model (official weights load; matches official code)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Science Advances* 2025, [doi:10.1126/sciadv.adu2488](https://doi.org/10.1126/sciadv.adu2488)
4
 
5
  `weatherai/models/orca_dl` re-implements ORCA-DL (*Science Advances* 2025) without `transformers`/`timm`/`global_land_mask`; parameter names equal the official ones. Weights: HF dataset `JayKuo/ORCA-DL-data` (`seed_1..8.bin`, 2.16 GB each; licence not stated, GitHub repo has no LICENSE file; never committed). Input: GODAS-style monthly fields on 128×360 (−63.5…63.5°), 16 levels, normalised by the shipped monthly anomaly statistics.
6
 
 
1
  # ORCA-DL — native PyTorch Swin ocean model (official weights load; matches official code)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Science Advances* 2025, [doi:10.1126/sciadv.adu2488](https://doi.org/10.1126/sciadv.adu2488)
4
 
5
  `weatherai/models/orca_dl` re-implements ORCA-DL (*Science Advances* 2025) without `transformers`/`timm`/`global_land_mask`; parameter names equal the official ones. Weights: HF dataset `JayKuo/ORCA-DL-data` (`seed_1..8.bin`, 2.16 GB each; licence not stated, GitHub repo has no LICENSE file; never committed). Input: GODAS-style monthly fields on 128×360 (−63.5…63.5°), 16 levels, normalised by the shipped monthly anomaly statistics.
6
 
models/pangu.md CHANGED
@@ -1,11 +1,11 @@
1
  # Pangu-Weather — 3-D Earth-specific transformer (PyTorch)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature* 2023, [doi:10.1038/s41586-023-06185-3](https://doi.org/10.1038/s41586-023-06185-3)
4
 
5
  `weatherai.models.pangu` is adapted from [WeatherLearn](https://github.com/lizhuoq/WeatherLearn) (MIT): patch embedding of the
6
  surface (4 + 3 masks) and upper-air (5 variables × 13 levels) fields, a U-shaped stack of 3-D Earth-specific window-attention
7
  blocks with an Earth position bias, and patch recovery. `Pangu` is the paper size (721 × 1440, embed 192); `Pangu_lite`
8
- uses 8 × 8 patches for smaller grids. The same blocks are reused by [CAS-Canglong](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\canglong.md).
9
 
10
  ```python
11
  from weatherai.models import Pangu, Pangu_lite
 
1
  # Pangu-Weather — 3-D Earth-specific transformer (PyTorch)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature* 2023, [doi:10.1038/s41586-023-06185-3](https://doi.org/10.1038/s41586-023-06185-3)
4
 
5
  `weatherai.models.pangu` is adapted from [WeatherLearn](https://github.com/lizhuoq/WeatherLearn) (MIT): patch embedding of the
6
  surface (4 + 3 masks) and upper-air (5 variables × 13 levels) fields, a U-shaped stack of 3-D Earth-specific window-attention
7
  blocks with an Earth position bias, and patch recovery. `Pangu` is the paper size (721 × 1440, embed 192); `Pangu_lite`
8
+ uses 8 × 8 patches for smaller grids. The same blocks are reused by [CAS-Canglong](canglong.md).
9
 
10
  ```python
11
  from weatherai.models import Pangu, Pangu_lite
models/pbc.md CHANGED
@@ -1,46 +1,46 @@
1
- # PBC: probabilistic subseasonal calibration
2
 
3
- [Model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · [Three-paper integration](https://github.com/GISWLH/WeatherAI/blob/main/docs\research_20261019.md)
 
 
 
4
 
5
- Guan et al., **Advancing subseasonal forecasting with machine learning**, Nature,
6
- 14 October 2026. [Paper](https://doi.org/10.1038/s41586-026-11126-x).
7
- [Official MIT code](https://github.com/mouatadid/pbc), inspected at
8
- `d59e18f720bac712e18563bd469d4679c120f9e5`;
9
- [archived release](https://zenodo.org/records/22238796).
10
-
11
- PBC is statistical post-processing, not an autoregressive neural forecaster.
12
- It combines Debias++ (historical CDF bias subtraction), Persistence++ (per-grid,
13
- per-threshold least squares on intercept, rolling climatology, two observable
14
- lag indicators and raw forecast CDF), clip-then-isotonic CDF projection, and
15
- equal-weight averaging. Forecasts in the paper cover days 19-25 and 26-32.
16
- These are target-relative lead windows, **not** WeatherPre's standard weeks 3/4
17
- (days 14-21/21-28); do not interchange the labels.
18
 
19
  ```python
20
- from weatherai import PBC, PBC_lite
21
- model = PBC() # PBC_lite() uses the same statistical algorithm
22
- model.fit(raw_train, observed_train, clim_train, lag1_train, lag2_train,
23
- debias_raw=seasonal_raw, debias_observed=seasonal_observed)
24
- cdf = model(raw, climatology, lag1, lag2, debias_target=raw_observed_thresholds)
25
- probabilities = model.predict_proba(raw, climatology, lag1, lag2)
 
 
 
26
  ```
27
 
28
- Tensor axes: `(sample, *grid, K-1)` in training; either that layout or `(*grid,
29
- K-1)` in prediction. Four CDF thresholds produce five category probabilities.
30
- `ensemble_cdf` uses strict `<`, including ties at zero precipitation. Save/load
31
- the fitted parameters with `state_dict`; there is no universal pretrained
32
- checkpoint in the official release. SVD least squares also handles collinear
33
- predictors on CUDA. The lite variant reduces data/grid size, not the algorithm.
 
 
 
 
 
34
 
35
- Scope: the numerical kernels are implemented. Users must prepare the paper's
36
- 20-year climatologies, model/observed quintile thresholds, hindcasts, seasonal
37
- training windows and observable lags; the rolling three-year selection over
38
- Debias++ spans 14/28/35 and issuance-date ensembles is not automated here.
39
- Pass separately prepared Debias and Persistence CDFs when their thresholds
40
- differ (as they do in several official experiments). Fit only on observations
41
- available at issuance. No live data download, paper RPSS replication, POET,
42
- AIFS-SUBS retraining, or MicroDuet operational product is claimed.
43
 
44
- Tests compare the isotonic result to SciPy and the fitted coefficients to
45
- independent NumPy least squares, and check category normalization, rank
46
- deficiency, strict ties, invalid inputs and checkpoint round trips.
 
 
1
+ # PBC — official probabilistic subseasonal calibration
2
 
3
+ Guan et al., *Advancing subseasonal forecasting with machine learning*, Nature
4
+ 2026. [Paper](https://doi.org/10.1038/s41586-026-11126-x) ·
5
+ [Author code](https://github.com/mouatadid/pbc/tree/d59e18f720bac712e18563bd469d4679c120f9e5) ·
6
+ [Archived release](https://zenodo.org/records/22238796).
7
 
8
+ `PBC` executes the unchanged official scripts pinned to
9
+ `d59e18f720bac712e18563bd469d4679c120f9e5`. The paper combines Debias++,
10
+ Persistence++, CDF projection and probabilistic ensembling. It is statistical
11
+ post-processing, not a new autoregressive neural forecaster.
 
 
 
 
 
 
 
 
 
12
 
13
  ```python
14
+ from weatherai import PBC
15
+ model = PBC(work_dir="pbc-work")
16
+ # Official baseline, runnable without reforecast archives:
17
+ result = model.run("climatology", "era5-f1_mslp", 19, target_dates="20200106")
18
+ # After preparing author-format historical data and component forecasts:
19
+ # model.run("ecmwfpp", "era5-f1_mslp", 19, target_dates="20200106", arguments=(...))
20
+ # model.run("perpp_ecmwf", "era5-f1_mslp", 19, target_dates="20200106", arguments=(...))
21
+ # model.run("projector", "era5-mslp", 19, target_dates="20200106", arguments=(...))
22
+ # model.run("pbc_ecmwf", "era5-f1_mslp", 19, target_dates="20200106")
23
  ```
24
 
25
+ There is **no universal pretrained checkpoint in the official release**.
26
+ Coefficients depend on variable, grid, target date and historical observations/
27
+ reforecasts. `PBC.from_pretrained` raises an explanatory error. Author code/data
28
+ layout and selection configuration are preserved; missing data never trigger a
29
+ lite substitute. Original scripts sometimes exit successfully after skipping
30
+ missing inputs, so inspect their logs and expected NetCDF outputs.
31
+
32
+ Recommended runtime: author's Linux/Python 3.13 environment; the tested CPU
33
+ stages also ran on Windows/Python 3.13. Dependencies include NumPy, pandas,
34
+ xarray, netCDF4, dask, scikit-learn, ttictoc and setuptools with pkg_resources.
35
+ The full archive preparation uses further tools listed upstream.
36
 
37
+ `python scripts/validate_official_pbc.py` actually runs the original climatology,
38
+ projector and equal-weight PBC scripts. A synthetic fixture verifies clipping,
39
+ isotonic projection and final CDF `[0.05, 0.35, 0.55, 0.95]`.
40
+ [Execution report](https://github.com/GISWLH/WeatherAI/blob/main/docs/results/pbc_official_execution.json). This is not a
41
+ historical fit or reproduction of the paper's skill scores.
 
 
 
42
 
43
+ `PBC_lite()` is the earlier independent tensor approximation, with `.fit`,
44
+ `.predict_proba` and `state_dict`. It requires explicitly prepared CDF histories;
45
+ it is not the official pipeline. Official target windows are days 19–25 and
46
+ 26–32, distinct from WeatherPre's generic weekly windows. Code license: MIT.
models/stormcast.md CHANGED
@@ -1,6 +1,6 @@
1
  # StormCast — native PyTorch regression U-Net + EDM diffusion U-Net (official weights load; bit-identical to PhysicsNeMo on CPU)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Science Advances* 2026, [doi:10.1126/sciadv.adv0423](https://doi.org/10.1126/sciadv.adv0423)
4
 
5
  `weatherai/models/stormcast` re-implements NVIDIA's StormCast v1 (ERA5 → HRRR, 3 km, 512×640 grid, 99 state + 26 conditioning + 2 invariant channels)
6
  without any PhysicsNeMo dependency. Weights: HF [`nvidia/stormcast-v1-era5-hrrr`](https://huggingface.co/nvidia/stormcast-v1-era5-hrrr) (Apache-2.0, ungated; not committed here).
 
1
  # StormCast — native PyTorch regression U-Net + EDM diffusion U-Net (official weights load; bit-identical to PhysicsNeMo on CPU)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Science Advances* 2026, [doi:10.1126/sciadv.adv0423](https://doi.org/10.1126/sciadv.adv0423)
4
 
5
  `weatherai/models/stormcast` re-implements NVIDIA's StormCast v1 (ERA5 → HRRR, 3 km, 512×640 grid, 99 state + 26 conditioning + 2 invariant channels)
6
  without any PhysicsNeMo dependency. Weights: HF [`nvidia/stormcast-v1-era5-hrrr`](https://huggingface.co/nvidia/stormcast-v1-era5-hrrr) (Apache-2.0, ungated; not committed here).
models/translocation.md CHANGED
@@ -1,6 +1,6 @@
1
  # Gray-swan translocation: forecast diagnostics
2
 
3
- [Model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · [Three-paper integration](https://github.com/GISWLH/WeatherAI/blob/main/docs\research_20261019.md)
4
 
5
  Sun et al., **Predicting regional gray swans via translocation: AI weather models
6
  and Dubai's unprecedented 2024 rainfall**, Science Advances, 7 October 2026.
@@ -16,9 +16,10 @@ Official resources from the paper:
16
  | Resource | Code / model |
17
  |---|---|
18
  | GraphCast | [DeepMind code and weight links](https://github.com/google-deepmind/graphcast) |
19
- | FuXi | [Code and ONNX weight links](https://github.com/tpys/FuXi) |
20
  | AIFS | [ecmwf/aifs-single-1.0](https://huggingface.co/ecmwf/aifs-single-1.0) |
21
- | Pangu / FourCastNet-v2 | [ECMWF ai-models integrations](https://github.com/ecmwf-lab/ai-models) |
 
22
  | GraphCast fine-tuning | [csubich/graphcast](https://github.com/csubich/graphcast) |
23
  | Dynamically similar event tracking | [cyclonetracking](https://github.com/alexcrawford0927/cyclonetracking) |
24
  | Paper ERF notebook and data | [Zenodo 18111128](https://zenodo.org/records/18111128) |
@@ -26,22 +27,41 @@ Official resources from the paper:
26
  Zenodo lists the 19,427-byte notebook `graphcast-ERF-cleaned.ipynb`, README,
27
  and three ERF data ZIP files (~3.14, 2.03 and 2.41 GB). These are diagnostics,
28
  not newly released model checkpoints. Model licenses follow each backbone.
 
 
 
 
 
29
 
30
  ```python
31
- from weatherai import Translocation, Translocation_lite
32
- model = Translocation_lite() # random-weight small GraphCast
33
- # For scientific use: model = Translocation(your_differentiable_forecaster).eval()
34
- result = model.effective_receptive_field(x, channel=0, location=(3, 4))
35
- # Or region_mask=(H,W): select the strongest predicted value inside the region.
36
- signed_erf = result['gradient'] # B,C,H,W, with original signs preserved
 
 
37
  ```
38
 
39
- The wrapper delegates forecasts to its supplied backbone. The diagnostic uses
40
- the derivative of one selected output with respect to every input variable,
41
- as in the paper's Dubai analysis. It does not change/accumulate parameter
42
- gradients. Tests check signed analytical derivatives and a GraphCast directional
43
- finite-difference oracle. Real use requires the correct physical output,
44
- normalization, precipitation channel and coordinates. Lite channel 0 has no
45
- physical variable identity. A nonlocal ERF is not by itself evidence of causal
46
- translocation. The paper's weather reconstruction, event tracking and
47
- fine-tuning ablations have not been reproduced.
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  # Gray-swan translocation: forecast diagnostics
2
 
3
+ [Model pages](README.md) 路 [Three-paper integration](https://github.com/GISWLH/WeatherAI/blob/main/docs/research_20261019.md)
4
 
5
  Sun et al., **Predicting regional gray swans via translocation: AI weather models
6
  and Dubai's unprecedented 2024 rainfall**, Science Advances, 7 October 2026.
 
16
  | Resource | Code / model |
17
  |---|---|
18
  | GraphCast | [DeepMind code and weight links](https://github.com/google-deepmind/graphcast) |
19
+ | FuXi | [Original cascade code (pinned)](https://github.com/tpys/FuXi/tree/9292fe0692156a01cd3d62bcb427cc3798cf8add), [author ONNX models / sample data](https://doi.org/10.5281/zenodo.10401602) |
20
  | AIFS | [ecmwf/aifs-single-1.0](https://huggingface.co/ecmwf/aifs-single-1.0) |
21
+ | Pangu | [Original author code and four ONNX weight links](https://github.com/198808xc/Pangu-Weather), [6h checkpoint](https://drive.google.com/file/d/1a4XTktkZa5GCtjQxDJb_fNaqTAUiEJu4/view) |
22
+ | FourCastNet-v2 | [NVIDIA inference code](https://github.com/NVIDIA/earth2mip/blob/main/earth2mip/networks/fcnv2_sm.py), [official NGC v0.2 weights + statistics](https://api.ngc.nvidia.com/v2/models/nvidia/modulus/modulus_fcnv2_sm/versions/v0.2/files/fcnv2_sm.zip) |
23
  | GraphCast fine-tuning | [csubich/graphcast](https://github.com/csubich/graphcast) |
24
  | Dynamically similar event tracking | [cyclonetracking](https://github.com/alexcrawford0927/cyclonetracking) |
25
  | Paper ERF notebook and data | [Zenodo 18111128](https://zenodo.org/records/18111128) |
 
27
  Zenodo lists the 19,427-byte notebook `graphcast-ERF-cleaned.ipynb`, README,
28
  and three ERF data ZIP files (~3.14, 2.03 and 2.41 GB). These are diagnostics,
29
  not newly released model checkpoints. Model licenses follow each backbone.
30
+ The current FuXi repository advertises FuXi 2.1; that newer release must not be
31
+ silently substituted for the original cascade used in this study. The table
32
+ therefore links the original code snapshot and its author-hosted archive.
33
+ Only the GraphCast route was executed for this integration; the other official
34
+ backbone artifacts are discovery references, not newly validated implementations.
35
 
36
  ```python
37
+ from weatherai import Translocation
38
+ model = Translocation.from_pretrained() # official 0.25-degree, 13-level GraphCast
39
+ inputs, targets, forcings = model.prepare(era5_dataset, lead_time="6h")
40
+ forecast = model(inputs, targets * float("nan"), forcings)
41
+ erf = model.effective_receptive_field(
42
+ inputs, targets_template=targets * float("nan"), forcings=forcings,
43
+ variable="total_precipitation_6hr",
44
+ location={"batch": 0, "time": -1, "lat": 260, "lon": 220})
45
  ```
46
 
47
+ The main entry downloads the official GraphCast checkpoint and normalization
48
+ statistics from `dm_graphcast` and imports unchanged author code at
49
+ `08cf73625c9d12bd9aaa038868bcb2fe488f2a22`. This historical revision retains the
50
+ `graphcast` package before the upstream repository's WeatherNext reorganization.
51
+ Variants are `operational` (0.25°, 13 levels), `37levels` (0.25°, 37 levels), and
52
+ `small` (official pretrained 1°, 13 levels; not the paper's main resolution).
53
+ Code is Apache-2.0; GraphCast weights retain CC-BY-NC-SA-4.0.
54
+
55
+ The wrapper order follows the study's released ERF notebook: GraphCast,
56
+ Bfloat16Cast, InputsAndResiduals and autoregressive.Predictor. Signed input
57
+ sensitivities use JAX reverse-mode AD on a selected physical output. Coordinates,
58
+ variable, lead and normalization must match the intended event. ERF alone does
59
+ not establish causality. Full 0.25° operational forecasting and the actual Small-model ERF adapter passed
60
+ ZeroGPU execution. [Current test evidence](https://github.com/GISWLH/WeatherAI/blob/main/docs/research_20261019.md).
61
+
62
+ `Translocation_lite()` retains the independent random-weight PyTorch toy model.
63
+ For custom PyTorch diagnostic experiments, import `TranslocationLite` from
64
+ `weatherai.models.translocation`. Neither is the pretrained main entry.
65
+ The complete Dubai reconstruction, event tracking and fine-tuning ablations
66
+ have not been reproduced. No study-specific fine-tuned weights were found in
67
+ the cited ERF archive; its large ZIP files contain diagnostic data.
models/tropicyclonenet.md CHANGED
@@ -1,6 +1,6 @@
1
  # TropiCycloneNet (TCN_M) — native PyTorch, official weights load, bit-exact vs official code
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature Communications* 2025, [doi:10.1038/s41467-025-61087-4](https://doi.org/10.1038/s41467-025-61087-4)
4
 
5
  `weatherai/models/tropicyclonenet` re-implements the TCN_M generator (*Nature Communications* 2025): `Unet3D` (8 observed 64×64 500 hPa geopotential frames → 12), two `EnvNet`s (tabular environment + geopotential, 2-layer Transformer), two LSTM encoders, six LSTM decoders (generators) and a class chooser; 6-hourly steps, 8 observed → 4 forecast steps (6–24 h) of (Δlon, Δlat, Δpressure, Δwind). Checkpoint: Zenodo 15024028 `checkpoint_with_model_16000.pt` (CC-BY-4.0, 4,767,195 parameters; not committed). Data: public TCND release (Zenodo 15009527, CC-BY-4.0); `scripts/tcn_prepare_ni.py` builds the official loader layout (NI test split).
6
 
 
1
  # TropiCycloneNet (TCN_M) — native PyTorch, official weights load, bit-exact vs official code
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature Communications* 2025, [doi:10.1038/s41467-025-61087-4](https://doi.org/10.1038/s41467-025-61087-4)
4
 
5
  `weatherai/models/tropicyclonenet` re-implements the TCN_M generator (*Nature Communications* 2025): `Unet3D` (8 observed 64×64 500 hPa geopotential frames → 12), two `EnvNet`s (tabular environment + geopotential, 2-layer Transformer), two LSTM encoders, six LSTM decoders (generators) and a class chooser; 6-hourly steps, 8 observed → 4 forecast steps (6–24 h) of (Δlon, Δlat, Δpressure, Δwind). Checkpoint: Zenodo 15024028 `checkpoint_with_model_16000.pt` (CC-BY-4.0, 4,767,195 parameters; not committed). Data: public TCND release (Zenodo 15009527, CC-BY-4.0); `scripts/tcn_prepare_ni.py` builds the official loader layout (NI test split).
6
 
models/unicm.md CHANGED
@@ -1,6 +1,6 @@
1
  # UniCM — native PyTorch, bit-exact vs official code (random weights, no public checkpoint)
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature Machine Intelligence* 2026, [doi:10.1038/s42256-026-01245-5](https://doi.org/10.1038/s42256-026-01245-5)
4
 
5
  `weatherai/models/unicm` re-implements UniCM (*Nature Machine Intelligence* 2026): two encoder–decoder transformers (spatio-temporal field branch and climate-mode series branch) on a 12×72 ocean grid with 5 channels (SST, taux, tauy, HTC300, T20D), 2×2 patches, d=256, 4 heads, 4+4 layers, 12 monthly steps in → 24 out; 12,726,037 parameters, state-dict keys identical to the official `UniCM`. `climate_modes` computes the mode series from the fields (region means; dipoles; WWV proxy), `unicm_loss` is the teacher-forced SST + mode loss (the skill term is not ported), `load_official(path)` strict-loads an official `model_best.pkl`. Official-code quirks are reproduced (spatial attention ignores its mask; decoder FFN reuses `sublayer[1]`; `time_emb` shared between branches).
6
 
 
1
  # UniCM — native PyTorch, bit-exact vs official code (random weights, no public checkpoint)
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature Machine Intelligence* 2026, [doi:10.1038/s42256-026-01245-5](https://doi.org/10.1038/s42256-026-01245-5)
4
 
5
  `weatherai/models/unicm` re-implements UniCM (*Nature Machine Intelligence* 2026): two encoder–decoder transformers (spatio-temporal field branch and climate-mode series branch) on a 12×72 ocean grid with 5 channels (SST, taux, tauy, HTC300, T20D), 2×2 patches, d=256, 4 heads, 4+4 layers, 12 monthly steps in → 24 out; 12,726,037 parameters, state-dict keys identical to the official `UniCM`. `climate_modes` computes the mode series from the fields (region means; dipoles; WWV proxy), `unicm_loss` is the teacher-forced SST + mode loss (the skill term is not ported), `load_official(path)` strict-loads an official `model_best.pkl`. Official-code quirks are reproduced (spatial attention ignores its mask; decoder FFN reuses `sublayer[1]`; `time_emb` shared between branches).
6
 
models/weathernext_cyclones.md CHANGED
@@ -1,6 +1,6 @@
1
  # WeatherNext Cyclones Mini — native PyTorch network (official Mini weights load), JAX wrapper kept as oracle
2
 
3
- [← Model zoo](..\README.md) · [All model pages](https://github.com/GISWLH/WeatherAI/blob/main/docs\models\README.md) · **Paper:** *Nature* 2026, [doi:10.1038/s41586-026-10953-2](https://doi.org/10.1038/s41586-026-10953-2)
4
 
5
  WN-C / WeatherNext 2 (FGN, a *Functional Generative Network*) is the same family as GenCast/GraphCast (grid → icosahedral mesh → 16-layer k-hop
6
  transformer → grid) but with its own wiring: instead of a diffusion noise level, a 32-channel N(0,1) vector conditions every normalisation, the
 
1
  # WeatherNext Cyclones Mini — native PyTorch network (official Mini weights load), JAX wrapper kept as oracle
2
 
3
+ [← Model zoo](../README.md) · [All model pages](README.md) · **Paper:** *Nature* 2026, [doi:10.1038/s41586-026-10953-2](https://doi.org/10.1038/s41586-026-10953-2)
4
 
5
  WN-C / WeatherNext 2 (FGN, a *Functional Generative Network*) is the same family as GenCast/GraphCast (grid → icosahedral mesh → 16-layer k-hop
6
  transformer → grid) but with its own wiring: instead of a diffusion noise level, a 32-channel N(0,1) vector conditions every normalisation, the