This study examined whether early-lead ECMWF Artificial Intelligence Forecasting System (AIFS) total column water (TCW) fields at +0, +6, and +12 h could be mapped to later-valid MIMIC-like total precipitable water (TPW) fields at +24 and +30 h using a U-Net. Separate models were developed from 59–60 six-hourly samples spanning 16 days. Equal-input references included +12 h persistence, linear-tendency extrapolation, and three-input linear regression, whereas target-lead AIFS fields were retained as information-advantaged operational references. In the six-field chronological model-selection subsets, the U-Net produced the lowest cosine-latitude-weighted MAE and RMSE among the examined equal-input references, with reductions of approximately 26–29% relative to the three-input linear regression. The frozen models were subsequently applied without refitting to one forecast initialized at 00 UTC on March 10, 2026, and similar relative improvements over the equal-input references were observed. Target-lead AIFS calibrations generally produced lower errors than the U-Net. The absolute TPW discrepancy increased with the magnitude of integrated vapor transport (IVT). These results suggest that early AIFS forecast trajectories contain useful information for later-valid MIMIC-like TPW mapping, although the short development period and single out-of-period case limit conclusions regarding broader temporal generalization.

Abstract
1. Introduction
2. Data and Methods
    2.1. Data
    2.2. Development Dataset and Preprocessing
    2.3. U-Net Model Architecture
    2.4. Training Procedure
    2.5. Evaluation Metrics and ComparativeEvaluation Design
    2.6. Equal-Input Reference Methods
    2.7. Target-Lead Statistical Calibration References
3. Results
    3.1. Development-Set and Model-SelectionResults
    3.2. Equal-Input and Target-Lead ReferenceComparisons
    3.3. Single Out-of-Period Case Study withEarly-Window and Target-Lead References
    3.4. Integrated Vapor Transport Calculationand IVT-Stratified Error Diagnosis
4. Discussion
5. Conclusion
References

• Ju-Hyeon Park(Department of Environmental Education, Korea National University of Education, Cheongju 28173, Korea)
• Yun Seob Moon(Department of Environmental Education, Korea National University of Education, Cheongju 28173, Korea, Chungbuk Carbon Neutrality Center, Korea National University of Education, Cheongju 28173, Korea) Corresponding author
• Seong Un Kim(Chungbuk Carbon Neutrality Center, Korea National University of Education, Cheongju 28173, Korea)