EarthAlign: Learning Language-Aligned Representations of Earth System Data

Fadillah Adamsyah Maani¹, Numan Saeed¹, Jiaqi Ma¹, Dat Tien Nguyen¹, Youssef Wehbe², Duncan Watson-Parris³, Salman Khan^1,2

¹Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), ²ADIA Lab, ³University of California, San Diego (UC San Diego)

---

🌍 Overview

EarthAlign is a large-scale Earth system-language alignment dataset and benchmark for learning and evaluating multimodal representations of Earth system data. It contains 553,738 samples spanning 2000-2024, with broad geographic coverage and diverse natural phenomena and normal weather conditions. Each sample pairs temporal and geospatial information and multivariate ERA5 data with four metadata-grounded caption styles (natural-event-focused, impact-focused, physical framing, and atmospheric pattern) that describe the same Earth system condition from different perspectives.

Building on EarthAlign, MeteoLIP is a CLIP/SigLIP-based dual-encoder framework for aligning meteorological data with natural language. Its meteorology tokenizer, MetTok, aggregates variable-specific representations into compact meteorology tokens while incorporating temporal and geospatial information. MeteoLIP is evaluated on cross-modal retrieval, weather-condition classification, and physical variable estimation.

This Hugging Face repository hosts the EarthAlign dataset and the pretrained MeteoLIP weights. The code for preprocessing, training, evaluation, and linear probing is available in the GitHub repository.

---

📦 Repository Contents

MBZUAI/EarthAlign
├── EarthAlign/                        # EarthAlign dataset (~241 GB)
│   ├── webdatasets_train/             # Training data in WebDataset format (.tar shards, ~212 GB)
│   ├── data_era5_processed/           # Test/evaluation data (~27.6 GB, tar archives to extract)
│   │   ├── npz/                       #   Preprocessed ERA5 samples (.npz)
│   │   ├── json/                      #   Multi-style captions (.json)
│   │   └── info.json
│   └── data/
│       └── record_json/               # Full dataset records: per-source natural event and
│                                      #   weather-normal (negsamp) samples with captions and metadata
└── MeteoLIP/                          # Pretrained MeteoLIP weights (~17.2 GB)
    ├── Meteo-CLIP/
    │   ├── ViT-B-4_Full-MetTok/
    │   └── ViT-L-4_Full-MetTok/
    └── Meteo-SigLIP/
        ├── ViT-B-4_Full-MetTok/
        └── ViT-L-4_Full-MetTok/

MeteoLIP Weights

Model	Meteorology Encoder	MetTok Configuration	Weights
Meteo-CLIP ViT-B	ViT-B/4	Full MetTok	Link
Meteo-CLIP ViT-L	ViT-L/4	Full MetTok	Link
Meteo-SigLIP ViT-B	ViT-B/4	Full MetTok	Link
Meteo-SigLIP ViT-L	ViT-L/4	Full MetTok	Link

---

⚖️ License

This repository is released under CC BY-NC 4.0. EarthAlign is built on ERA5 reanalysis data and natural event records from multiple public sources, each with its own license or terms of use. Please refer to the natural event record sources for details.

---

🙏 Acknowledgement

This material is based on work supported by ADIA Lab, Abu Dhabi under a collaborative research project titled Climate Co-Pilot: A Multimodal AI Agent for Climate Resilience in the MENA Region.

Our codebase is built upon the OpenCLIP repository, and we thank sincerely its developers and contributors for making their work publicly available.

---

📜 Citation

If you find EarthAlign or MeteoLIP useful in your research, please consider citing:

@misc{maani2026earthalignlearninglanguagealignedrepresentations,
      title={EarthAlign: Learning Language-Aligned Representations of Earth System Data}, 
      author={Fadillah Adamsyah Maani and Numan Saeed and Jiaqi Ma and Dat Tien Nguyen and Youssef Wehbe and Duncan Watson-Parris and Salman Khan},
      year={2026},
      eprint={2610.14624},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.14624}, 
}