# WeatherNext Future Development Plans: Roadmap for AI Weather Forecasting

> Explore WeatherNext future development plans including AI weather forecasting, higher-resolution models, unified APIs, and enhanced cyclone tracking. See the roadmap.

- Repository: [Google DeepMind/weathernext](https://github.com/google-deepmind/weathernext)
- Tags: architecture
- Published: 2026-08-10

---

**Google DeepMind's WeatherNext is evolving toward higher-resolution models, unified APIs, and improved cyclone tracking capabilities.**

The WeatherNext repository represents a modular, open-source research platform for global medium-range weather and cyclone forecasting. This article examines the concrete future development plans signaled by the DeepMind team through source code comments, documentation updates, and architectural decisions in the `google-deepmind/weathernext` codebase.

## Current Architecture: Three Model Families

WeatherNext currently hosts three distinct model families that share common infrastructure:

- **WeatherNext 2 (WN2)** – Probabilistic weather forecasting via Flow Generative Networks (`weathernext/weathernext2/`)
- **WeatherNext Cyclones** – Operational tropical cyclone tracking (`weathernext/cyclones/`)
- **Legacy models** – GraphCast (Science 2023) and GenCast (Nature 2024) in `weathernext/weathernext1_graph/` and `weathernext/weathernext1_gen/`

Shared utilities in `weathernext/utils/` (autoregressive rollouts, normalization, mesh-transformer, sparse-transformer) enable rapid experimentation across all three families.

## Six Priority Areas in the WeatherNext Roadmap

### Higher-Resolution, Larger-Ensemble Models

The current "Mini" configurations target cheap TPU/GPU inference. **TODO comments** in [`weathernext/weathernext2/architecture_utils.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/weathernext2/architecture_utils.py) and [`weathernext/utils/update_blocks.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/utils/update_blocks.py) explicitly discuss unifying APIs to support higher-resolution training without extensive code modifications. This signals a move toward operational-scale deployments matching the resolution of traditional numerical weather prediction systems.

### Unified Block API for Flexible Experimentation

Multiple **TODO notes** in [`update_blocks.py`](https://github.com/google-deepmind/weathernext/blob/main/update_blocks.py) and [`rollout.py`](https://github.com/google-deepmind/weathernext/blob/main/rollout.py) indicate plans to consolidate block interfaces. This architectural simplification will enable researchers to swap Graph Neural Network (GNN) or diffusion blocks with minimal friction, accelerating the research cycle for new model architectures.

### Extended Cyclone Tracking Capabilities

The cyclone tracker currently processes 6-hourly forecasts. Comments in [`weathernext/cyclones/direct_tracker.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/cyclones/direct_tracker.py) reference adding **"future centers"** and improving handling of post-processed best-track data. This suggests enhanced temporal resolution and accuracy for operational tropical cyclone warnings.

### Expanded Data Pipeline Integrations

[`weathernext/cyclones/data_pipeline_utils.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/cyclones/data_pipeline_utils.py) contains **placeholder code** for future integration of additional observational datasets. Satellite-derived wind and precipitation products are specifically targeted, which would improve initialization quality and verification capabilities.

### Model-Agnostic Inference Back-Ends

The README explicitly warns that **"future updates may introduce breaking changes"** and recommends release pinning. This indicates an intentional evolution of the public API toward greater hardware flexibility and deployment options while preserving scientific reproducibility.

### Commercial-Use Licensing for Model Weights

An August 2026 documentation update announced a **commercial-use-friendly license** for model weights. This policy shift points toward broader community adoption and potential public model-hosting infrastructure, reducing barriers for operational meteorological services.

## Code Example: Current WeatherNext 2 Inference Workflow

The following pattern demonstrates the extensible design that supports these future developments:

```python
import jax
import xarray as xr
from weathernext.weathernext2 import fgn, architecture_utils

# Load initial conditions

init_state = xr.open_zarr('gs://weatherbench2/hres/2022-01-01.zarr')

# Build model with configurable architecture

model = fgn.FGN(architecture_utils.get_config('WeatherNext2_<2025'))

# Initialize with pretrained weights

params = model.initialize_parameters(jax.random.PRNGKey(0))

# Autoregressive rollout: 6-day forecast (48 steps at 3-hour intervals)

predictor = architecture_utils.Predictor(model, params)
forecast = predictor.rollout(init_state, num_steps=48)

```

*Source: [`weathernext/weathernext2/fgn.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/weathernext2/fgn.py), [`weathernext/weathernext2/architecture_utils.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/weathernext2/architecture_utils.py)*

## Code Example: Cyclone Tracking Integration

The tracker framework is designed for the planned enhancements:

```python
from weathernext.cyclones.tracker_base import CycloneTracker
from weathernext.cyclones import ibtracs_processing_utils as ibtracs

tracker = CycloneTracker()
tracks = tracker.track(forecast)  # Currently 6-hourly; higher frequency planned

# Evaluation against best-track data

ibtracs_data = ibtracs.load_ibtracs('2022')
metrics = tracker.evaluate_against_ibtracs(tracks, ibtracs_data)

```

*Source: [`weathernext/cyclones/direct_tracker.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/cyclones/direct_tracker.py)*

## What Developers Should Expect

The WeatherNext codebase is designed for **rapid research iteration**. The development trajectory emphasizes:

1. **Scalability** – Higher-resolution operational runs with larger ensembles
2. **Extensibility** – New data modalities and interchangeable model components
3. **Operational robustness** – Enhanced cyclone tracking and stable, versioned APIs

Researchers and operational users should monitor [`README.md`](https://github.com/google-deepmind/weathernext/blob/main/README.md) for API change announcements and consider pinning to specific releases for production deployments.

## Summary

- **WeatherNext future development** centers on higher-resolution models, unified APIs, and improved cyclone tracking
- **TODO comments** in [`architecture_utils.py`](https://github.com/google-deepmind/weathernext/blob/main/architecture_utils.py) and [`update_blocks.py`](https://github.com/google-deepmind/weathernext/blob/main/update_blocks.py) reveal active work on scalable architectures
- **Cyclone tracker enhancements** are planned for finer temporal resolution and better best-track integration
- **Data pipeline expansion** will incorporate satellite-derived observations
- **Commercial licensing changes** signal intent for broader operational adoption
- **API evolution** is expected; release pinning is recommended for reproducibility

## Frequently Asked Questions

### Will WeatherNext support higher-resolution forecasts than the current Mini versions?

Yes. The source code contains explicit **TODO comments** in [`weathernext/weathernext2/architecture_utils.py`](https://github.com/google-deepmind/weathernext/blob/main/weathernext/weathernext2/architecture_utils.py) discussing unified APIs for higher-resolution training. The modular design of shared utilities in `weathernext/utils/` is specifically architected to accommodate scaling without fundamental code rewrites.

### Is there a formal public roadmap for WeatherNext?

No formal roadmap file exists in the repository. However, **future development plans** are clearly signaled through inline TODO comments, README warnings about breaking changes, and documentation updates regarding licensing. These sources collectively indicate priorities: scalability, extensibility, and operational deployment.

### When will commercial use of WeatherNext model weights be permitted?

According to the August 2026 README update in [`docs/weathernext1_graph/README.md`](https://github.com/google-deepmind/weathernext/blob/main/docs/weathernext1_graph/README.md), **commercial-use-friendly licensing** for model weights has already been announced. This change removes previous restrictions and enables operational meteorological services to deploy WeatherNext models in production environments.