Yang et al. (2026) A hybrid physics–ML framework for integrating groundwater dynamics into land surface modeling: a ParFlow-based demonstration in headwater watersheds of the Pearl River Basin, China
Identification
- Journal: Geoscientific model development
- Year: 2026
- Date: 2026-11-11
- Authors: Chen Yang, Hui Huang, Zeyu Tang, Yongjiu Dai, Xiaobin Chang
- DOI: 10.5194/gmd-19-10719-2026
Research Groups
- School of Atmospheric Sciences, Sun Yat-sen University, Zhuhai, China (Chen Yang, Hui Huang, Zeyu Tang, Yongjiu Dai)
- School of Artificial Intelligence, Sun Yat-sen University, Zhuhai, China (Xiaobin Chang)
Short Summary
This study develops a hybrid physics-machine learning framework to integrate three-dimensional groundwater dynamics into land surface modeling, demonstrating an 18.24x computational speedup in headwater watersheds of the Pearl River Basin while maintaining high accuracy and preserving groundwater-land surface interactions.
Objective
- To develop and demonstrate a hybrid physics-machine learning framework for integrating three-dimensional groundwater dynamics into land surface modeling at substantially reduced computational cost, while preserving groundwater-land surface interactions and providing a practical pathway for Earth system simulations.
- To assess error accumulation over time and the effective forecast horizon of the hybrid system.
- To evaluate the hybrid system's ability to faithfully reproduce groundwater-land surface interactions.
- To determine if the hybrid system provides meaningful improvement over a simplified lower-bound representation (free drainage) and how this can be used to constrain acceptable model error.
Study Configuration
- Spatial Scale: Three headwater watersheds (Zhenjiang River, Wujiang River, Nanshui River) in the Pearl River Basin, China, covering approximately 34 000 km². The model domain consists of 252 × 146 grid cells at a horizontal resolution of approximately 1 km. Vertically, the model extends to a depth of 103.43 m, discretized into 11 layers.
- Temporal Scale: Hourly time steps for flux exchanges. A four-year spin-up was performed using meteorological data from water year (WY) 2019. The model was then driven by forcings from WY2020 and WY2021 for simulations. Surrogate training used data from WY2019–2020, validation used the first 720 hours of WY2021, and the main evaluation period spanned 8040 hours (hours 721–8760) of WY2021. Segmented predictions were also performed over 720-hour (approximately 1 month) windows.
Methodology and Data
- Models used:
- Physics-based reference: Coupled CoLM/ParFlow model (Common Land Model coupled with ParFlow).
- Land Surface Model: Common Land Model (CoLM).
- Physics-based Groundwater Model: ParFlow.
- Deep Learning Surrogate: A PredRNN-based deep learning architecture adapted from FSTR (a variant of PredRNN), designed to emulate the 3D pressure head field from ParFlow.
- Hybrid System: CoLM coupled with the ParFlow deep learning surrogate.
- Benchmark: PFFD (ParFlow Free Drainage) configuration, where lateral groundwater flow is disabled and the bottom boundary is free-drainage.
- Data sources:
- Meteorological forcing: ERA5-Land reanalysis for WY2019, WY2020, and WY2021.
- Soil texture (upper 10 layers): GSDE global dataset.
- Permeability (bottom 100 m layer): GLHYMPS global dataset, adjusted using an e-folding decay function.
- Land cover: MODIS MCD12Q1.061 product.
- Terrain slope: Calculated from a digital elevation model (DEM) after hydrological processing.
- Initial conditions and spin-up data: Multi-year mean precipitation minus evapotranspiration (P-ET) from the national-scale CONCN model.
Main Results
- The hybrid system achieved an 18.24x computational speedup compared to the fully physics-based CoLM/ParFlow model.
- Over the 8040-hour evaluation period, the hybrid system maintained strong agreement with the physics-based reference, with a median water table depth (WTD) error within 0.5 m.
- The domain-averaged latent heat flux from the hybrid model achieved a Kling–Gupta efficiency (KGE) of 0.959 and a mean bias of 1.88 W m⁻² relative to the physics-based model.
- The free-drainage (PFFD) configuration systematically underestimated latent heat flux (KGE of 0.564, mean bias of −24.59 W m⁻²), demonstrating the hybrid system's superior ability to capture groundwater feedback.
- WTD bias increased systematically with increasing WTD (deeper water tables), and latent heat bias showed a positive, nonlinear relationship with WTD bias, indicating error propagation from subsurface to surface fluxes.
- Segmented simulations (e.g., 720-hour windows) effectively suppressed autoregressive error accumulation, maintaining high accuracy (e.g., KGE of 1.000 for latent heat flux in a high-variability segment), supporting multi-timescale simulation strategies.
Contributions
- Presents a novel hybrid coupling framework that integrates a deep learning surrogate for groundwater dynamics into land surface modeling, replacing a physics-based solver while preserving the original coupling interface.
- Demonstrates a practical and computationally efficient pathway (18.24x speedup) for incorporating three-dimensional groundwater processes into Earth system simulations.
- Introduces an error-control strategy based on a free-drainage lower bound, allowing for application-dependent fidelity criteria.
- Provides one of the first efforts to couple an independent hydrologic surrogate model with a land surface model, advancing the paradigm of hybrid physics-ML Earth system modeling.
- Shows the feasibility of multi-timescale, on-demand simulations, where reinitialization can mitigate error accumulation for improved short-term accuracy.
Funding
- National Natural Science Foundation of China (grant nos. 2025HY00260101 and 2025HYSPT0705)
- Fundamental Research Funds for the Central Universities – Young Faculty Development Program (grant no. 25hytd008)
- Guangdong Major Project of Basic and Applied Basic Research (grant no. 2021B0301030007)
Citation
@article{Yang2026hybrid,
author = {Yang, Chen and Huang, Hui and Tang, Zeyu and Dai, Yongjiu and Chang, Xiaobin},
title = {A hybrid physics–ML framework for integrating groundwater dynamics into land surface modeling: a ParFlow-based demonstration in headwater watersheds of the Pearl River Basin, China},
journal = {Geoscientific model development},
year = {2026},
doi = {10.5194/gmd-19-10719-2026},
url = {https://doi.org/10.5194/gmd-19-10719-2026}
}
Original Source: https://doi.org/10.5194/gmd-19-10719-2026