PhD: Explainable AI for Sustainable Sewers under Climate and Urban Stress

Project team

Exeter – Dr Jawad Fayaz

UQ – Professor Steven Kenway

Project description

Urban sewer and stormwater systems face escalating failures due to climate extremes and urbanisation. Ageing infrastructure, designed for historical rainfall patterns, now struggles with frequent “1-in-100-year” storms and urban sprawl, which increase toxic overflows, breach environmental regulations, and disproportionately harm marginalised communities. Traditional models like SWMM are computationally slow and lack scalability, while opaque AI methods risk biased outcomes. This project addresses these gaps by developing a responsible machine-learning framework that integrates climate resilience, equity, and cost-effectiveness into infrastructure management, aligning with UN SDGs 6 (clean-water) and 11 (sustainable-cities).

Objectives

• Develop physics-based ML models to simulate sewer networks as dynamic systems, targeting ≥90% modelling accuracy.
• Train an explainable decision-making agent to optimize interventions (e.g., pipe upgrades), balancing cost, equity, and compliance.
• Resolve data harmonization challenges across utility systems to ensure tool functionality.
• Validate outcomes through case studies in Brisbane and Exeter, targeting ≥20% overflow reduction.
• Deliver open-source tools and training modules for global utility adoption.

The framework combines physics-informed graph-neural-networks (GNNs), diffusion model, and explainable reinforcement learning (XRL) to simulate sewer/stormwater system behaviour, predict risks, and optimize interventions. GNNs act as surrogate digital twins, embedding hydraulic principles to model how land-use changes and extreme weather impact flows. Nodes (junctions, tanks) and edges (pipes) encode hydraulic and climate data, predicting vulnerabilities like overflows.

A generative AI diffusion model synthesizes high-resolution climate-urban scenarios by downscaling global climate maps (e.g., CMIP6) and integrating urban growth projections. Combined with Bayesian uncertainty analysis, the simulated scenarios are used alongwith GNNs to identify overflow hotspots and pressure deficits.

A multi-objective explainable reinforcement learning (XRL) engine then optimizes interventions against environmental, financial, regulatory, and equity goals. Explainability tools—saliency maps, counterfactual analyses, and GNNExplainer—quantify trade-offs and clarify how actions reduce risks, building trust, ensuring regulatory compliance, and minimizing service disparities. Auditable decision trails mitigate bias.

DevOps/API pipelines automate deployment, while interactive dashboards visualize risks, policy impacts, and intervention outcomes. This end-to-end approach balances technical precision with transparency, enabling utilities to preempt failures and prioritize equitable, low-carbon solutions.

Institutional Expertise:

• Exeter: Dr. Fayaz (physics-informed ML) and Prof. Javadi (hydraulic modelling) advance AI development using IDSAI’s GPU clusters, ISCA HPC, and Southwest Water’s utility datasets. Collaboration with HRWallingford enhances industry adoption and testing.
• UQ: Prof. Kenway (urban water systems) and Dr. Moravej (water networks) provide SWMM integration and field validation via ACWEB, utilizing Urban Utilities data. Dr. Gibbes (hydroinformatics) guides scenario generation and policy formulations.
• Collaboration: Joint workshops align tools with utility needs. Exeter develops the ML architecture; UQ validates models and deploys case studies.

Collaboration Phases:

• Months 1–18 (Exeter): Data collation, GNN/diffusion model development.
• Months 18–30 (UQ): XRL policy optimisation and historical validation.
• Months 30–36 (Both): Case study deployment in Brisbane/Exeter, open-source release.
• Months 36–42 (Exeter): Thesis completion, digital twin deployment.

Deliverables:

• Python code for a modular framework integrating physics-informed GNNs, diffusion model, and XRL agent, with DevOps/APIs for utility integration.
• Report analysing overflow reduction, cost savings, and equity improvements.
• Dashboard visualizing overflow hotspots and intervention impact.
• >2 peer-reviewed publications (e.g., Water Research) and training modules.

Contact

Questions about this project should be directed to Dr Jawad Fayaz J.Fayaz@exeter.ac.uk
Apply here: Award details | Funding and scholarships for students | University of Exeter