Joint inversion of hydraulic and thermal parameter fields using a CNN-based framework and transient hydraulic tomography
Journal
Journal of Hydrology
Journal Volume
668
Start Page
134991
ISSN
0022-1694
Date Issued
2026-04
Author(s)
Abstract
Accurate reconstruction of hydraulic and thermal parameter fields is essential for understanding coupled groundwater flow and heat transport. This study develops a Thermal–Hydraulic Tomography Neural Network (THT-NN), a DenseNet-based convolutional encoder–decoder that jointly estimates permeability ((Formula presented) ), porosity ((Formula presented) ), thermal conductivity ((Formula presented) ), and specific heat capacity ((Formula presented) ) from transient hydraulic and thermal tomography data. A dataset of 20,000 coupled groundwater–heat transport simulations generated with TOUGH2 is used for training and evaluation. Under noise-free conditions, THT-NN reconstructs heterogeneity in all four parameters with high accuracy, especially for (Formula presented) and (Formula presented). When realistic measurement noise is introduced, estimation accuracy decreases, particularly for (Formula presented) and (Formula presented), because information relevant to these storage-related parameters is primarily encoded in early-time hydraulic and thermal responses, where signal-to-noise ratios are lower. Noise-adaptive training improves robustness and partially recovers the degraded accuracy. Bootstrap-based deep ensembles show low epistemic uncertainty for all parameters, indicating stable convergence of the learned inverse mapping. Generalization experiments show that THT-NN remains robust to moderate perturbations of geostatistical moments and correlation scales for (Formula presented), (Formula presented), and (Formula presented), whereas (Formula presented) shows degraded accuracy under both amplitude and structural perturbations, reflecting its strong dependence on early-time thermal dynamics. Overall, the results suggest that THT-NN provides a data-driven alternative for hydro-thermal parameter inversion and a promising basis for future field-scale applications.
Subjects
Contaminated site remediation
Convolutional neural networks
Ground source heat pump
Heterogeneous parameter estimation
Hydraulic tomography
Thermal and hydraulic properties
Publisher
Elsevier BV
Type
journal article