Sync from GitHub 67168d0
Browse files- README.md +33 -32
- models/README.md +24 -24
- models/aardvark.md +2 -2
- models/ace2.md +1 -1
- models/arches.md +1 -1
- models/aurora.md +2 -2
- models/canglong.md +3 -3
- models/fengwu.md +2 -2
- models/fuxi.md +2 -2
- models/fuxi_ens.md +1 -1
- models/fuxi_s2s.md +1 -1
- models/gencast.md +2 -2
- models/genfocal.md +45 -59
- models/graphcast.md +2 -2
- models/neuralgcm.md +2 -2
- models/neuralgcm_precip.md +1 -1
- models/orca_dl.md +1 -1
- models/pangu.md +2 -2
- models/pbc.md +38 -38
- models/stormcast.md +1 -1
- models/translocation.md +38 -18
- models/tropicyclonenet.md +1 -1
- models/unicm.md +1 -1
- models/weathernext_cyclones.md +1 -1
README.md
CHANGED
|
@@ -17,14 +17,14 @@ tags:
|
|
| 17 |
|
| 18 |
<div align="center">
|
| 19 |
|
| 20 |
-
<img src="
|
| 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 |
[](https://github.com/GISWLH/WeatherAI/blob/main/LICENSE)
|
| 25 |
[](https://github.com/GISWLH/WeatherAI/blob/main/pyproject.toml)
|
| 26 |
[](https://pytorch.org)
|
| 27 |
-
[](
|
| 28 |
[](#model-zoo)
|
| 29 |
[](https://huggingface.co/spaces/LonghaoWang/weatherai-graphcast-smoke)
|
| 30 |
[](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
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
[
|
|
|
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 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](
|
| 114 |
-
| **GenFocal** | `from weatherai
|
| 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](
|
| 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](
|
| 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](
|
| 123 |
|
| 124 |
-
More: [all model pages](
|
| 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
|
| 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
|
| 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
|
| 173 |
<!-- /readme-visual-refresh:artwork-source -->
|
| 174 |
|
| 175 |
|
| 176 |
<!-- readme-visual-refresh:execution -->
|
| 177 |
## Demonstration captured on EHPC · 17 October 2026
|
| 178 |
|
| 179 |
-
 · [日本語](https://github.com/GISWLH/WeatherAI/blob/main/README_ja.md)
|
| 23 |
|
| 24 |
[](https://github.com/GISWLH/WeatherAI/blob/main/LICENSE)
|
| 25 |
[](https://github.com/GISWLH/WeatherAI/blob/main/pyproject.toml)
|
| 26 |
[](https://pytorch.org)
|
| 27 |
+
[](models/neuralgcm.md)
|
| 28 |
[](#model-zoo)
|
| 29 |
[](https://huggingface.co/spaces/LonghaoWang/weatherai-graphcast-smoke)
|
| 30 |
[](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 |
+

|
| 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](..
|
| 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](
|
| 11 |
-
| 🔬 Extreme-event diagnostics | Gray-swan translocation (existing backbones) | [translocation.md](
|
| 12 |
-
| 🌍 Medium-range | Pangu-Weather | [pangu.md](
|
| 13 |
-
| 🌍 Medium-range | FuXi | [fuxi.md](
|
| 14 |
-
| 🌍 Medium-range | FengWu | [fengwu.md](
|
| 15 |
-
| 🌍 Medium-range | GraphCast | [graphcast.md](
|
| 16 |
-
| 🌍 Medium-range | Aurora | [aurora.md](
|
| 17 |
-
| 🌍 Medium-range | NeuralGCM | [neuralgcm.md](
|
| 18 |
-
| 🌍 Medium-range | Aardvark Weather | [aardvark.md](
|
| 19 |
-
| 🎲 Ensemble | GenCast | [gencast.md](
|
| 20 |
-
| 🎲 Ensemble | FuXi-ENS | [fuxi_ens.md](
|
| 21 |
-
| 🎲 Ensemble | ArchesWeather / ArchesWeatherGen | [arches.md](
|
| 22 |
-
| 🎲 Ensemble | NeuralGCM precipitation | [neuralgcm_precip.md](
|
| 23 |
-
| 🌊 S2S / climate / ocean | CAS-Canglong 🆕 | [canglong.md](
|
| 24 |
-
| 🌊 S2S / climate / ocean | FuXi-S2S | [fuxi_s2s.md](
|
| 25 |
-
| 🌊 S2S / climate / ocean | UniCM | [unicm.md](
|
| 26 |
-
| 🌊 S2S / climate / ocean | ORCA-DL | [orca_dl.md](
|
| 27 |
-
| 🌊 S2S / climate / ocean | ACE2-ERA5 | [ace2.md](
|
| 28 |
-
| 🌊 S2S / climate / ocean | GenFocal | [genfocal.md](
|
| 29 |
-
| 🌀 Regional / cyclones | StormCast | [stormcast.md](
|
| 30 |
-
| 🌀 Regional / cyclones | TropiCycloneNet | [tropicyclonenet.md](
|
| 31 |
-
| 🌀 Regional / cyclones | WeatherNext Cyclones | [weathernext_cyclones.md](
|
| 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
|
|
|
|
| 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](..
|
| 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
|
| 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](..
|
| 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](..
|
| 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](..
|
| 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
|
| 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](..
|
| 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
|
| 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
|
| 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](..
|
| 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
|
| 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](..
|
| 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](
|
| 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](..
|
| 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](..
|
| 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](..
|
| 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
|
|
|
|
| 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
|
| 2 |
|
| 3 |
-
|
| 4 |
|
| 5 |
-
`
|
| 6 |
-
|
| 7 |
-
The
|
| 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 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
|
|
|
|
|
|
| 19 |
```
|
| 20 |
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 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](..
|
| 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
|
| 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](..
|
| 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
|
| 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](..
|
| 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](..
|
| 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](..
|
| 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](
|
| 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
|
| 2 |
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
| 4 |
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 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
|
| 21 |
-
model = PBC()
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
| 26 |
```
|
| 27 |
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 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 |
-
|
| 45 |
-
|
| 46 |
-
|
|
|
|
|
|
| 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](..
|
| 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](
|
| 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 | [
|
| 20 |
| AIFS | [ecmwf/aifs-single-1.0](https://huggingface.co/ecmwf/aifs-single-1.0) |
|
| 21 |
-
| Pangu
|
|
|
|
| 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
|
| 32 |
-
model =
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
| 37 |
```
|
| 38 |
|
| 39 |
-
The
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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](..
|
| 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](..
|
| 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](..
|
| 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
|