The amount of precipitable water (PW) in the atmospheric column is a key limiting factor for precipitation amount, and NWP models often fail in rainfall forecasts due to underestimation of PW. Therefore, improving PW forecasts serves as a critical step towards achieving accurate precipitation forecasts. Environmental variations at remote moisture sources, which govern evaporation, inherit the potential to regulate the amount of moisture supplied to and transported in the atmosphere, eventually influencing PW variability. Machine learning models which lack physical knowledge of such atmospheric phenomena can benefit in further enhancing their performance by explicitly providing physical relations between predictors and the predictand. A machine learning model based on the XGBoost algorithm was set up to postprocess daily-mean NWP PW forecasts at 1 to 14-day lead times for the Kyushu region. The model used raw PW forecasts from NCEP GFS together with spatial distributions of surface temperature and surface wind speed from NCEP FNL as input features, while ERA5 total column water vapor served as the target variable. A novel methodology was developed to further improve forecast skill by constraining surface input features to moisture sources, identified using the Lagrangian-based particle transport model FLEXPART. RMSE variability with 5-fold cross-validation for events during summer (June to September) from 2016 to 2023 showed significant improvements at medium-term forecasts (5 to 13-day lead times), achieving a maximum additional RMSE reduction of 10.8% with the novel methodology. While NSE and the correlation coefficient also showed consistent improvements, it did not lead to improvements in bias. The positive results gained are an indication that machine learning models related to atmospheric processes with multiple input features can benefit from explicitly ingesting physical knowledge.
DANTANARAYANA et al. (2026) studied this question.