Machine Learning Post-Processing of Atmospheric River Persistence Forecasts: A Pre-Trained Tabular Transformer Across Mid-Latitude West Coasts

Abstract

Atmospheric river (AR)-driven flooding is a natural hazard that causes severe damage in many regions, and the damage escalates sharply once an AR persists beyond a certain duration. Existing studies and numerical weather prediction models, however, have focused mainly on AR occurrence and on the intensity of integrated vapor transport (IVT) at individual time steps, paying little attention to duration. California and Chile are both exposed to severe AR-related hazards, yet for the period since 2000, the Global Ensemble Forecast System (GEFS) forecast at a two-day lead, compared against the regional IVT threshold, correctly predicts persistence for only about 30% of the ARs that actually persisted for 24 h or longer. This study retains the predictive capability of the physics-based GEFS forecast while correcting its weak performance on persistence using TabPFN, a pre-trained transformer for tabular data. Using single-control-member forecasts from the GEFS v12 reforecast (2000 to 2019), the model predicts the minimum IVT over the target window and compares it with the regional threshold. In the performance evaluation, the F1 score, which combines the precision and recall of AR persistence prediction into a single measure, rose from 0.414 to 0.502 and 0.601 in the two regions, and TabPFN outperformed machine learning models such as 1D-CNN and LGBM. Moreover, on the California coast, the proposed model, through its persistence decisions, captured 66.3% of the rainfall that fell during persistent atmospheric river events, up from 30.7% for the raw forecast. The proposed model offers a forecast post-processing method for predicting AR persistence and can contribute meaningfully to flood disaster prevention.