“Artificial intelligence can basically rebuild tunnel engineering.”

A key field of civil engineering, tunnel engineering often involves the design, construction and maintenance of underground structures in complex and uncertain geological conditions. Traditionally, this field has relied on empirical methods, deterministic models, and manual data interpretation. This may limit our ability to address underground environment variability.

The integration of artificial intelligence (AI) is revolutionizing tunnel engineering by providing powerful data-driven solutions. Techniques such as machine learning, deep learning, neural networks, and computer vision are applied to optimize geological engineering investigations, control tunnel boring machines (TBMs) and improve real-time risk assessment. These AI-driven approaches capture nonlinear patterns, adapt to new information, support rapid decision-making, and lead to more accurate and resilient tunnel designs.

In TBM tunnels, where machines interact with very variable and anisotropic rationale, AI tools can analyze large datasets from sensors and operational logs to predict ground behavior and machine performance. This not only increases operational efficiency and safety, but also reduces costs and delays. The hybrid model, combining artificial neural networks (ANNs) with support vector machines (SVMs) or particle swarm optimization (PSOs), demonstrated improved performance and further pushed the boundaries of predictive capabilities in modern tunnel structures.

Predicting TBM Performance

One of the most prominent applications of AI in tunnel engineering is modelling and prediction of TBM performance. AI models are used to predict machine behavior based on both historical and live data. The key performance indicators (KPIs) are:

• Intrusion Rate (PR): The speed at which TBM progresses with each revolution
• Advanced Rate: Total distance per unit time
• Torque and Thrust: Mechanical force required for excavation
• Specific Energy: Energy consumption per cubic meter of excavated material
• Cutter wear prediction: When to replace the prediction tool, and key cost factors.

Traditional methods for estimating penetration rates, torque, and thrust rely heavily on expert judgment or empirical correlations derived from site-specific data. These methods are often lacking in precise consideration of rock properties, operational parameters of TBM, and interactions between environmental conditions.

ANNS has proven that they can model these interactions by learning complex patterns from historical and sensor data without the need for predefined mathematical relationships. This is a computational model inspired by the structure and function of biological neural networks, as seen in the human brain. Anne is one of the most widely used AI technologies, especially in tunnel engineering. This is due to the ability to model complex, nonlinear relationships between input and output data. These models are often developed using multi-layer perceptron (MLP) architectures. This approximates a very nonlinear function and can be generalized across invisible data scenarios. Unlike traditional models, ANN does not require predefined mathematical formulas. Instead, they learn patterns and relationships directly from the data. As more data is fed into the model, its performance will improve and become more and more accurate over time.

The structure of an ANN consists of several important components. First, there is an input layer that receives raw data from the environment or dataset. This is followed by one or more hidden layers, and data is processed through interconnected nodes using mathematical operations. Finally, the output layer generates predictions or classifications of the model based on the processed information.

Each node, also known as a neuron, receives input, applies a mathematical transformation via an activation function (sigmoid or relu function, etc.), and passes the result to the next layer. Connections between neurons have associated weights to determine the strength of the signal that passes. During training, these weights are adjusted in a way that minimizes errors, allowing the network to learn and make more and more accurate predictions.

For example, the ANN model uses input features such as rockless compressive strength (UCS), brittleness index, soil aggregation, internal friction angle, water table conditions, and TBM control settings to predict operational metrics with high accuracy.

Early studies in the late 1990s (e.g., Grima et al, 2000; Kim et al, 2001) led to further research by Benardos and Kaliampakos (2004) who applied two hidden layers to ANNs to predict TBM performance. Their model used variables such as rock mass rating (RMR), overload depth, UCS, rock weathering, permeability, and water surface surfaces. This study demonstrated that Anne can help select the optimal tunnel alignment and ground improvement strategy. In particular, this model achieves relative prediction errors of less than 10%, indicating the accuracy and reliability of practical engineering use.

These results demonstrated not only the feasibility of AI-based predictions, but also improved performance under actual variability compared to deterministic methods.

Earth's pressure and stability modelling

Maintaining the stability of the tunnel's face is one of the most important safety concerns during mechanized tunneling, especially in soft soils and ground mixing conditions. Face pressure refers to the pressure applied to the tunnel surface, that is, the exposed end of a tunnel where excavation is actively carried out. This pressure is important in maintaining the stability of the tunnel surface and preventing collapse and ground movement. Misregulated facial pressure can lead to catastrophic consequences such as face collapse, ground collapse, or excessive ground bumps. Shielded TBMs are equipped to adjust facial pressure using systems such as slurries and earth's pressure balance mechanisms, but manual adjustments of these systems are difficult, especially when ground conditions change rapidly.

AI has proven to be extremely effective in global pressure modeling and control. Using machine learning algorithms such as ANN, Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and SVM, engineers can develop predictive models that estimate optimal facial pressure based on real-time input data from TBM sensors and geological logs. Input parameters include pre-speed, slurry injection volume, cutter head torque, soil plasticity, and groundwater depth.

These AI systems allow for dynamic, real-time adjustments and face pressure that essentially functions as a digital co-pilot for a TBM operator. There are two advantages. Improved excavation safety and reduced surface impact. Several studies cited in both Baghbani et al (2022) and Liu et al (2023) have demonstrated that such models can maintain facial pressure within optimal safety margins at various hierarchies and reduce the risk of ground failure.

Future tunnel systems can embed AI-based pressure modeling in closed loop control systems. Here, facial pressure, slurry flow and TBM speed are continuously adjusted based on predictive feedback rather than static operator settings. This opens the door to more autonomous tunnel operations.

Payment forecast

Tunnel-induced settlements are affected by many interrelated variables such as tunnel shape, depth, groundwater conditions, soil stiffness, and construction methods. AI models trained with geotechnical monitoring data including standard penetration test (SPT) values, moisture content, unit weight, elastic modulus of elasticity, and Poisson ratio, were used to predict surface deformation with impressive reliability. ANN is also the most widely used method here, but hybrid models that integrate PSO or SVMs show enhanced prediction capabilities by improving network parameters optimization and more effectively dealing with nonlinear classification boundaries, thereby increasing the accuracy and generalization of the model in complex geological engineering conditions.

Anne dominates the landscape, but other AI approaches emerge with special features. ANFI, which combines fuzzy logic and neural networks, is used when interpretability and rule-based inference are required. This is useful in scenarios where data is limited or where decision logic is important for engineering validation. Convolutional neural networks (CNNS) and long-term memory (LSTM) networks have also been investigated in limited studies. CNNs are suitable for analysis of spatial or image-based data, such as tunnel face images and rock discontinuity patterns. LSTM networks, on the other hand, can be applied to monitoring sensor data from TBMS or Grand Instruments during tunnel advance.

It's not that there's no such challenge

Of course, AI is not the ultimate solution. Baghbani et al and Liu et al note that they rely on a “black box” model that lacks transparency. Engineers still need to understand the geodynamics behind numbers, especially when model predictions conflict with field observations.

Data quality remains a hurdle. AI models are as good as the data they are trained. Inconsistent sensor calibrations, missing records, or biased datasets can loosen predictions. Therefore, many researchers have recommended a hybrid approach to combining physics-based modeling with AI to balance interpretability and performance.

AI is essentially restructuring tunnel engineering by providing powerful tools for prediction, automation and adaptive decision-making. From optimizing TBM performance and stabilizing facial pressure, AI improves both the safety and efficiency of underground construction. While challenges remain particularly surrounding data quality and model transparency, traditional engineering expertise and integration of AI demonstrate a shift towards smarter, more resilient tunnel systems. As research advances and real-time data become more accessible, AI is set up to become an essential asset in meeting the growing demand for underground infrastructure development.

reference

Baghbani, Abolfazl, et al. “Applications of artificial intelligence in geotechnology: a cutting-edge review.” Earth Science ReviewVol. 228, May 1, 2022, p. 103991, www.sciencedirect.com/science/article/abs/pii/s00128252222000757?via%3dihub, https://doi.org/10.1016/j.earsirev.2022.103991 [Accessed 5 Jun 2025].

Liu, Lianbaichao, et al. “Artificial intelligence in tunnel structures: a comprehensive review of hotspots and frontier topics.” Geohazard MechanicsDecember 1, 2023, https://doi.org/10.1016/j.ghm.2023.11.004 [Accessed 5 Jun 2025].

• Amy Daniel studied civil engineering at Bath University and was selected as one of five early career contributors. Ground Engineering Magazine for 2025.