⚡ electrical engineering

From Wetting-Front Coalescence to Urban Road Collapse: Interpretable Machine Learning of Hydro-Mechanical Failure under Coupled Rainfall and Pipeline Leakage

This study elucidates the hydro-mechanical failure mechanism of urban road collapse caused by coupled rainfall and pipeline leakage through physical testing and numerical simulation, revealing that leakage pressure is the dominant driver while rainfall lowers failure thresholds, and establishes an interpretable machine learning framework to accurately predict collapse risks and guide targeted infrastructure maintenance.

Original authors: Liming Xue, Qiangbing Huang, Jiawei Liang, Yuxuan Gou, Jun Dong

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Liming Xue, Qiangbing Huang, Jiawei Liang, Yuxuan Gou, Jun Dong

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the ground beneath our feet not as a solid, unchanging rock, but as a giant, invisible sponge. In the world of geotechnical engineering—the science of how soil and rock behave under pressure—this sponge is constantly being squeezed, stretched, and soaked. Two main things happen to it: sometimes water trickles down from the sky (rainfall), and sometimes water leaks out from hidden pipes buried deep below. When these two water sources meet, they can turn the soil from a sturdy foundation into a mushy, weak mess. This isn't just a puddle problem; if the soil gets too soft, the ground can suddenly give way, swallowing roads, cars, and buildings in a terrifying event called urban road collapse. While scientists have long known that leaks and rain are bad news, figuring out exactly how they team up to cause a collapse has been like trying to solve a puzzle in the dark.

This is where a team of researchers from Chang'an University steps in with a clever new approach. They wanted to understand the secret handshake between rain and leaking pipes that leads to disaster. Instead of just guessing or waiting for a real road to fall in, they built a massive digital playground. They created a computer model that simulates how soil gets wet, weakens, and eventually fails, running 120 different scenarios to see what happens when you change the depth of the pipe, the strength of the rain, and the pressure of the leak. To make sense of all this data, they didn't just use a calculator; they taught a smart computer program (a type of machine learning) to spot the patterns. Their goal was to create a "crystal ball" that could predict which underground pipes are most likely to cause a collapse, helping city planners fix the right problems before the ground gives way.

The Great Soil Sponge Showdown

So, what did our digital detectives find? They discovered that the collapse of a road isn't a random accident; it's a four-step dance of doom involving water, soil, and pressure.

Step 1: The Top-Down Soak. It starts with the rain. As rain falls, it soaks into the top layer of the soil, like a sponge absorbing water from above. This makes the soil heavy and weakens its grip, kind of like how a wet towel is easier to tear than a dry one.

Step 2: The Bottom-Up Leak. Meanwhile, a broken pipe deep underground is leaking water under pressure. This creates a bubble of wet, mushy soil around the pipe, pushing outward like an inflating balloon.

Step 3: The Great Merge. Here is the critical moment. The wet zone from the rain (coming down) and the wet zone from the leak (pushing out) eventually meet. When they connect, they form a continuous, super-soft pathway from the pipe all the way to the surface. It's like two armies of water meeting in the middle of a fortress and opening the gates.

Step 4: The Collapse. Once this soft path is formed, the soil above it can no longer hold the weight of the road and traffic. It sinks, cracks, and eventually collapses.

The Big Reveal: Who is the Boss?

The researchers ran their simulations to see which factor was the real villain: the rain, the depth of the pipe, or the pressure of the leak. The answer was surprisingly clear.

The Leak is the King. The study found that the pressure of the water leaking from the pipe is the single most important factor. In fact, it was responsible for about 86.42% of the changes in how deep the collapse went, 87.79% of the size of the mushy zone, and a whopping 92.77% of how deep the wetness spread. If the leak pressure is low, even a lot of rain might not cause a disaster. But if the pressure is high, the ground is in serious trouble.

Rain is the Amplifier. Rain doesn't usually start the collapse on its own, but it makes everything worse. It acts like a volume knob, turning up the danger. By soaking the topsoil, it lowers the threshold for failure, making it much easier for the leak to push the soil over the edge.

Depth is the Regulator. How deep the pipe is buried changes the game, too. The researchers found a "sweet spot" of danger. Pipes buried at an intermediate depth (around 2 to 3 meters) were the most likely to cause a big surface collapse. If the pipe is too shallow, the water doesn't have far to travel, but if it's too deep (like 6 meters), the soil above is thick enough to absorb the shock. But at that middle depth, the rain from above and the leak from below meet perfectly to create a disaster.

The Magic Crystal Ball

To make this useful for real cities, the team used a special kind of artificial intelligence called machine learning. They trained two different types of "brains" on their 120 simulation cases.

One brain, called GPR/Kriging, was incredibly precise, predicting the results of their simulations with a score of 0.997 to 0.998 (where 1.0 is perfect). It's like a super-accurate map.

The other brain, XGBoost, was the detective. It didn't just predict numbers; it explained why. Using a tool called SHAP, it confirmed the story the simulations told: the leak pressure is the boss, the rain is the helper, and the depth controls the path.

Why This Matters for Your City

The researchers didn't stop at just understanding the science; they built a "Risk Map." This map shows that if it rains a lot, the danger zone shifts. Suddenly, pipes that were previously safe (with lower leak pressure) become risky.

The big takeaway for city planners is simple: Don't just check every pipe equally. Instead, focus your attention on the "danger trio": pipes with high leak pressure, located at intermediate depths (2–3 meters), during periods of heavy rain. These are the spots where the "sponge" is most likely to fail. By using this smart, AI-powered tool, cities can prioritize their inspections, fix the leaks before the ground opens up, and keep our roads safe from sudden, scary collapses.

It's a reminder that even though the ground feels solid, it's always listening to the water above and below. And now, thanks to this study, we have a better way to listen back.

Technical Summary: From Wetting-Front Coalescence to Urban Road Collapse

Problem Statement
Urban road collapse poses a severe threat to public safety and infrastructure in rapidly urbanizing areas. While previous research has extensively studied pipeline leakage-induced erosion or rainfall-induced landslides separately, the hydro-mechanical coupling mechanism between rainfall infiltration and pipeline leakage remains insufficiently quantified. Specifically, the interaction between the downward-migrating rainfall wetting front and the outward-expanding leakage-induced wetted zone, and how their convergence transforms hidden hydraulic disturbances into surface collapse, lacks a comprehensive description. Existing methods face limitations: physical model tests are constrained by scale and cost; high-fidelity numerical simulations are computationally expensive for rapid screening of numerous pipe sections; and many machine learning models prioritize prediction accuracy over physical interpretability and mechanism consistency.

Methodology
The study proposes a mechanism-constrained dual-surrogate modelling framework that integrates physical testing, numerical simulation, and interpretable machine learning.

Physical Model Testing: A large-scale physical model (1.5 m × 1.0 m × 1.2 m) was constructed to validate the collapse evolution process. The test simulated intermittent rainfall (15 mm/h) and continuous pipeline leakage (0.01 MPa) to observe wetting-front migration, zone convergence, and surface settlement.
Hydro-Mechanical Numerical Modeling: A 2D plane-strain finite element model was developed, incorporating saturation-dependent strength and stiffness degradation. Key parameters (internal friction angle, cohesion, elastic modulus) were defined as functions of the degree of saturation ( $S_r$ ) using empirical relationships derived from triaxial tests on Hangzhou silty clay. The hydraulic behavior was modeled using the van Genuchten-Mualem (VGM) model, implemented via a USDFLD subroutine to update material properties dynamically.
Dataset Construction: Based on the validated numerical model, a full-factorial database of 120 numerical cases was generated. The input variables were pipe burial depth ( $D$ : 1–6 m), rainfall intensity ( $R$ : 20–80 mm/d), and leakage pressure ( $L$ : 50–90 kPa). The output responses were surface collapse depth ( $S$ ), near-saturated zone area ( $A_s$ ), and wetting-front depth ( $Z_w$ ).
Dual-Surrogate Modelling:
- GPR/Kriging: Used as a benchmark for high-accuracy interpolation and uncertainty quantification in the small-sample, low-dimensional parameter space.
- XGBoost: Selected for response prediction, risk classification, and SHAP (SHapley Additive exPlanations) interpretation to ensure the model's predictions align with physical mechanisms.
Risk Assessment: A comprehensive risk index ($RI$) was developed by normalizing and averaging the three response variables to classify scenarios into low, medium, and high-risk categories.

Key Results

Failure Mechanism: The study identifies a four-stage failure process: (1) shallow-soil pre-wetting by rainfall; (2) expansion of the leakage-induced wetted zone; (3) convergence of the upper (rainfall) and lower (leakage) wetting fronts, forming a continuous softened pathway; and (4) upward propagation of deformation leading to surface collapse.
Dominant Factors:
- Leakage Pressure ( $L$ ): The dominant factor, contributing 86.42%, 87.79%, and 92.77% to the variance in collapse depth ( $S$ ), near-saturated zone area ( $A_s$ ), and wetting-front depth ( $Z_w$ ), respectively.
- Rainfall Intensity ( $R$ ): Acts primarily as an amplifier, dissipating matric suction and lowering the failure threshold, contributing ~9–10% to variance.
- Burial Depth ( $D$ ): Regulates the spatial convergence path. Intermediate burial depths (2–3 m) were found to produce the most significant surface responses due to optimal wetting-front convergence, while deeper burial attenuates the response.
Model Performance:
- GPR/Kriging achieved exceptional accuracy on the independent test set ( $R^2$ = 0.997–0.998).
- XGBoost demonstrated strong predictive capability ( $R^2$ = 0.962–0.990) and, crucially, its SHAP analysis confirmed the physical mechanism: $L$ -domination, $R$ -amplification, and $D$ -regulation.
- The XGBoost classifier achieved 96.7% accuracy in distinguishing risk levels, with no high-risk samples missed in the test set.
Risk Zoning: Risk maps revealed that increasing rainfall intensity shifts high-risk conditions toward lower leakage pressures. High-risk zones are most prevalent at intermediate burial depths under conditions of high leakage pressure and strong rainfall.

Significance and Claims
The paper claims to provide an interpretable and efficient tool for identifying vulnerable pipeline sections and supporting targeted inspection and collapse prevention. Its primary contributions are:

Mechanistic Insight: It clarifies that rainfall and leakage are not merely additive water sources but interact to form a continuous softened pathway through wetting-front coalescence.
Methodological Framework: It introduces a dual-surrogate approach that balances high-accuracy interpolation (GPR) with interpretable risk classification and mechanism verification (XGBoost-SHAP), addressing the "black box" limitation of standard ML in geotechnical engineering.
Engineering Application: The proposed framework serves as a graded auxiliary tool for "preliminary screening-priority verification-field calibration." It enables engineers to rapidly screen pipe sections based on available data (depth, pressure, rainfall forecast) to prioritize resources for high-risk segments (e.g., those with high leakage pressure, strong rainfall exposure, and intermediate burial depth).

The authors modestly note that the risk levels represent relative rankings within the specific parameter space studied and are not absolute probabilities of field failure. Practical application requires further calibration using local geological conditions, specific pipe defect geometries, and field monitoring data. The study does not claim to replace detailed site-specific analysis but rather to optimize the allocation of inspection resources.

The Great Soil Sponge Showdown

The Big Reveal: Who is the Boss?

The Magic Crystal Ball

Why This Matters for Your City

Technical Summary: From Wetting-Front Coalescence to Urban Road Collapse

More like this