Neural Network Predicts Dangerous Water Fractures Above Repeated Coal Mining

Deep beneath some of China’s most productive coalfields, mining engineers face a hidden threat that has nothing to do with roof collapses or gas explosions. When coal seams lying close together are excavated one after another, the rock above them fractures in ways that can open hidden pathways to groundwater aquifers. If those pathways reach a water-bearing layer, the result can be a catastrophic inrush of water into the mine. A new study published in Discover Artificial Intelligence shows that a compact artificial neural network, trained on anonymized engineering records from multiple mining projects, can predict the height of these dangerous fracture zones with remarkable precision, offering a potential lifeline for mines operating above vulnerable aquifers.

The zone in question is known as the water-conducting fracture zone, or WCFZ. It is the vertically connected portion of caved and fractured rock that forms above a mined-out panel, through which water can migrate from overlying aquifers down into workings, or from mine water upward into drinking supplies. Its height is the single most important number in deciding how much solid coal must be left in place as a waterproof pillar between a working face and an aquifer. Underestimate it, and the safety margin shrinks, raising the risk of a sudden flood. Overestimate it, and vast quantities of otherwise recoverable coal are sterilized, driving up production costs. Getting this number right is, in the researchers’ framing, not merely an academic exercise in rock mechanics but a decision variable that links geology, mine design, and water-hazard management.

What makes close-distance coal seam groups so treacherous is that repeated mining is fundamentally different from mining a single seam in isolation. When the upper seam is extracted, the roof caves, bedding planes separate, and a network of fractures spreads upward through the overburden. Some of those fractures later close as broken rock compacts, but closure does not restore the original stiffness or continuity of the rock mass. Residual discontinuities and loosened interfaces remain frozen into the structure. When a lower seam is then mined, the new stress disturbance is superimposed on this already damaged overburden. Old tensile cracks can reopen, shear displacement can mobilize along weakened bedding planes, and newly generated fractures can coalesce with inherited ones into a single continuous pathway. The researchers describe this as a three-stage process: inherited-damage formation, stress reactivation, and finally either fracture coalescence or mechanical separation, depending on whether the intervening rock layer retains enough bending capacity to stay intact.

Two variables emerge as the central characters in this story. The first is cumulative mining thickness, denoted M, which controls the total void space, the roof-subsidence demand, and the energy available for deforming the key strata above the workings. The second is interlayer spacing, D, the thickness of rock separating the two mining horizons. If D is small, the lower seam’s fracture zone can punch straight through the interburden and connect with the damage left by the upper seam, so the two disturbance zones evolve as a coupled system. If D is larger, part of the secondary disturbance is absorbed through bending and internal deformation of the intervening rock, reducing the chance of complete connection. Crucially, the study argues that the final fracture-zone height is governed by a conditional, nonlinear interaction between these two quantities rather than by their independent linear contributions, which is precisely why a simple empirical formula tends to fail.

To capture that interaction, the team built a back-propagation neural network with six inputs: mining depth, cumulative mining thickness, interlayer spacing, an overburden hardness coefficient, face length, and seam dip angle. All six can be estimated before mining begins, which is essential for a pre-mining prediction tool. The hidden layer uses hyperbolic tangent activation functions that allow smooth nonlinear transformations, while a linear output node produces the predicted fracture-zone height in meters. Training was performed with the Levenberg–Marquardt damped least-squares algorithm, a method well suited to small and medium datasets because it combines the stability of gradient descent with the fast convergence of Gauss–Newton optimization. After comparing hidden-layer sizes of 5, 7, 9, 11, and 13 neurons, the authors settled on a parsimonious 6-9-1 architecture, since adding more neurons improved validation error by less than one percent while multiplying the number of adjustable weights that had to be estimated from only 21 training cases.

The data behind the model came from 30 anonymized engineering cases drawn from close-distance coal seam mining projects, spanning mining depths of roughly 129 to 663 meters, cumulative thicknesses of 3.5 to 15.5 meters, and interlayer spacings of about 3.8 to 44 meters. The cases were divided into 21 training, 5 validation, and 4 test records using a stratified split that ensured low, intermediate, and high fracture heights appeared in every subset. On the four hold-out test cases, the network achieved a mean relative error of just 1.02 percent, a root mean square error of 0.71 meters, and a coefficient of determination of 0.9987. Recognizing that four cases is a thin basis for celebration, the team also ran repeated five-fold cross-validation with three repeats, which yielded a mean relative error of 4.03 percent with a standard deviation of 0.93 percent, a more honest picture of how the model behaves across different data partitions.

The comparison against rival methods was equally telling. On identical test cases, a support vector machine with a radial basis kernel achieved a mean relative error of 3.04 percent, a random forest managed 8.55 percent, and a traditional empirical formula trailed at 17.19 percent. The empirical formula’s weakness was systematic rather than random: it consistently underestimated the highest fracture zones because it lacks any term describing the coupling between cumulative thickness and interlayer spacing. In water-hazard assessment, the authors note, such persistent underprediction is more dangerous than symmetric random error, because it silently shrinks the calculated distance between the fracture zone and the aquifer above. The random forest struggled for a different reason: tree-based models approximate continuous responses with locally constant regions, and with only 21 training cases, some combinations of large thickness and small spacing were represented by very few samples in the terminal nodes.

Perhaps the most scientifically valuable part of the work is its effort to open the black box. Three complementary analyses converged on the same conclusion. An ablation study, in which each variable was removed and the network retrained under the same cross-validation protocol, showed that deleting cumulative mining thickness raised the mean relative error from 4.03 to 27.29 percent, by far the largest degradation, while deleting interlayer spacing raised it to 10.34 percent. Garson connection-weight analysis, which partitions the magnitudes of the trained network’s weights, assigned 43.5 percent of the contribution to cumulative thickness and 15.6 percent to interlayer spacing. Finally, a three-dimensional response surface revealed the interaction directly: fracture height rises steeply when cumulative thickness is large and spacing is small, but the gradient with respect to thickness flattens as the interburden thickens, exactly as the mechanical picture of disturbance superposition predicts.

To test whether the model could transfer beyond its training data, the researchers applied it, without retraining, to 25 entirely separate engineering cases that had played no role in normalization, hyperparameter selection, or weight estimation. The network achieved a mean relative error of 2.67 percent and a coefficient of determination of 0.9834, comfortably outperforming the support vector machine at 4.34 percent, the random forest at 8.80 percent, and the empirical formula at 13.28 percent. Residuals from the repeated cross-validation were used to construct an approximate empirical 95 percent prediction interval of plus or minus 6.35 meters, shifting the output from a single deterministic number to a range that engineers can compare against the thickness of available aquicludes and safety pillars. The authors are candid about the limitations: the sample is modest, the records come from heterogeneous projects, and factors such as mining sequence, key-stratum position, faults, and aquifer pressure are not explicitly included. They position the model as an engineering prediction aid rather than a sole basis for safety-critical design, to be checked against borehole observations, water-level changes, and microseismic monitoring as mining proceeds. Even so, the study offers a template for how machine learning can be applied to geotechnical prediction responsibly, pairing mechanism-informed inputs with resampling-based validation, interpretability analysis, and honest uncertainty estimates rather than a single seductive accuracy figure.

Subject of Research: Machine-learning prediction of water-conducting fracture zone height in repeated mining of close-distance coal seam groups

Article Title: Prediction of water-conducting fracture zone height under repeated mining in close-distance coal seam groups using a BP neural network

Article References: Zhang, D., Guo, J., Liu, F., & Pu, Z. (2026). Prediction of water-conducting fracture zone height under repeated mining in close-distance coal seam groups using a BP neural network. Discover Artificial Intelligence, 6(1), Article 1247. https://doi.org/10.1007/s44163-026-02321-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02321-w

Keywords: water-conducting fracture zone, close-distance coal seams, repeated mining, BP neural network, Levenberg-Marquardt algorithm, water inrush, overburden failure, interlayer spacing, cumulative mining thickness, ablation analysis, cross-validation, mine water hazard