Machine learning reveals hidden danger in China’s gently dipping highway slopes

Deep in the mountainous terrain of southwestern China, a 40-meter-high highway cutting hides a geological trap that engineers have long struggled to see. The slope along the G5 Beijing-Kunming Expressway near Guangyuan looks deceptively ordinary: layers of grey sandstone and purplish-red mudstone dip gently, at only 13 degrees, into the hillside. Yet it is precisely this gentle inclination, combined with thin clay-rich seams barely half a meter thick, that makes such slopes so dangerous. A new study published in Environmental Earth Sciences by Le Zhang, Junyun Zhang and colleagues at Southwest Jiaotong University and the Sichuan Highway Planning, Survey, Design and Research Institute has now dissected exactly how these slopes fail, and has built a machine-learning system that can warn of trouble before disaster strikes.

Gently dipping bedding slopes are defined by a subtle geometry: the rock strata dip in the same direction as the slope face, but at an angle below 20 degrees. Because the dip is so shallow, these structures are frequently overlooked during geological hazard investigations and early monitoring campaigns. The consequences of that oversight can be catastrophic. When failure occurs, it can sever highway infrastructure and trigger substantial economic losses. Previous research has shown that the strength of weak interlayers plays a controlling role in the stability of such slopes, but most investigations have concentrated on slopes containing a single weak plane. Slopes threaded with multiple weak seams, like the one studied here, remain poorly understood, and the assumption of simple planar sliding used in conventional Chinese limit equilibrium analysis clearly diverges from the actual deformation behavior observed in the field.

The research team began with the slope itself, located in the deep excavation cutting section K8+280 to K8+660 of the Guangyuan-Mianyang segment of the expressway. Drilling revealed Quaternary residual and colluvial silty clay deposits roughly 4.5 to 5.5 meters thick, underlain by the Jurassic Lianhuakou Formation, an alternating sequence of light-grey feldspar-quartz sandstone and purplish-red silty mudstone, both moderately to strongly weathered and prone to disintegration in water. Between the sandstone and mudstone beds, clay-dominated weak interlayers 0.3 to 0.6 meters thick had developed. Critically, the relatively impermeable silty mudstone acts as a natural aquiclude, so these shallow weak seams remain saturated under continuous recharge from perched groundwater and sandstone fissure water. The rock strata dip at 13 degrees while the excavation slope stands at 53 degrees, a geometry that practically guarantees bedding slip once unloading begins.

To quantify how dangerous those saturated seams become, the researchers collected undisturbed samples using a static-pressure method, pressing PVC tubes into the exposed weak interlayers and sealing them with wax to preserve their natural structure and moisture. Laboratory direct shear tests were then run on specimens 61.8 millimeters in diameter and 20 millimeters high under normal stresses of 100 to 400 kilopascals. The natural-state soil, with a water content of 20.4 percent, exhibited strain-hardening behavior, with shear stress rising in a stepwise increase-stabilization-further increase-restabilization pattern. The team then immersed the samples for 1, 12 and 24 hours to simulate groundwater wetting. As water content climbed to 27.9, 36.4 and finally 39.1 percent, the shear strength curves shifted progressively downward. Fitting the results to the Mohr-Coulomb criterion showed cohesion falling from 24.06 kilopascals to 12.19 kilopascals and the internal friction angle dropping from 15.92 to 8.22 degrees. Both parameters decayed exponentially with water content, with coefficients of determination above 0.96. Remolded control samples proved dramatically weaker still, confirming that only undisturbed soil captures the true strength of these seams.

With reliable strength parameters in hand, the team built a two-dimensional finite element model in PLAXIS 2D spanning 220 meters, discretized with 15-node triangular elements into 10,586 elements and 85,585 nodes, with the mesh refined near the weak interlayers, excavation surfaces and protective structures. The construction process was simulated in four excavation steps and four protection steps, each including fully coupled flow-deformation analysis, with heavy rainfall represented by an equivalent infiltration flux of 15 millimeters per hour over 24 hours. The simulations revealed a striking pattern: after excavation-induced unloading, displacement developed preferentially along the bedrock-overburden interface and the weak interlayers within the slope. Once the first weak seam was exposed at the second excavation step, the maximum horizontal displacement shifted downward to that layer, and by the third step, obvious displacement ran along all four exposed seams. The slip surface took the form of a stepped composite path, alternating between gently dipping weak interlayers and steeply dipping tensile cracks, defining what the authors call a stepped sliding-tensile cracking failure mode.

The simulations also exposed a critical safety moment. The factor of safety dropped after each excavation step, but only at the second step did it fall below the requirement, reaching 1.13 against a threshold of 1.30, corresponding to a basically stable state. Once the prestressed anchor cable and frame beam system was installed at that stage, deformation was effectively restrained and the factor of safety rebounded to 1.56. This demonstrates that staged excavation with concurrent support, explicitly required in the construction specifications, is a genuine safeguard rather than a bureaucratic formality. Comparing simulated maximum horizontal displacements with field monitoring data showed generally consistent trends with a maximum difference of only 5 millimeters, validating the numerical model.

The real innovation came next. Rather than relying solely on computationally expensive simulations, the researchers ran 324 finite element analyses varying the cohesion, internal friction angle and elastic modulus of the weak interlayers, the construction stage, and the number of weak interlayers, generating a dataset with the factor of safety and maximum horizontal displacement as outputs. An orthogonal sensitivity analysis with five factors and three levels, evaluated through range analysis and analysis of variance, ranked the influences as number of weak interlayers greater than construction stage greater than friction angle greater than cohesion greater than stiffness, with the first three highly significant. The team then trained a back-propagation neural network on 260 samples, testing it on 64 held-out cases. After trialing hidden-layer sizes from 1 to 10 neurons across 30 repeated training runs each, the optimal architecture emerged as 5-10-2, yielding a root mean square error of roughly 0.10 for the factor of safety and a coefficient of determination near 0.97. Predicted factor-of-safety errors stayed below 0.1 and displacement errors below 1 millimeter.

The researchers then translated the trained network into a practical early-warning tool. Fixing the elastic modulus at 30 megapascals, the construction stage at completion, and the slope at five weak interlayers, consistent with the actual service condition, they predicted the factor of safety for different strength values and fitted its relationship with maximum horizontal displacement. Below 3.03 millimeters, the factor of safety is discretely distributed with no clear one-to-one correlation, but at or above that threshold a negative linear relationship emerges, described by the equation Fs equals -0.174 times the maximum displacement plus 2.178, with R-squared of 0.80. At 5 millimeters of displacement, the factor of safety equals 1.30, so the team adopted 5 millimeters as the safety warning threshold for the project. Field engineers can now read a displacement value from an inclinometer borehole and immediately know the slope’s condition.

An engineering application demonstrated the method in action. A deep inclinometer borehole installed at the monitoring section recorded deep displacement from January to April 2026, with maximum horizontal displacements at or above 5 millimeters throughout. Substituting these values into the analytical formula, which embeds the exponential strength-degradation relationships and accounts for water uplift and hydrostatic thrust in rear-edge tensile cracks, showed the slope remained in the yellow, basically stable state defined by Chinese highway subgrade design specifications. The prescribed response is daily inspection of retaining structures and close attention to displacement data, a proportionate rather than alarmist conclusion.

The study closes with a staged roadmap for practitioners. During feasibility studies and preliminary design, the analytical limit equilibrium formula corresponding to the stepped sliding-tensile cracking failure mode suffices. At detailed design, material properties should be accurately determined and stability evaluated through numerical simulation. During operation and maintenance, extensive simulations combined with field monitoring can feed machine-learning models that meet the practical need for rapid, convenient field assessment. The authors caution that their neural network cannot be directly transplanted to other projects, since slope geometry, geological conditions and weak-interlayer properties all govern model reliability. But the framework itself, pairing laboratory characterization of strength degradation with simulation-generated training data and field-validated thresholds, offers a template for taming one of engineering geology’s most quietly treacherous slope types.

Subject of Research: Earth Sciences

Article Title: Safety performance evaluation of a gently dipping bedding slope containing multiple weak interlayers in Southwest China

Article References: Zhang, L., Zhang, J., Gong, Z., He, C., Feng, J., Xiang, B., Zhang, Y., & Gao, F. (2026). Safety performance evaluation of a gently dipping bedding slope containing multiple weak interlayers in Southwest China. Environmental Earth Sciences, 85(17), Article 435. https://doi.org/10.1007/s12665-026-13178-8

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13178-8

Keywords: bedding slope, weak interlayers, slope stability, machine learning, neural network, finite element method, direct shear test, landslide hazard, geotechnical engineering, Southwest China, Safety, performance