Matus-Bello et al. (2026) Deep-learning-based prediction of precipitable water vapor in the Chajnantor area

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Springer Link (Chiba Institute of Technology)
• Year: 2026
• Date: 2026-01-06
• Authors: Alison Matus-Bello, Silvia E. Restrepo, Ricardo Bustos, Yi Hu, Fujia Du, Jaime Cariñe, Pablo García, Javier Sanchez Maldonado, R. Reeves, Zhaohui Shang
• DOI: 10.1051/0004-6361/202556107/pdf

Research Groups

• Astronomical Observatories in the Chajnantor area (e.g., Atacama Pathfinder Experiment)
• Research groups focused on deep learning and atmospheric science for astronomical applications.

Short Summary

This paper develops and validates a Long Short-Term Memory (LSTM) deep-learning model to predict precipitable water vapor (PWV) in the Chajnantor area, demonstrating significantly improved prediction accuracy for 12- and 24-hour horizons compared to traditional methods.

Objective

• To develop and validate a deep-learning-based model for predicting precipitable water vapor (PWV) at various forecasting horizons (12, 24, 36, and 48 hours) to support radio telescope operations in the Chajnantor area.

Study Configuration

• Spatial Scale: Chajnantor area, northern Chile.
• Temporal Scale: Forecasting horizons of 12 hours, 24 hours, 36 hours, and 48 hours.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) deep-learning-based model.
• Data sources: Historical data from two 183 GHz radiometers and a weather station in the Chajnantor area.

Main Results

• The LSTM method predicts PWV for 12- and 24-hour forecasting horizons with a mean absolute percentage error (MAPE) of approximately 22%.
• This performance significantly improves upon the traditional Global Forecast System (GFS) method, which exhibited a MAPE of approximately 36%.
• The root mean square error (RMSE) for the LSTM method was reduced by approximately 50% compared to traditional methods for the 12- and 24-hour horizons.

Contributions

• Presents the first application of deep learning techniques for preliminary predictions of precipitable water vapor (PWV) in the Chajnantor area.
• Demonstrates significant improvements in PWV prediction performance for 12- and 24-hour time windows compared to traditional methods.
• Proposes strategies for further improving the method on shorter (<12 hours) and longer (>36 hours) forecasting timescales.

Funding

• Not specified in the provided text.

Citation

@article{MatusBello2026Deeplearningbased,
  author = {Matus-Bello, Alison and Restrepo, Silvia E. and Bustos, Ricardo and Hu, Yi and Du, Fujia and Cariñe, Jaime and García, Pablo and Maldonado, Javier Sanchez and Reeves, R. and Shang, Zhaohui},
  title = {Deep-learning-based prediction of precipitable water vapor in the Chajnantor area},
  journal = {Springer Link (Chiba Institute of Technology)},
  year = {2026},
  doi = {10.1051/0004-6361/202556107/pdf},
  url = {https://doi.org/10.1051/0004-6361/202556107/pdf}
}