⚡ electrical engineering

Sensitivity Analysis of Shield Tunnel-Induced Ground Settlement to Stratigraphic Spatial Variability: From Random Field Simulation to Causal Decoupling

This paper presents an integrated framework combining random-field simulation, controlled numerical experiments, and symbolic regression to quantify how geological spatial variability in weathered granite affects shield tunnel-induced settlements, identifying the deformation modulus of the excavated and overburden strata as the dominant factors and deriving a physically interpretable predictive model.

Original authors: Chenlong Yue, Jian Zhang, Tugen Feng, Haibo Wang, Bin Zheng

Published 2026-08-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chenlong Yue, Jian Zhang, Tugen Feng, Haibo Wang, Bin Zheng

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 block of rock, but as a giant, messy layer cake. Sometimes the frosting is thick, sometimes thin; sometimes the sponge is soft and squishy, other times it's hard as a brick. This is the reality of geology, especially when we try to dig tunnels for subways through it. For decades, engineers have tried to predict how much the ground will sink (settle) when a giant tunnel boring machine (TBM) eats its way through. They used to guess based on simple formulas or look at the ground as if it were perfectly uniform, like a single flavor of ice cream. But we now know the ground is full of surprises. It has "spatial variability," which is a fancy way of saying that the soil changes its properties from one spot to the next, even over short distances. If you dig in one spot, the ground might be stiff; move a few meters, and it might be squishy. This unpredictability is a huge headache for engineers because if they guess wrong, the ground could sink too much, cracking the sidewalks and buildings above. So, the big question is: how do we predict this sinking when the ground is so messy and changeable?

This paper dives into that messy reality, specifically looking at shield tunnels (the kind that use a giant rotating cutter head) digging through weathered granite in Shenzhen, China. The researchers wanted to figure out exactly how much the ground's natural "messiness" affects the sinking, and more importantly, which specific part of the ground matters the most. They didn't just guess; they built a sophisticated digital twin of the tunnel using a powerful computer program called PLAXIS 3D. They simulated the ground as a "random field," which is like rolling dice to create thousands of different, realistic versions of the underground layers, each with slightly different strengths and stiffness. They found that just by changing the natural layout of the rocks and soil, the amount the ground sank could vary by about 3.38 mm. That might sound tiny, but in the world of tunneling, it's a huge difference that could mean the difference between a safe tunnel and a cracked building.

The most exciting part of their discovery is how they figured out why the ground sinks. They treated the ground like a two-part team: the "excavated stratum" (the layer the tunnel is actually cutting through) and the "overburden" (the layers of dirt and rock sitting on top of the tunnel). They ran 37 different computer experiments, changing one thing at a time, like a scientist tweaking a recipe. They found that the stiffness (how hard or soft it is) of the layer the tunnel is cutting through is the absolute boss. If that layer is stiffer, the ground sinks less. The stiffness of the layers on top matters too, but it's the second most important thing. Surprisingly, they found that other factors, like how "sticky" the soil is (cohesion) or the angle at which it slides (friction angle), barely made a difference at all. Even the exact number of layers or how they were stacked didn't matter much as long as the overall stiffness was the same. It's like realizing that when you're pushing a heavy box, the weight of the box matters way more than the color of the carpet it's sitting on.

To make this useful for real-world engineers, the team used a clever math trick called "symbolic regression." Think of this as a robot that looks at all their computer data and tries to write a simple, readable sentence (an equation) that explains the results, rather than just giving a black-box answer. They came up with a formula that uses just three numbers—the stiffness of the tunnel layer, the stiffness of the top layers, and the stickiness of the tunnel layer—to predict how much the ground will sink. This formula was surprisingly accurate, getting the prediction right within less than half a millimeter on average. They also used a tool called SHAP to double-check their work, which confirmed that the stiffness of the tunnel layer is indeed the superstar of the show.

The paper essentially argues against the old idea that we need to know every single tiny detail of the ground to make a good prediction. Instead, they suggest that if we focus on getting the "stiffness" of the right layers correct, we can make much better predictions without getting lost in the details. They explicitly ruled out the idea that the friction angle or the exact number of layers are major players in this specific type of granite ground. While their results are based on computer simulations and not physical digging in the real world, the consistency between their controlled experiments and their new formula gives them high confidence in their findings. They aren't claiming to have solved every tunneling problem in the universe, but they have provided a much clearer, simpler, and more accurate way to understand how the ground's natural variability affects tunnel safety, helping engineers build safer subways in the future.

Technical Summary: Sensitivity Analysis of Shield Tunnel-Induced Ground Settlement to Stratigraphic Spatial Variability

Problem Statement
Shield tunnelling in complex geological environments, particularly weathered granite formations characterized by irregular stratigraphic interfaces and pronounced stiffness contrasts, introduces significant uncertainty into ground settlement predictions. Traditional deterministic analyses often assume homogeneous soil properties or rely on mean values, failing to capture the dual nature of geological uncertainty: the discrete variability of stratigraphic sequences and the continuous spatial variability of geotechnical parameters within strata. Furthermore, existing sensitivity analyses frequently treat the entire soil profile as a unified system, overlooking the fundamental mechanical distinction between the excavation stratum (the source of deformation) and the overburden (the pathway for deformation transmission). This study addresses the need to quantify, explain, and predict settlement uncertainty arising from these spatial variabilities, specifically within the context of the Shenzhen Intercity Railway.

Methodology
The authors developed an integrated framework combining random field simulation, controlled numerical experiments, and symbolic regression:

Numerical Modeling and Validation: A 3D finite element model was established using PLAXIS 3D, simulating an Earth Pressure Balance (EPB) shield tunnel with a diameter of 8.0 m and an average burial depth of 20 m. The model utilized the Mohr–Coulomb constitutive model for soil and linear elastic models for structural components. It was validated against field monitoring data and the Peck empirical method, demonstrating high accuracy in reproducing surface settlement troughs in weathered granite.
Dual-Uncertainty Coupling Framework: To simulate geological spatial variability, the study employed a two-module approach:
- Stratigraphic Simulation: Coupled Markov Chains (CMC) were used to generate 11 independent geological realizations, capturing the discrete uncertainty of stratigraphic sequences (e.g., fluctuating weathering interfaces).
- Parameter Assignment: Conditional Random Fields (CRF) with exponential autocorrelation functions were applied within each realization to assign spatially varying values for deformation modulus ( $E$ ), cohesion ( $c$ ), and friction angle ( $\phi$ ), capturing continuous parameter variability.
Controlled Parametric Experiments: A series of 37 one-variable-at-a-time (OVAT) numerical experiments were conducted to decouple the effects of geological parameters. The study independently varied $E$ , $c$ , and $\phi$ for both the excavated stratum and the overburden, as well as the number of layers, to isolate their causal contributions to maximum surface settlement ( $S_{max}$ ).
Symbolic Regression and Interpretation: The dataset was expanded using Latin Hypercube Sampling (LHS) to cover the parameter space more uniformly. Symbolic regression was then employed to derive an explicit, interpretable prediction equation. Finally, SHAP (SHapley Additive exPlanations) analysis was applied to the resulting model to verify the sensitivity ranking and ensure physical interpretability.

Key Results

Impact of Spatial Variability: The 11 random-field realizations produced maximum surface settlements ranging from $-11.25$ mm to $-14.63$ mm, a variation of $3.38$ mm. This confirms that geological spatial variability alone can cause practically significant settlement uncertainty, which deterministic models based on mean values may underestimate.
Sensitivity Ranking: The controlled experiments established a clear sensitivity hierarchy: $E_{exc}$ $E_{e x c}$ (excavated stratum deformation modulus) $> E_{upper}$ $> E_{u pp er}$ (overburden deformation modulus) $> c_{exc}$ $> c_{e x c}$ (excavated stratum cohesion).
- Variations in $E_{exc}$ (35–75 MPa) caused a settlement change of $3.72$ mm.
- Variations in $E_{upper}$ (5–30 MPa) caused a change of $2.57$ mm.
- Variations in $c_{exc}$ (25–48 kPa) caused a change of $1.09$ mm.
- Friction angle ( $\phi$ ) and stratification complexity (layering) had negligible effects, with settlement variations below $0.18$ mm.
Mechanical Roles: The study identified that the excavated stratum acts as the deformation source, where both stiffness and, to a lesser extent, strength govern deformation generation. Conversely, the overburden acts primarily as a deformation transmission pathway, where stiffness is the sole governing factor; strength parameters ( $c$ and $\phi$ ) in the overburden showed no measurable influence on surface settlement.
Predictive Model: The derived symbolic regression equation, incorporating $E_{exc}$ , $E_{upper}$ , and $c_{exc}$ , achieved an $R^2$ of $0.806$, MAE of $0.265$ mm, and RMSE of $0.454$ mm. SHAP analysis on this model reproduced the sensitivity ranking ( $E_{exc} \gg E_{upper} > c_{exc}$ ), confirming the model's physical consistency.

Significance and Claims
The paper claims to advance shield tunnel settlement analysis from simple uncertainty characterization to a mechanism-based, causal decoupling approach. Its primary contributions include:

Methodological Innovation: The development of a dual-uncertainty framework that simultaneously models stratigraphic sequence uncertainty (via CMC) and within-stratum parameter variability (via CRF), offering a more realistic representation of weathered granite formations than deterministic or single-source stochastic models.
Mechanistic Insight: The explicit decoupling of the excavated stratum and overburden reveals that they play fundamentally different mechanical roles. This challenges conventional global sensitivity analyses that treat the soil profile as a single unit, providing a more physically meaningful basis for parameter evaluation.
Practical Utility: The study demonstrates that the deformation modulus is the dominant parameter controlling settlement, significantly more so than cohesion or friction angle. It proposes an explicit, interpretable symbolic regression equation that serves as a transparent alternative to "black-box" machine learning models. This tool facilitates rapid preliminary settlement prediction, geological parameter screening, and risk assessment without sacrificing physical interpretability.

The authors position this work as a step toward physics-informed predictive modeling in geotechnical engineering, bridging the gap between complex stochastic simulations and practical engineering decision-making.

More like this