This repo is the official PyTorch implementation of Triton_Earth: TritonCast: Advanced Long-term Earth System Forecasting.

📘Documentation | 🛠️Installation | 🚀Model Zoo | 🤗Huggingface | 👀Visualization | 🆕News

📑Open-source Plan

• [✅] Project Page
• [✅] Paper

🛠️Repository Structure

TritonCast-main/
├── tritoncast/                               # Shared package for common utilities and model builders
├── experiments/                             # Paper experiments grouped under one parent directory
│   ├── exp1_medium_range_weather_forecasting/
│   ├── exp2_long_term_stability_test/
│   ├── exp3_multi_year_climate_simulation/
│   ├── exp4_global_ocean_simulation_and_forecasting/
│   ├── exp6_high_fidelity_eddy_forecast/
│   └── exp7_isotropic_turbulence/
└── Readme.md                                # This document

Below is a guide to the experiments presented in our paper and their corresponding code directories.

Experiment Description	Directory	Quick Start
Medium-Range Weather Forecasting (on WeatherBench 2)	./experiments/exp1_...	Instructions
Long-Term Atmospheric Stability Test (Year-long forecast)	./experiments/exp2_...	Instructions
Multi-Year Climate Simulation	./experiments/exp3_...	Instructions
Global Ocean Simulation & Forecasting	./experiments/exp4_...	Instructions
High-Fidelity Ocean Eddy Forecast	./experiments/exp6_...	Instructions
Isotropic Turbulence Benchmark	./experiments/exp7_...	Instructions

Shared Package

The root-level tritoncast/ package centralizes common utilities that were previously copied across experiments:

• shared config parsing in tritoncast.utils.params
• shared logging and checkpoint loading in tritoncast.utils
• a common model factory in tritoncast.models.build_model()

Experiment folders now primarily keep task-specific data adapters, configs, and entry scripts. For readability, the shared API uses descriptive task names such as medium_range_weather and multi_year_climate; the exp1 / exp2 / exp3 naming is now only kept inside the experiments/ folder to stay aligned with the paper.

🚀Architecture

Figure: The V-cycle architecture of TritonCast. It integrates a Multi-Grid Hierarchy for multi-scale processing, a stable Latent Dynamical Core (LDC) for long-term evolution, and Skip-Connections to retain high-fidelity details. This design effectively mitigates error accumulation in long-term forecasts.

🌟 Highlights

TritonCast establishes a new state-of-the-art in long-term Earth system forecasting. Our key contributions include:

• 🌀 Unprecedented Long-term Stability: Achieves stable, year-long, purely autoregressive global atmospheric forecasts without any drift or model collapse, accurately capturing seasonal cycles.
• 🌊 High-Fidelity Ocean Forecasting: Extends the skillful forecast of ocean eddies to an unprecedented 120 days, preserving fine-scale structures that other models lose.
• 🏆 State-of-the-Art Performance: Matches or exceeds leading AI models (like Pangu-Weather, GraphCast) and operational systems on the WeatherBench 2 benchmark for medium-range forecasting.
• 🌐 Zero-Shot Generalization: Demonstrates a remarkable ability to generalize across resolutions—a model trained on 0.25° data can produce physically realistic forecasts on unseen 0.125° grids, proving it has learned the underlying physical laws.