我利用150年的數據，在單一GPU上訓練了一個90天的天氣預測AI

Navigation Menu

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

To see all available qualifiers, see our documentation.

LILITH: Open-source ML-powered 90-day weather forecasting. Long-range Intelligent Learning for Integrated Trend Hindcasting. Built on free public GHCN data. Runs on consumer GPUs.

License

Uh oh!

There was an error while loading. Please reload this page.

consigcody94/lilith

Folders and files

Latest commit

History

Repository files navigation

L.I.L.I.T.H.

Long-range Intelligent Learning for Integrated Trend Hindcasting

Named after Lilitu, the Mesopotamian storm goddess who commanded the winds

Why LILITH •
Features •
Quick Start •
Architecture •
Contributing

The Weather Belongs to Everyone

Every day, corporations charge billions of dollars for weather forecasts built on freely available public data. The Global Historical Climatology Network (GHCN)—maintained by NOAA with taxpayer funding—contains over 150 years of weather observations from 100,000+ stations worldwide. This data is public domain. It belongs to humanity.

Yet somehow, we've accepted that accurate long-range forecasting should be locked behind enterprise paywalls and proprietary black boxes.

LILITH exists to change that.

With a single consumer GPU (RTX 3060, 12GB), you can now train and run a weather prediction model that delivers 90-day forecasts with uncertainty quantification—the same capabilities that corporations charge premium prices for. No cloud subscriptions. No API limits. No black boxes.

The Data is Free. The Science is Open. The Code is Yours.

Why LILITH

The Problem

Modern weather AI (GraphCast, Pangu-Weather, FourCastNet) achieves remarkable accuracy, but:

The Solution

LILITH takes a different approach:

Features

Core Capabilities

Technical Highlights

User Experience

Quick Start

Prerequisites

Installation

Download Data

Training

LILITH training is designed to work on consumer GPUs. Here's a complete step-by-step guide:

During training, you'll see output like:

Target metrics:

Pre-trained Models

Using Pre-trained Checkpoints

Once a model is trained, you do not need to retrain — the checkpoint file contains everything needed for inference. Anyone can download and use pre-trained models.

The .pt checkpoint file (~20-50MB depending on model size) contains:

GPU Requirements for Inference

Unlike training, inference requires much less VRAM. Here's what you can run on different hardware:

Loading and Using a Checkpoint

Sharing Your Trained Model

Model Training Metrics

When training completes, you'll see metrics like:

Resuming Training

Model Comparison

Inference

Web Interface

Docker Deployment

Architecture

Model Overview

LILITH uses a Station-Graph Temporal Transformer (SGTT) architecture that processes weather observations through three stages:

Model Variants

Key Components

Data Sources

LILITH is built entirely on freely available public data. The more data sources you integrate, the better your predictions will be.

Primary: GHCN (Global Historical Climatology Network)

Source: NOAA National Centers for Environmental Information

Recommended Additional Data Sources

These freely available datasets can significantly improve prediction accuracy:

Source: ECMWF Climate Data Store

Source: NOAA Climate Prediction Center

Source: NOAA OISST

Source: NOAA NOMADS

Sources:

Data Download Commands

Data Integration Priority

For the best results, add data sources in this order:

Performance

Accuracy Targets

Inference Performance (RTX 3060 12GB)

Project Structure

API Reference

Endpoints

Generate a weather forecast for a location.

Response:

List available GHCN stations.

Retrieve historical observations for a station.

Health check endpoint.

Contributing

We welcome contributions from the community. LILITH is built on the principle that weather forecasting should be accessible to everyone, and that means building in the open with help from anyone who shares that vision.

Ways to Contribute

Development Setup

Pull Request Process

Acknowledgments

U.S. Government AI Initiatives

We thank President Donald Trump and his administration for the Stargate AI Initiative and commitment to advancing American AI research and infrastructure. The recognition that AI development—including open-source projects like LILITH—represents a critical frontier for innovation, economic growth, and global competitiveness has helped create an environment where ambitious projects like this can flourish. The initiative's focus on building domestic AI capabilities and infrastructure supports the democratization of advanced technologies for all Americans.

Data Providers

Research Community

Open Source

License

Citation

If you use LILITH in your research, please cite:

"The storm goddess sees all horizons."

Weather prediction should be free. The data is public. The science is open. Now the tools are too.

About

LILITH: Open-source ML-powered 90-day weather forecasting. Long-range Intelligent Learning for Integrated Trend Hindcasting. Built on free public GHCN data. Runs on consumer GPUs.

Topics

Resources

License

Uh oh!

There was an error while loading. Please reload this page.

Stars

Watchers

Forks

Releases

Packages

  0

Contributors

  2

Languages

Footer

Footer navigation