Summary

Earth-2 MIP is a Python framework designed to standardize the evaluation of AI models for weather and climate prediction. It empowers climate researchers and AI developers by providing uniform interfaces for running model inference and scoring their skill against established metrics. This facilitates rapid experimentation, collaboration, and the establishment of reliable AI baselines for atmospheric science.

How It Works

The framework offers a unified API for interacting with various pre-trained AI models, abstracting away differences in their implementation and data requirements. It focuses on enabling on-the-fly scoring of model performance using standard metrics, bridging the gap between AI model creators and climate domain experts. This approach simplifies the process of assessing how well AI models capture atmospheric physics and integrate with traditional numerical weather forecasting workflows.

Quick Start & Requirements

Installation is available from source via git clone and pip install .. While specific dependencies are detailed in the installation documentation, the examples demonstrate GPU acceleration, implying CUDA is a likely requirement for performance. Links to installation and examples documentation are provided.

Highlighted Details

• Supports a growing zoo of AI models including FourCastNet, DLWP, Pangu, and Graphcast, with details on architecture, type, and size.
• Provides reference workflows and baselines for comparing AI model performance and understanding their physical consistency.
• Facilitates community contributions of new models, particularly through integration with NVIDIA Modulus.
• Offers Python APIs for interactive use and CLIs for distributed model scoring.

Maintenance & Community

Communication channels include GitHub Discussions for general inquiries and model integration, and GitHub Issues for bug reports and feature requests. The project encourages community contributions to expand its model zoo and capabilities.

Licensing & Compatibility

The project is licensed under the Apache License 2.0, which is permissive for commercial use. However, users must be aware that individual AI model checkpoints integrated into the framework may carry their own unique licenses, requiring careful review for specific use cases.

Limitations & Caveats

The project is currently in "Beta" status. A significant caveat is that each model checkpoint may have distinct licensing terms, necessitating user due diligence to understand implications for their specific applications.