IAP-25-086

Modelling urban infrastructure interdependencies under flood conditions via dynamic graph learning

Urban transport and drainage systems are tightly coupled, both physically and operationally, and their interdependencies become most critical during floods. When heavy rainfall overwhelms drainage capacity, inundated roads disrupt mobility, block emergency access, and delay maintenance operations [2]. At the same time, the transport network actively shapes the behaviour of the drainage system: road surfaces built of asphalt increase surface runoff, alter flow paths, and concentrate discharges into specific inlets [3]. Traffic congestion further amplifies these effects by delaying pumping, clearance, and repair activities. These physical and operational feedbacks form bidirectional loops that evolve rapidly during storm events, yet are poorly represented in existing models that treat infrastructure systems as static or independent [1,4].

This PhD will develop a dynamic graph learning framework to capture and predict these co-evolving processes. Each subsystem, transport and drainage, will be represented as a time-varying graph, whose structure (e.g. link closures, flow diversion) and attributes (e.g. flood depth, travel time, discharge rate) change through time [5]. The two layers will be dynamically coupled, so that flooding modifies road connectivity and mobility patterns, while the spatial and material characteristics of the road network (e.g. impermeable surfaces, slope gradients) influence drainage loading and flood propagation [6]. Temporal graph neural networks (GNNs) with attention mechanisms will be employed to learn these evolving dependencies from simulation and observational data [7], enabling the identification of cascading effects and key points of failure. The resulting framework will provide both predictive capability and interpretability, helping planners test interventions that reduce cross-network vulnerability and maintain emergency accessibility during extreme rainfall.

In recent years, the convergence of network science and machine learning has opened new avenues for modelling the complex dynamics of urban infrastructure systems. Yet, most applications in transport and flood modelling still rely on static or siloed representations, neglecting the evolving, bidirectional dependencies between networks such as roads and drainage. Traditional models are well suited to representing hydraulic or traffic processes independently, but they fail to capture how these systems co-evolve and influence one another during extreme events. Dynamic graph learning offers a novel framework to bridge this gap by combining time-varying graph representations with deep learning architectures capable of learning structural and temporal dependencies directly from data [1–3].

A dynamic graph represents a system whose topology and attributes evolve over time. In this project, the coupled transport–drainage infrastructure will be represented as a multilayer dynamic network, in which each layer (transport and drainage) is a time-varying graph that evolves with changing conditions. Within each layer, nodes such as junctions, manholes, pumps, and intersections, and edges such as pipes, conduits, or road segments, will vary in state according to flooding, operational interventions, or loss of connectivity. The two layers will interact through dynamic cross-layer connections that represent their bidirectional dependencies. For example, flooding in the drainage layer may deactivate or degrade edges in the transport layer, while impermeable road surfaces and traffic congestion in the transport layer can increase runoff and delay drainage recovery [4–6].

The use of dynamic GNNs enables the learning of these complex temporal dependencies. Unlike conventional deep learning models, which treat data as independent samples, GNNs operate directly on the graph structure, capturing spatial and relational patterns among components [7]. Temporal encoders such as gated recurrent units and attention mechanisms will enable the model to learn how relationships evolve over time and how cascading effects emerge across layers. The resulting network will offer both predictive capability and interpretability, allowing the identification of key nodes or links whose disruption disproportionately affects urban resilience.

The research will pursue four main objectives, each with corresponding activities:

1. Development of a dynamic graph representation of coupled infrastructure systems. Tasks: (1) Formally define the structure and temporal properties of the drainage and transport networks; (2) Curate datasets from hydrodynamic simulations (e.g., CityCAT, SWMM) and traffic data sources (e.g., OpenStreetMap, DfT feeds, satellite imagery), along with data from Glasgow Open Data Hub and the Newcastle Urban Observatory; (3) Represent the two systems as temporal graphs where nodes and edges evolve in structure and attributes; (4) Establish inter-layer connections that encode physical (e.g., impermeable surface runoff, drainage overflow) and operational (e.g., delayed maintenance response) dependencies; (5) Construct a database of time-stamped graph snapshots for model training and validation.

2. Design and implementation of the dynamic graph learning framework. Tasks: (1) Develop a graph neural network architecture capable of processing sequential graph states; (2) Employ temporal attention or recurrent mechanisms to capture both short-term and long-term dependencies; (3) Incorporate graph perturbation analysis to quantify the influence of specific nodes or links on cascading failures; (4) Implement coupling functions to simulate bidirectional interactions between drainage and transport layers; (5) Train the model using a combination of simulated and observed flood events, employing standard metrics such as prediction accuracy, resilience index, and recovery time.

3. Model validation and resilience analysis. Tasks: (1) Evaluate model performance under synthetic and real-world flood scenarios; (2) Conduct sensitivity analyses to examine how variations in permeability, rainfall intensity, or network connectivity affect outcomes; (3) Perform cascading failure analysis using attention weights and perturbation experiments; (4) Compare the proposed framework with conventional coupled models (hydrodynamic + traffic) to demonstrate computational efficiency and interpretability; (5) Quantify resilience indicators, including accessibility to critical services and drainage recovery speed.

4. Decision-support framework for resilient urban planning. Tasks: (1) Translate model outputs into interpretable resilience maps and accessibility indicators; (2) Develop scenario-analysis tools to assess interventions such as drainage retrofitting, permeable pavements, or adaptive traffic routing; (3) Collaborate with city stakeholders (e.g., Glasgow City Council, Newcastle City Council, water utilities) to test and refine the decision-support prototype; (4) Publish and disseminate the findings through open-source tools and academic outputs, ensuring reproducibility and practical impact.

Year 1

– Conduct a literature review on interdependent infrastructure, flood modelling, and graph learning.
– Gather and preprocess drainage and transport data from Glasgow Open Data and the Newcastle Urban Observatory, and/or other City Council and water utilities data sources.
– Build an initial multilayer dynamic network and run exploratory hydrodynamic simulations (CityCAT / SWMM).
– Present a short paper or poster at an internal or national event.

Year 2

– Develop the dynamic graph learning architecture and test initial training runs.
– Apply the model to a pilot case study (Newcastle) and refine based on results.
– Extend the framework to Glasgow for gaining impact and cross-city validation.
– Begin drafting the first journal article and a detailed methods chapter for the thesis.

Year 3

– Complete full model calibration and validation for both case studies.
– Carry out resilience analysis and scenario testing with city partners.
– Integrate decision-support tools and produce visual outputs for stakeholder workshops.
– Submit a second journal article and begin systematic thesis writing (introduction, literature, and methods chapters finalised).

Year 3.5

– Final synthesis of results and cross-city comparison.
– Complete and submit the PhD thesis.
– Disseminate findings through stakeholder presentations, code release, and conference talks.

The project will provide a comprehensive training programme, equipping the student with both specialist expertise and transferable skills for a career in academia or industry. The interdisciplinary nature of the research (combining hydrology, artificial intelligence, and urban systems analysis) requires targeted support for advanced technical and professional development.

Specialist Skills: The student will develop cutting-edge expertise in:
– Dynamic graph learning: Practical experience with time-varying graph representations and deep learning architectures, including Graph Neural Networks (GNNs) with temporal encoders.
– Hydrodynamic modelling: Proficiency in advanced flood simulation tools such as CityCAT and SWMM for data generation and model validation.
– Computational methods: Use of high-performance computing (HPC) and cloud platforms to train and test data-intensive models on large, complex datasets.
– Data curation: Integration and processing of heterogeneous datasets from simulations, transport feeds (e.g. DfT, OpenStreetMap), and urban observatories.

Professional Skills: Beyond technical development, the project will foster key professional competencies:
– Interdisciplinary collaboration: Close cooperation with supervisors and partners at Newcastle University and the University of Glasgow, complemented by links to an extended international research network working on digital water and flood resilience (e.g., collaborations with the Federal University of Minas Gerais, Brazil).
– Stakeholder engagement: Collaboration with city councils in Newcastle and Glasgow to co-design case studies and ensure that research outcomes address practical resilience challenges.
– Scientific communication: Presentation of research results at national and international conferences (e.g. IAHR, IWA, EGU) and preparation of manuscripts for peer-reviewed journals.
– Project management: Experience in planning, coordinating, and delivering a multi-stage research project with defined milestones and outcomes.