A Transformer-Based Surrogate for Distributed Groundwater Dynamics with Zero-Shot Cross-Basin Transfer and Precipitation-Regime Data Augmentation
Abstract
Large-scale physically based groundwater models generate spatially distributed simulations, but their cost limits the regional datasets available for training artificial-intelligence-based surrogates. Here, we developed and systematically evaluated a groundwater-adapted PredFormer surrogate for regional ParFlow-CLM simulations. The framework uses temporal-first factorized attention, pressure-residual prediction, prescribed precipitation-minus-evapotranspiration (PME) forcing, and static hydrogeological attributes. The model was developed and evaluated over the 34,000 km² upper Beijiang River basin (UBRB), with zero-shot transfer tested over the 24,000 km² Hejiang basin. Across 12 independent 720 h test segments, PredFormer maintained decimeter-scale water-table-depth errors. Upper-tail errors were associated with precipitation statistics, whereas abrupt wet–dry transitions coincided with temporal error fluctuations. PME reduced the overall mean absolute error from 0.151 to 0.099 m. Depth-stratified analysis showed that PME was most beneficial in shallow-water-table areas, whereas static attributes contributed most at intermediate depths and little in deep areas. PredRNN was more accurate initially, but PredFormer performed better at extended lead times, reducing the across-segment mean error at 720 h from 0.232 to 0.152 m. Without target-basin training, fine-tuning, or recalculation of standardization statistics, zero-shot transfer produced an overall mean error of 0.106 m in Hejiang, close to 0.099 m in the UBRB. Precipitation-regime data augmentation produced modest overall changes but clearer benefits when added samples matched prediction conditions; in a precipitation-active segment, the mean error decreased from 0.141 to 0.095 m. Together, these results show how groundwater state, training-data coverage, and cross-basin transfer shape the performance of Transformer-based groundwater surrogates.
Related
- Input-data perturbation in groundwater level prediction: a controlled experiment on accuracy–reliability trade-offs among data augmentation and Bayesian-type uncertainty estimation methods
- Trade-offs between predictive accuracy and distributional fidelity in surrogate modeling of extreme precipitation
- Anthropogenic groundwater declines may exert far-reaching climatic impacts
- The GRACE satellite mission and transformation of global groundwater science
- Global projections of groundwater sustainability through the end of the century
BibTeX
@misc{huang2026transformerbased,
title = {{A Transformer-Based Surrogate for Distributed Groundwater Dynamics with Zero-Shot Cross-Basin Transfer and Precipitation-Regime Data Augmentation}},
author = {Hui Huang and Aoqi Sun and Hoang Tran and Yongjiu Dai and Chen Yang},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x5qv3p},
url = {https://eartharxiv.org/repository/view/14828/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A Transformer-Based Surrogate for Distributed Groundwater Dynamics with Zero-Shot Cross-Basin Transfer and Precipitation-Regime Data Augmentation
AU - Hui Huang
AU - Aoqi Sun
AU - Hoang Tran
AU - Yongjiu Dai
AU - Chen Yang
PY - 2026
DA - 2026/09/05
PB - EarthArXiv
DO - 10.31223/x5qv3p
UR - https://eartharxiv.org/repository/view/14828/
AB - Large-scale physically based groundwater models generate spatially distributed simulations, but their cost limits the regional datasets available for training artificial-intelligence-based surrogates. Here, we developed and systematically evaluated a groundwater-adapted PredFormer surrogate for regional ParFlow-CLM simulations. The framework uses temporal-first factorized attention, pressure-residual prediction, prescribed precipitation-minus-evapotranspiration (PME) forcing, and static hydrogeological attributes. The model was developed and evaluated over the 34,000 km² upper Beijiang River basin (UBRB), with zero-shot transfer tested over the 24,000 km² Hejiang basin. Across 12 independent 720 h test segments, PredFormer maintained decimeter-scale water-table-depth errors. Upper-tail errors were associated with precipitation statistics, whereas abrupt wet–dry transitions coincided with temporal error fluctuations. PME reduced the overall mean absolute error from 0.151 to 0.099 m. Depth-stratified analysis showed that PME was most beneficial in shallow-water-table areas, whereas static attributes contributed most at intermediate depths and little in deep areas. PredRNN was more accurate initially, but PredFormer performed better at extended lead times, reducing the across-segment mean error at 720 h from 0.232 to 0.152 m. Without target-basin training, fine-tuning, or recalculation of standardization statistics, zero-shot transfer produced an overall mean error of 0.106 m in Hejiang, close to 0.099 m in the UBRB. Precipitation-regime data augmentation produced modest overall changes but clearer benefits when added samples matched prediction conditions; in a precipitation-active segment, the mean error decreased from 0.141 to 0.095 m. Together, these results show how groundwater state, training-data coverage, and cross-basin transfer shape the performance of Transformer-based groundwater surrogates.
ER -