Energy piles—foundation elements that simultaneously provide structural support and exchange heat with the ground—offer a direct response to the growing heating and cooling demands of north-western Europe, where chalk is the primary bearing stratum for pile foundations. Despite the widespread occurrence of chalk and its frequently unsaturated state, its coupled thermo-hydraulic behaviour has not yet been characterised as a host medium for energy piles: existing estimates of chalk thermal conductivity correspond to discrete saturation states rather than to a continuous conductivity–saturation relationship established experimentally, representing a gap in geothermal foundation design. This study addresses that gap by delivering the first chalk-specific experimental dataset, constitutive framework, and validated numerical model for energy pile performance in unsaturated chalk. Thermal conductivity was measured on intact chalk specimens using a transient needle-probe across controlled degrees of saturation, and the data were used to calibrate a thermal conductivity function. A fully coupled thermo-hydraulic formulation was implemented in the finite-element framework COMSOL Multiphysics and verified against two independent benchmarks: a field thermal response test on an energy pile in saturated clayey sand and a laboratory model energy-pile test in unsaturated sand. The validated model was then applied to a three-dimensional field-scale energy pile in three chalk formations, and compared with representative sand, silt and clay, to examine the effect of saturation on thermal performance under a seasonal heating–cooling cycle. For the investigated configuration, increasing GWT depth from 0 to 29 m reduces the end-of-extraction heat rate by 6.1%, 8.8% and 11.9% for Estreux, Château-Landon and Seaford chalk, respectively, compared with 3.6% for clay, 8.6% for silt and 26.0% for sand. At 29 m, the chalk cases retain 159.4–169.9 Wm−1, compared with 135.5 Wm−1 for sand. This crossover highlights the potential of water-retentive chalk under deep groundwater conditions, subject to the adopted material properties and modelling conditions.
Fundamental challenges for improving supervised machine learning (ML) models for predicting geothermal resources include: 1) available training data being mostly positive (e.g., hot springs or operating power plants), so negative training sites must be selected, and in particular, negative training sites must represent the full range of conditions where no resources exist; and 2) input features potentially containing signals of several geologic processes, some of which may not be directly informative for modeling the availability of geothermal resources. We consider the artificial neural network (ANN) developed by the Nevada Machine Learning project (NVML), propose new selection strategies for negatives, and use signal separation to enhance the information content of input features. Using these strategies, we develop ML models using arguably simpler algorithmic approaches (Logistic Regression and eXtreme Gradient Boosting [i.e., XGBoost]) and compare the new results with the original ANN results.In particular, we evaluate the influence of selection strategy for negative training sites by comparing two training strategies: 1) using the original negative sites defined by NVML (NVML negatives); and 2) randomly selecting negatives from outside ellipsoids defined by NVML as having favorable structural settings, under the assumption that areas far from favorable settings are likely negative. Next, we explore the added value of splitting the topography input feature into two features to separate signals into regional elevation trend and local relative elevation, under the assumption that regional trends are correlated to regional geology (e.g., crustal thickness) and that local relative elevation (e.g., basin and range topography) is correlated to local geology (e.g., faulting). All new models using logistic regression and XGBoost were run 60 times each using Monte Carlo cross-validation to evaluate predictive uncertainty.Models using the randomly selected negatives outperform models using the negatives selected for the NVML ANN; models using XGBoost outperform models using logistic regression; and models using separated elevation features outperform models using the original elevation data. The results showing improved performance with random sampling of negatives demonstrates the value of sampling the entire input feature space when fitting a model (i.e., providing the ML workflow with all types of negatives and positives). The results showing improved performance with separated topography demonstrates that ML models cannot always find important patterns when they are obscured by larger trends, so feature engineering that enhances these patterns can be beneficial.
The Paris Basin hosts one of the longest-operating deep geothermal systems in Europe, with approximately 50 geothermal doublets exploiting the Middle Jurassic Dogger carbonate and Lower Cretaceous Albian siliciclastic aquifers and supplying nearly 2 TWh of renewable heat annually. Despite extensive geological and geophysical investigations, predicting reservoir facies, heterogeneity, and transmissivity remains highly uncertain. This study presents a physics-guided, multi-scale workflow that integrates petrophysics, rock physics, seismic inversion, and stochastic reservoir modeling to characterize the Dogger, Oxfordian, and Triassic geothermal reservoirs of the Paris Basin. The workflow combines four components: (i) estimation of mineral volumes and total porosity (ϕtotal) using a hybrid statistical–machine learning approach calibrated with X-ray diffraction (XRD) measurements; (ii) joint Monte Carlo inversion of water saturation (Sw) and formation water resistivity (Rw) under petrophysical constraints; (iii) generation of a rock-physics-guided synthetic pseudo-well catalogue and corresponding pre-stack Zoeppritz angle gathers, in which XRD-calibrated mineral end-members and petrophysically consistent elastic trends impose physical structure absent from a purely data-driven training set; and (iv) convolutional neural network (CNN) prediction of clay volume (Vclay), total porosity, and effective porosity (ϕeff) from pre-stack seismic data, pretrained on this synthetic catalogue and fine-tuned via transfer learning on seven real wells. Three additional wells, excluded from transfer learning, were reserved for semi-blind validation. At the calibration wells, R2 ranges from 0.858 to 0.985 across the three properties, with normalized mean-square error (NMSE) of 1.8%–18.0% and root-mean-square error (RMSE) below 0.043. At the semi-blind wells, performance varies with local well-to-seismic tie quality: one well reproduces calibration-level statistics (R2>0.92, NMSE <10%), while the other two show progressively degraded performance (R2 as low as 0.41, NMSE up to 113%), consistent with documented tie-quality issues at those locations rather than a general failure to generalize. The predicted properties reproduce major depositional trends, including Oxfordian lateral continuity and Dogger reservoir compartmentalization, without explicit stratigraphic conditioning. Effective porosity predicted along 1,990 km of interpreted 2D seismic profiles was then integrated into stochastic three-dimensional reservoir modeling, generating 100 equiprobable realizations that represent reservoir heterogeneity and uncertainty while remaining consistent with well observations and regional seismic trends.
