The central theme of our research is integrating fundamental physics into AI models to improve reliability, interpretability, and generalisability for real-world infrastructure and construction problems.
Our work embeds first-principles civil engineering physics directly into machine learning models, enabling AI systems that remain grounded in the physical behaviour of the built environment. To achieve this, we combine methods from continuum and geotechnical mechanics, multimodal sensing, computer vision, machine learning, and multi-scale simulation. This integrated approach allows us to develop trustworthy AI systems capable of operating in complex, data-limited, and safety-critical infrastructure environments.
Some of the major contributions from our group include:
An AI-based digital inspector for new and ageing infrastructure, leveraging physics-based models for pre-training, which led to the spinout company InfraMind Labs
Game-changing “physics-supervised” learning approaches enabling dataset-free deep learning training
Integration of physical AI with ground penetrating radar to “see beyond the surface” by reconstructing subsurface structures
Novel patented sensing technologies combining AI and fibre-optic sensing, leading to two patents and the spinout company Sophone Force
Synthetic data simulators for pre-training infrastructure AI models
Physics-informed world models enabling reliable and generalisable robotic execution of complex construction tasks
Our Research Pillars
Sample Research Projects
Subsurface 3D Reconstruction Using Physics-Informed ML and GPR
Team: Haibing Wu (lead), Dr Guangming Wang, Dr Yixiong Jing, Dr Olaf Wysocki
This work introduces physics-supervised deep learning, synthetic data simulators and novel computer vision techniques method for 3D reconstruction beyond the surface. This work has led to zero-shot and few-shot reconstruction models as well as techniquesto generalize to unseen waveforms, overcoming one of the major bottlenecks for building GPR foundation models. The approach provides a data-efficient pathway toward scalable, physics-informed foundation models for subsurface imaging
Physics-Informed World Models for Generalisable Construction Robotics
Team: Dr Guangming Wang (lead), Dr Yixiong Jing, Dr Olaf Wysocki, Haibing Wu, Zhengyang Ling et al.
This work explores the use of synthetic simulators and domain adaptation to advance structural health monitoring (SHM) of tunnels. High-fidelity simulators are used to generate large-scale synthetic datasets of tunnel point clouds and defect scenarios, providing abundant and controllable training data for machine learning models. To bridge the gap between synthetic and real-world conditions, domain adaptation techniques are applied, enabling models trained on synthetic data to generalise effectively to noisy, incomplete, or variable-quality field measurements. This physics-informed, data-centric approach provides a scalable pathway for reliable tunnel inspection and anomaly detection, reducing dependence on expensive labelled datasets while improving the robustness of SHM systems in real underground environments.
Experimental Modelling of Polymer Flow in Sands
Team: Marina Bortolotto (lead), Daniel McNamara, Prof. Stephan Jefferis, Imperial, University of Oxford
This research investigates the permeation of polymer-based support fluids through sandy soils to better understand their role in stabilising excavations such as bored piles and diaphragm walls. Using a bespoke rigid-wall permeameter and complementary rheological testing, the study examines polymer flow through sands with different particle sizes and pore structures. Our research shows that polymer conductivity is typically at least two orders of magnitude lower than water due to their high viscosity and non-Newtonian shear-thinning behaviour, explaining their ability to provide effective excavation support even in highly permeable sands. Our findings also highlight the strong influence of soil density on seepage behaviour and provide new experimental insights that help reduce design uncertainty and support the wider adoption of polymer support fluids in geotechnical construction.
Physical Reasoning for Few-Shot Point Cloud Segmentation
Team: Tao Xinghui (lead), Dr Guangming Wang
This work explores the use of reasoning models with multi-agent systems for tunnel point cloud segmentation. The proposed framework is applied to tunnel point cloud segmentation in the first instance and reformulates tunnel segmentation as a multi-agent reasoning system. Large language models (LLMs) with chain-of-thought prompting, domain knowledge retrieval, tool use, and self-reflection are embedded as agents that adapt parameters, generate code, and iteratively improve results. Unlike traditional supervised or foundation-model pipelines, R4Tun reduces dependence on expert tuning and produces interpretable reasoning traces. Tested on the Seg2Tunnel benchmark, it demonstrates improved generalisation across varied tunnel geometries and emergent self-reflective behaviours. The approach represents a paradigm shift where AI agents not only replicate but also augment expert workflows, opening pathways to adaptive, interpretable, and insight-generating structural health monitoring.
Synthetic Simulators and Domain Adaptation for Pre-Training Infrastructure Foundation Models
Team: Wanru Yang (lead – tunnels), Dr Olaf Wysocki (lead – facades), Dr Yixiong Jing, Dr Guangming Wang et al.
This work explores the use of synthetic simulators and domain adaptation to advance modelling and interpretation of infrastructure. For tunnels, high-fidelity simulators are used to generate large-scale synthetic datasets of tunnel point clouds and defect scenarios, providing abundant and controllable training data for machine learning models. To bridge the gap between synthetic and real-world conditions, domain adaptation techniques are applied, enabling models trained on synthetic data to generalise effectively to noisy, incomplete, or variable-quality field measurements. This physics-informed, data-centric approach provides a scalable pathway for reliable tunnel inspection and anomaly detection, reducing dependence on expensive labelled datasets while improving the robustness of SHM systems in real underground environments.
Diffusion Models for Infrastructure
Team: Dr Yixiong Jing (lead), Dr Olaf Wysocki, Haibing Wu, Dr Guangming Wang
InfraDiffusion introduces a zero-shot framework for enhancing masonry point clouds. The method projects sparse and noisy LiDAR scans into depth maps using virtual camera projection and restores them with an adapted diffusion-based model (DDNM) enhanced by boundary masking. This produces clean, geometrically consistent depth maps without task-specific training or ground-truth supervision. By coupling these restored maps with the Segment Anything Model (SAM), InfraDiffusion significantly improves brick-level segmentation in tunnels and bridges. The approach addresses the challenges of fine-grained masonry inspection in low-light or data-sparse environments and opens new pathways for scalable, automated structural health monitoring.
Physics-Informed Machine Learning for Underground Construction
Team: Nandeesh Babanagar (lead), Xindong Zhai, Haibing Wu
Advanced 3D modelling and live site monitoring are common for urban basement construction, yet they remain largely disconnected. The absence of feedback between monitored behaviour and design models currently limits optimisation during construction. This challenge is particularly critical as the industry increasingly seeks to reuse old inner-city foundations, where monitored performance is essential both (a) to justify reuse and (b) to inform new superstructure design.
This project will, for the first time, develop physics-informed machine learning (PIML) digital twins for basements to enable design and construction optimisation. By embedding soil–structure mechanics within data-driven models, the digital twin will autonomously link 3D design models, structural analyses (e.g. finite element models) and state-of-the-art real-time monitoring. Monitored data will be assimilated “on-the-fly” through PIML to continually update and refine the design process, while lifetime behavioural records will be stored in the digital twin to support future foundation reuse. The full workflow will be validated on a live basement construction project in the UK.
Construction Site Monitoring & Modelling
Team: Dr Danny Murguia (lead), Prof Cam Middleton, Dr Zhengyang Lin, Dr Ashan Asmone, Dr Wanli Ma, Dr Olaf Wysocki
This research focuses on intelligent monitoring and modelling of construction sites using multi-modal data to provide real-time insights into project progress and emerging risks. The work integrates diverse data sources, including CCTV imagery, laser scans, audio recordings, and textual site reports, to create a richer understanding of construction activities and site conditions. The first stage of the project involved the acquisition of an unprecedented multi-modal dataset through close collaboration with industry partners, capturing real construction environments over extended periods. Building on this dataset, the second stage applies artificial intelligence methods to automatically interpret site activity, detect potential issues, and forecast problems before they escalate, ultimately enabling proactive decision-making and providing recommendations to improve construction performance, safety, and productivity.
Physics-Informed Machine Learning for Microtunnelling
Student: Malachi Philipson
This research explores how data-driven approaches, when combined with mechanistic or physics-based models, can improve predictions of soil–structure interaction, construction performance, and infrastructure resilience. Traditional empirical or purely numerical models often fall short due to limited accuracy, sensitivity to input assumptions, or a lack of generalisability across sites. By incorporating machine learning with physical constraints, we aim to create predictive models that are both more reliable with small datasets and more interpretable in real-world engineering contexts.
Trustworthy Digital Twins
Postdoc: Linjun Lu
This project is aiming to develop robust methods and frameworks to measure, ensure, and enhance the trustworthiness and ethical integrity of digital twins applied to highway infrastructure, facilitating safer, more reliable, and ethically responsible management of road networks.
Digital twins for as-built shield tunnels
Student: Mr Wei Lin
Deformation and defects in as-built shield tunnels present significant challenges for tunnel managers and engineers, particularly in urban environments. Determining the current mechanical status (e.g. stress level) of shield tunnels becomes critical to guide repair and reinforcement strategies. However, existing evaluation and analysis methods either lack precision or are prohibitively costly. The development of BIM and AI have brought promising opportunities for automatic and digital evaluation and analysis of existing tunnels. The objective of this work is to achieve automatic mechanical evaluation and analysis of as-built shield tunnels using digital twins by combining laser scanning, AI algorithms, BIM and finite element analysis.
Monitoring deep excavation construction
Student: Mr Pete Hensman
Industry sponsor: Ward and Burke Construction Ltd
This research aims to improve understanding of soil-structure interaction applied to deep bunkers for the construction of energy-from-waste facilities. This will be achieved through the use of real-time site monitoring and complementary finite element modelling. Data acquired from the field monitoring will be used to calibrate and verify the developed numerical models, whilst also creating a database for the development of simplified empirical methods. This field data can also be used to develop a system which provides real time feedback to site personnel, actively informing the construction process, and promoting a proactive approach to risk management and mitigation.
Design & performance of pipe-jacked tunnel drives
Student: Mr Bryn Phillips
Industry sponsor: Ward and Burke Construction Ltd
Microtunnelling is an increasingly popular means of trenchless installation for tunnels of between 600 mm and 3500 mm internal diameter and drive lengths of less than 1500m (typically). One of the greatest risks in long-drive microtunnelling is the development of excessive soil-structure penetration resistances, to the point where they exceed the total available jacking thrust. This project seeks to address this shortcoming through the development and monitoring of instrumented ‘smart’ tunnel pipes as well as bespoke reduced-scale tunnel-soil interface testing in the laboratory.
Satellite-enabled early warning system for geotechnical structures
Student: Ms Maral Bayaraa
Sponsor: Royal Commission for the Exhibition of 1851
Industry sponsor: Satellite Applications Catapult
From hundreds of km in space, satellite-Interferometric Synthetic-Aperture Radar (InSAR) analysis allow detection of ground motion to a high level of precision. However, complexities in converting the raw data to actionable information has so far limited authorities’ ability to practically monitor these structures from space. To help solve this, this research is creating a framework for geotechnical validation of the satellite measurements by developing 3D numerical models capable of simulating the underpinning mechanics of the observed soil movements. Deep learning experiments will be used to create an end-to-end scalable early warning tool, tested on historical failures and other geotechnical applications. The research will be focused on tailings dams and deep excavations.
Damage detection for critical infrastructure using computer vision & deep learning
Student: Yixiong Jing
PI: Professor Sinan Acikgoz
Ageing infrastructure places a higher demand on both the frequency and scale of inspections. These inspections are time-consuming, costly and are prone to human error. This project seeks to develop an automated, flexible, accurate and efficient process to identify existing damage in critical infrastructure. This will be achieved using a combination of state-of the art computer vision and deep learning techniques.
Behaviour of construction support fluids
Students: Mr Daniel McNamara, Mr Bryn Phillips, Ms Alicia Leong, Mr Diarmid Xu
Sponsors: Ward and Burke Construction Ltd (bentonite support fluids), EPSRC and Arup (polymer support fluids)
Bentonite- and polymer-based support fluids are widely used in the construction industry to provide temporary hydraulic support to unstable ground, for example in tunnelling works, caisson sinking and deep excavations. Whilst most civil engineering designs must account for the performance of support fluids during construction, they often do so in an ad hoc manner, relying primarily on the development of empiricism from accumulated field experience. This research will result in new design rules and testing techniques for construction support fluids, with significant benefits for the construction industry in terms of risk, material use, and time to construct. These objectives will be achieved through fieldwork, laboratory testing and theoretical analysis.
Design method updating using Bayesian machine learning
Sponsor: Royal Academy of Engineering, EPSRC
While significant advances have been made in prediction modelling, uncertainty surrounding the assessment of geotechnical parameters and underground conditions (e.g. karst caverns, faults,coal veins) remains the main barrier to accurate prediction of construction behaviour. Back-analysis of soil parameters within an observational framework has been applied widely for geotechnical applications such as slope stability, pile capacity and braced excavations. This study presents a Bayesian approach to update uncertain model parameters to inform rational strategies for optimising construction operations using the latest monitoring data acquired during the project. The proposed framework is applied to pipe jacking and caisson construction.
Anomaly detection for microtunnelling
Sponsor: Royal Academy of Engineering, EPSRC
The proliferation of data collected by modern tunnel boring machines presents a substantial opportunity for the application of data-driven anomaly detection (AD) techniques that can adapt dynamically to site specific conditions. Based on jacking forces measured during microtunnelling, this research explores the potential for AD methods to provide more accurate and robust detection of incipient faults. A selection of the most popular AD methods proposed in the literature, comprising both clustering- and regression-based techniques, are considered.
3D ground modelling for tunnelling using probabilistic machine learning
Collaborator: Dr Stephen Suryasentana (University of Strathclyde)
An effective site investigation (SI) campaign is critical for establishing the ground parameters for tunnel design. Optimal SI planning and interpretation is therefore an important element for the future progress of cost-effective underground infrastructure. However, choosing the appropriate SI campaign (number of tests, test locations etc.) to minimise ground-related risk and uncertainty is challenging, as there is a lack of guidance in the design codes for SI planning. Therefore, the SI campaign is usually planned heuristically using ad-hoc engineering judgment. Given the large-scale nature of SI campaigns for some underground infrastructure projects, such as HS2 and the Thames Tideway Tunnel, considerable savings can be realised if a more optimal, automated and adaptive approach is adopted. This project will address these limitations through the integration of state-of-the-art probabilistic machine learning (ML) into the planning process of SI campaigns for tunnelling.
Laboratory testing of construction-induced settlements surrounding caisson shafts
Student: Mr Matthew Willoughby
There is very little guidance in the published literature relating to the soil settlements induced by the construction of large-diameter caisson shafts. This is particularly problematic if these shafts are to be located in close proximity to existing infrastructure. This research involves the development of a new apparatus to investigate the settlement induced by the construction of large-diameter caisson shafts at model scale. Construction-induced soil settlements are measured using soil imaging techniques. The laboratory measurements are being used to inform the development of new empirical design rules and numerical models
Design & development of smart fibre optic force sensors
Students: Ms Irinka Lamiquiz-Pratt, Mr Jack Templeman
Sponsor: Royal Academy of Engineering
Existing sensor designs use increasingly more complex structures and strain sensing arrangements to measure specific force combinations (e.g. axial force, shear force) independent of temperature effects. This increases sensor cost as well as manufacture and calibration time. This work aims to deliver a step change in force sensor design, manufacture, and operation by fusing state-of-the art fibre optic strain sensing with artificial intelligence techniques.
