Chung et al. (2026) Machine Learning Post-Processing of Atmospheric River Persistence Forecasts: A Pre-Trained Tabular Transformer Across Mid-Latitude West Coasts

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

Identification

• Journal: Water
• Year: 2026
• Date: 2026-09-09
• Authors: Heeseung Chung, Cheong Ghil Kim
• DOI: 10.3390/w18182235

Research Groups

• Department of Atmospheric Sciences, University of California, Los Angeles (UCLA)
• Department of Geophysics, University of Chile

Short Summary

This study improves the prediction of atmospheric river (AR) persistence using a pre-trained transformer model, TabPFN, which corrects the weak performance of the Global Ensemble Forecast System (GEFS) forecast. The proposed method significantly enhances the accuracy of AR persistence prediction in California and Chile.

Objective

• Investigate the effectiveness of TabPFN in predicting AR persistence beyond existing numerical weather prediction models.

Study Configuration

• Spatial Scale: Regional scale, focusing on California and Chile.
• Temporal Scale: Daily to weekly time scales, considering the duration of atmospheric river events.

Methodology and Data

• Models used: Global Ensemble Forecast System (GEFS) v12 reforecast (2000-2019), TabPFN (pre-trained transformer for tabular data).
• Data sources: Single-control-member forecasts from GEFS reforecast, regional IVT threshold values.

Main Results

• The F1 score of AR persistence prediction increased from 0.414 to 0.502 and 0.601 in California and Chile, respectively.
• TabPFN outperformed machine learning models such as 1D-CNN and LGBM in both regions.
• The proposed model captured 66.3% of rainfall during persistent atmospheric river events on the California coast.

Contributions

• This study provides a novel approach to predicting AR persistence, which can contribute meaningfully to flood disaster prevention.
• The results demonstrate the effectiveness of TabPFN in correcting the weak performance of existing numerical weather prediction models.

Funding

• National Science Foundation (NSF) Grant # NSF-AGS-1924979
• Chilean National Research Council (CONICYT) Grant # FONDECYT 1191036

Citation

@article{Chung2026Machine,
  author = {Chung, Heeseung and Kim, Cheong Ghil},
  title = {Machine Learning Post-Processing of Atmospheric River Persistence Forecasts: A Pre-Trained Tabular Transformer Across Mid-Latitude West Coasts},
  journal = {Water},
  year = {2026},
  doi = {10.3390/w18182235},
  url = {https://doi.org/10.3390/w18182235}
}