Research

Geotechnical infrastructure is increasingly threatened by the escalating impacts of climate change, including flooding, landslides, and dam failures, which result in damages amounting to billions of dollars globally. These threats are closely related to three fundamental challenges: (i) uncertainty arising from incomplete knowledge of subsurface conditions, (ii) heterogeneity due to the diverse composition and behaviour of geomaterials across different geological settings, and (iii) nonlinearity stemming from the complex interactions between soil, rock, and infrastructure. Climate change is exacerbating these challenges, making it difficult for traditional analytical and numerical methods to accurately predict geomaterial and infrastructure behaviours, which is critical for designing resilient infrastructure, making informed decisions, and mitigating geohazard risks. This has created a pressing need for innovative approaches to enhance geotechnical engineering methodologies in anticipation of an increasingly uncertain and complex future.

To address these challenges, Dr Zhang's work aims to develop cutting-edge artificial intelligence methodologies as alternatives to traditional methods for accurate geomaterial and infrastructure response prediction in geotechnical engineering. He aspires to position AI as an indispensable tool in addressing the complex challenges faced by geotechnical engineers, ultimately transforming the field and contributing to societal betterment. Specifically, He has made successful attempts as below

The Physics-informed neural network method (PINN) has shown promise in resolving unknown physical fields in solid mechanics, owing to its success in solving various partial differential equations. Nonetheless, effectively solving engineering-scale boundary value problems, particularly heterogeneity and path-dependent elastoplasticity, remains challenging for PINN. To address these issues, this work proposes a hybrid computational framework integrating finite element method (FEM) with PINN, known as FEINN. This framework employs finite elements for domain discretization instead of collocation points and utilizes the Gaussian integration scheme and strain-displacement matrix to establish the weak-form governing equation instead of the automatic differentiation operator. FEINN is an innovative framework that combines neural networks with finite element methods (FEM) to solve elastic and elastoplastic boundary value problems, incorporating the heterogeneity and uncertainty of geomaterials. Built using in-house Python-based FEM code and PyTorch for deep learning, the framework achieves high precision, fast computation, and adaptability to measured data. Unlike traditional FEM, which is limited to forward analysis, FEINN supports both forward and inverse analysis by integrating measured data constraints into its loss function. This capability enables real-time decision-making, predictiion, and optimization in hazard prevention and mitigation. Additionally, FEINN is 10 to 100 times faster than conventional FEM, thanks to GPU acceleration. These features make FEINN a powerful tool for large-scale simulations, such as assessing infrastructure responses to climate change, urban flooding, landslides from extreme rainfall, and building collapse under earthquakes. Its versatility allows engineers and researchers to address complex geotechnical challenges with improved efficiency, accuracy, and cost-effectiveness. 3 related research papers have been published in Computer Methods in Applied Mechanics and Engineering and International Journal of Mechanical Sciences.

Reference:

• Zhang, N., Xu, K.P., Yin, Z-Y., Jin, Y. F., Li, K.Q. (2025). Finite element-integrated neural network framework for elastic and elastoplastic solids. Computer Methods in Applied Mechanics and Engineering, 433, 117474. https://doi.org/10.1016/j.cma.2024.117474
  
• Xu, K.P., Zhang, N., Yin, Z-Y., Li, K.Q. (2025). Finite element-integrated neural network for inverse analysis of elastic and elastoplastic boundary value problems. Computer Methods in Applied Mechanics and Engineering, 436, 117695. https://doi.org/10.1016/j.cma.2024.117695
  
• Zhang, N., Xu, K.P., Yin, Z-Y., Li, K.Q. (2025). Transfer Learning-Enhanced Finite Element-Integrated Neural Networks. International Journal of Mechanical Sciences, 290, 110075. https://doi.org/10.1016/j.ijmecsci.2025.110075

Large language models (LLMs) have achieved remarkable success in various industrial and research fields, enhancing work efficiency by assisting machines in comprehending human language. In geotechnical design where extensive repetitive cross-checking of design codes consumes considerable time and labour, the utilization of LLMs to enhance design procedures has not been explored before. The challenge is to ensure that LLMs accurately comprehend professional geotechnical information from text and execute mathematical calculations correctly. This work makes the first attempt at developing a specialized LLM framework, GeoLLM, integrated with an innovative prompt engineering strategy to extract professional information from text and enable accurate mathematical calculations. The findings indicate the promising capacity of GeoLLM to address professional tasks in geotechnical design.

Reference:

• Xu, H.R., Zhang, N.*, Yin, Z-Y., Atangana Njock, P.G. (2024) GeoLLM: A Specialized Large Language Model Framework for Intelligent Geotechnical Design. Computers and Geotechnics, 77, 106849. https://doi.org/10.1016/j.compgeo.2024.106849
  

Determining the distribution of soil layers through borehole data is a pivotal aspect of planning and executing geotechnical engineering projects. Due to the complex spatial features of geological formation, it remains a significant challenge to accurately predict geological cross-sections from limited borehole data. We have developed a series of intelligent models to forecast geological cross-sections using sparse borehole data, e.g. multi-spatial receptive field (MSRF) XGBoost approach. MSRF XGBoost encompasses classification and identification modules. The classification module exclusively employs sparse borehole data to train a series of MSRF XGBoost models for soil classification. The identification module leverages all the trained models to generate potential predictions of unknown soil strata, automatically pinpointing the optimal one via Gaussian filtering and boundary similarity algorithms. A new boundary accuracy criterion is proposed to assess the prediction capacity of different models.

Reference:

• Qiu, Y.S., Zhang, N.*, Yin, Z.Y., Wang, Y., Xu, C.J., Zhang, P., (2024). Novel multi-spatial receptive field (MSRF) XGBoost method for predicting geological cross-section based on sparse borehole data. Engineering Geology, p.107604.
  
• Qiu, Y.S., Yin, J.N., Zhang, N.*, Liu, H.C., Xu, C.J. (2024). Novel graph convolutional network for geological profile prediction using non-equidistant borehole data. Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards, 1–15. https://doi.org/10.1080/17499518.2024.2422489
  

Occlusions of granular particles in images significantly affect the accuracy of evaluating particle morphology for granular materials. In this work, a novel framework of SOLO-PCNet is proposed, which can automatically segment all the particles and predict the complete contours of the occluded particles in the densely packed materials. Firstly, the instance segmentation model SOLOv2 is trained for the prediction of all the detectable particles. Then a self-supervised learning algorithm PCNet-M is introduced for the inference of the complete contours of the occluded particles so that the prediction of SOLOv2 can be directly input to PCNet-M for the subsequent completion. Thereafter, the particle morphology characteristics including elongation, equivalent mean size, convexity, and circularity are automatically calculated.

Reference:

• Zhang, H., Yin, Z.Y., Zhang, N.*, Wang, X. (2024) A rapid segmentation and occlusion completion method for packed granular particles considering uncertainty. Canadian Geotechnical Journal, Accepted.
  
• Zhang, H., Yin, Z.Y., Zhang, N.*, Wang, X. Ding, Z., (2024). A scale-adaptive Mask R-CNN strategy for foreground particle segmentation and geometrical analysis of granular aggregates. Applied Soft Computing, p.111931.
  

The representation of soil stress-strain response by using neural networks have received considerable attention as a promising data-driven method. Recently, a magnitude-related accuracy issue on stress-strain response was exposed for the neural network-based method, where the accuracy had an apparent decay when predicting the low-magnitude stress and strain data. This work proposes an enhanced deep learning method to tackle this issue by the fair reallocation of weight gradient. The enhanced method can significantly improve the accuracy, extrapolation capacity, robustness of neural network-based methods. A rationality investigation is also conducted via an insight into the weight gradient variation in neural networks. The effectiveness of this enhanced method is verified by three stress-strain responses of soil: a raw synthetic stress-strain response for accuracy assessment, a noised synthetic response with Gaussian noise for robustness, and a limited measured response from laboratory test.

Reference:

• Zhang, N., Zhou, A., Jin, Y. F., Yin, Z-Y., Shen, S. L. (2023). An enhanced deep learning method for accurate and robust modelling of soil stress-strain response, Acta Geotechnica, 1-23.
  
• Zhang, N., Shen, S. L., Zhou, A., Jin, Y. F. (2021). Application of LSTM approach for modelling stress–strain behaviour of soil. Applied Soft Computing, 100, 106959.
  

Shield tunneling is a crucial method in modern civil engineering and construction due to its numerous advantages and applications. Intelligent construction and management techniques can significantly enhance the efficiency, safety, and quality of shield tunnelling projects. Shield tunnelling faces significant challenges, including settlement control, stable excavation, and precise steering, due to complex soil-structure interactions. These challenges often result in low construction rates, high costs, and increased damage risks in complex ground conditions. To advance sustainable shield tunnelling, I have developed a series of AI algorithms that integrate deep neural networks with evolutionary optimization techniques, enhancing the life-cycle construction process from site investigation to monitoring. (i) Prior to construction, subsurface geological conditions are accurately reconstructed using a multi-spatial receptive field (MSRF) XGBoost approach, graph neural networks, and Bi-LSTM models. These methods leverage sparse borehole data from site investigations, account for soil property uncertainties, and support informed decision-making. (ii) During construction, ground settlement, earth pressure, and tunnel alignment are predicted and regulated more effectively using sequence neural networks. (iii) Post-construction, tunnel deformation is monitored through a transformer-based method that analyzes single-camera video data to measure underground structure deformations. These algorithms enable precise predictions of shield tunnel behaviors under complex geological conditions, contributing to intelligent and sustainable shield tunnelling practices. 10 related research papers have been published in Geotechnique, Canadian Geotechnical Journal, Automation in Construction, Tunnelling and Underground Space Technology, Measurement, Acta Geotechnica, etc.

Reference:

• Xu, H.R., Yin, J.N., Zhang, N.* (2025). Transformer-based deformation measurement of underground structure from a single-camera video. Automation in Construction, accepted.
• Zhang, N., Zhou, A., Pan Y., Shen, S. L., (2021). Measurement and prediction of tunnelling-induced ground settlement in karst region by using expanding deep learning method. Measurement 183, 109700. doi: 10.1016/j.measurement.2021.109700.