README.md

---
frameworks: JAX
language:
- en
license: apache-2.0
tags:
- OneScience
- Earth Science
- Weather Forecasting
- Climate Simulation
- Hybrid Physics-ML
- ERA5
- NeuralGCM
tasks: []
datasets:
- OneScience/ERA5
---

<p align="center">
  <strong>
    <span style="font-size: 30px;">NeuralGCM</span>
  </strong>
</p>

# Model Introduction

NeuralGCM (Neural General Circulation Models) is an open-source hybrid machine-learning and physics-based atmospheric model developed by Google Research for weather forecasting and climate simulation.

Paper: Neural General Circulation Models for Weather and Climate

https://arxiv.org/abs/2311.07222

# Model Description

NeuralGCM is built around a differentiable atmospheric dynamical core. Neural networks represent unresolved physical processes, the encoder, and the decoder, improving forecast efficiency while retaining physical constraints.

| Profile | Resolution | Type | Bundled official checkpoint |
| :--- | :---: | :--- | :--- |
| `weather_forecast` | 0.7 degrees (`512 x 256`) | Deterministic weather forecasting for approximately 2 to 15 days | `weight/models_v1_deterministic_0_7_deg.pkl` |
| `climate_scale` | 1.4 degrees (`256 x 128`) | Deterministic climate-scale simulation | `weight/models_v1_deterministic_1_4_deg.pkl` |
| `forecast_2_8_deg` | 2.8 degrees (`128 x 64`) | Deterministic weather forecasting | `weight/models_v1_deterministic_2_8_deg.pkl` |
| `stochastic_1_4_deg` | 1.4 degrees (`256 x 128`) | Stochastic weather forecasting | `weight/models_v1_stochastic_1_4_deg.pkl` |

# Use Cases

| Scenario | Description |
| :---: | :--- |
| Global weather forecasting | Train the 0.7-degree model on ERA5 data for short- to medium-range weather forecasting. |
| Climate-scale simulation | Train the 1.4-degree model on ERA5 data for longer atmospheric simulations. |
| Low-resolution experiments | Use the 2.8-degree data profile for lower-cost weather forecasting experiments. |
| Local quick validation | Generate HDF5 data with the required channel protocol using `scripts/fake_data.py` and validate the data, model, and checkpoint workflows. |
| ModelScope / OneCode execution | Download the standalone model package, install the OneScience and JAX dependencies, and run the scripts directly. |
| Multi-device training | Run synchronous data-parallel training on multiple local accelerators. |

# Usage Guide

## 1. OneCode Usage

Experience intelligent one-click AI4S programming through the OneCode online environment:

[Click to Experience Intelligent One-Click AI4S Programming](https://web-2069360198568017922-iaaj.ksai.scnet.cn:58043/home)

## 2. Manual Installation and Usage

**Hardware Requirements**

- A GPU or DCU is recommended.
- A CPU can be used for import checks and small-scale connectivity validation, but full training and inference will be slow.
- DCU users must install DTK in advance. DTK 25.04.2 or later, or the OneScience-recommended version compatible with the current cluster, is recommended.

### Download the Model Package

```bash
hf download OneScience-Group/NeuralGCM --local-dir ./NeuralGCM
cd NeuralGCM
```

### Install the Runtime Environment

**DCU Environment**

```bash
# Activate DTK and conda first.
conda create -n onescience311 python=3.11 -y
conda activate onescience311
# Installation with uv is also supported.
pip install onescience[earth-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai
```

**GPU Environment**

```bash
# Activate conda first.
conda create -n onescience311 python=3.11 -y libstdcxx-ng=12 libgcc-ng=12 gcc_linux-64=12 gxx_linux-64=12
conda activate onescience311
# Installation with uv is also supported.
pip install onescience[earth-gpu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai
```

### Training Data

The OneScience community provides an ERA5 data slice for training. Download it with the following command and confirm that the data path in `conf/config.yaml` is correct:

```bash
hf download --repo-type dataset OneScience-Group/ERA5 --local-dir ./data
```

### Generate Synthetic Data

```bash
python scripts/fake_data.py
```

The script creates yearly HDF5 files under `data/data/`, writes synthetic static fields to `data/static.nc`, and saves channel, time-window, and grid metadata to `data/metadata/dataset_card.json`. The synthetic fields use approximate physical units but are intended only for shape, loading, regridding, and numerical-stability checks.

### Training

Single device:

```bash
# 0.7-degree deterministic short- to medium-range weather forecasting
python scripts/train_weather_forecast.py
# 1.4-degree deterministic climate-scale simulation
python scripts/train_climate_scale.py
# 2.8-degree deterministic low-resolution weather forecasting
python scripts/train_forecast_2_8_deg.py
# 1.4-degree stochastic weather forecasting
python scripts/train_stochastic_1_4_deg.py
```

Multiple devices:

```bash
# 0.7-degree deterministic short- to medium-range weather forecasting
python scripts/train_weather_forecast.py --devices 8
# 1.4-degree deterministic climate-scale simulation
python scripts/train_climate_scale.py --devices 8
# 2.8-degree deterministic low-resolution weather forecasting
python scripts/train_forecast_2_8_deg.py --devices 8
# 1.4-degree stochastic weather forecasting
python scripts/train_stochastic_1_4_deg.py --devices 8
```

### Fine-tuning

Fine-tuning can start from either a checkpoint produced by local training or the bundled official checkpoint for the selected profile.

```bash
# Use the bundled official checkpoint for each profile.
python scripts/train_weather_forecast.py --finetune weight/models_v1_deterministic_0_7_deg.pkl
python scripts/train_climate_scale.py --finetune weight/models_v1_deterministic_1_4_deg.pkl
python scripts/train_forecast_2_8_deg.py --finetune weight/models_v1_deterministic_2_8_deg.pkl
python scripts/train_stochastic_1_4_deg.py --finetune weight/models_v1_stochastic_1_4_deg.pkl

# Alternatively, provide a local checkpoint explicitly.
python scripts/train_weather_forecast.py --finetune ./data/checkpoint/model_bak.pkl
```

For multi-device fine-tuning, add `--devices` to the corresponding command.

### Pre-trained Weights

This project includes the following official pre-trained checkpoints:

| Local file | Official release path |
| :--- | :--- |
| `weight/models_v1_deterministic_0_7_deg.pkl` | `gs://neuralgcm/models/v1/deterministic_0_7_deg.pkl` |
| `weight/models_v1_deterministic_1_4_deg.pkl` | `gs://neuralgcm/models/v1/deterministic_1_4_deg.pkl` |
| `weight/models_v1_deterministic_2_8_deg.pkl` | `gs://neuralgcm/models/v1/deterministic_2_8_deg.pkl` |
| `weight/models_v1_stochastic_1_4_deg.pkl` | `gs://neuralgcm/models/v1/stochastic_1_4_deg.pkl` |

### Inference

```bash
# 0.7-degree deterministic short- to medium-range weather forecasting
python scripts/inference.py --mode weather_forecast --checkpoint weight/models_v1_deterministic_0_7_deg.pkl
# 1.4-degree deterministic climate-scale simulation
python scripts/inference.py --mode climate_scale --checkpoint weight/models_v1_deterministic_1_4_deg.pkl
# 2.8-degree deterministic low-resolution weather forecasting
python scripts/inference.py --mode forecast_2_8_deg --checkpoint weight/models_v1_deterministic_2_8_deg.pkl
# 1.4-degree stochastic weather forecasting
python scripts/inference.py --mode stochastic_1_4_deg --checkpoint weight/models_v1_stochastic_1_4_deg.pkl
```

Without an explicit `--checkpoint`, inference first checks `./data/checkpoint/model_bak.pkl`. The default output is `results/predictions.nc`, containing pressure-level variables with their official names and rollout time coordinates.

### Evaluation and Visualization

```bash
python scripts/result.py
```

# Official OneScience Resources

| Platform | OneScience Main Repository | Skills Repository |
| --- | --- | --- |
| Gitee | https://gitee.com/onescience-ai/onescience | https://gitee.com/onescience-ai/oneskills |
| GitHub | https://github.com/onescience-ai/OneScience | https://github.com/onescience-ai/oneskills |

# Citation and License

- This repository is a reproduction of the original NeuralGCM paper.
- The repository code is provided under the Apache License 2.0.
- The trained model weights released by Google, including the four checkpoints in this directory, are licensed under the Creative Commons Attribution-ShareAlike 4.0 International license (CC BY-SA 4.0). Redistribution or adaptation of the weights must preserve attribution and use the same license as required by those terms.