PhD Studentship: Human-in-the-Loop AI for Equitable and Climate-Resilient Sewer Systems.

Project Description
Urban sewer and stormwater systems are increasingly stressed by climate change and rapid urbanisation. Intensifying rainfall, more frequent extreme events, and expanding impervious surfaces have increased sewer overflows that contaminate ecosystems, threaten public health, and breach environmental regulations. Infrastructure designed for historical rainfall regimes can no longer cope with the emerging “new normal”, where events once classified as rare now occur far more frequently. In England and Wales, sewer overflow impacts exceed £270 million annually and affect approximately 80,000 homes; in Australia, annual costs exceed AUD 982 million. These failures disproportionately affect vulnerable communities and require utilities to make timely, accountable intervention decisions under uncertainty.

Utilities must prioritise interventions such as pipe upgrades, storage expansion, real-time control, or maintenance under uncertain climate futures, constrained budgets, and regulatory obligations. Hydraulic simulators are physically detailed but computationally slow and calibration-intensive, limiting large-scale scenario exploration and optimisation. Purely data-driven approaches are faster but can produce opaque decisions and amplify inequities if trained on historically biased patterns or narrow objectives. There is therefore a clear methodological gap: the absence of a rapid, data-driven, physically informed and human-aligned decision framework capable of robust, equity-aware intervention planning under climate uncertainty.

This PhD will develop a decision-support framework integrating physics-informed machine learning, scenario generation, and human-in-the-loop preference-based reinforcement learning to prioritise climate-robust and equity-aligned interventions. The core innovation is embedding expert judgement, regulatory constraints, and equity objectives directly into policy learning using structured preference feedback and explicit constraints.

The framework has three coupled components. First, a physics-informed graph surrogate model will emulate network hydraulics at scale, representing pipes and assets as a graph and predicting flows, depths, surcharge conditions, and overflow likelihood under rainfall and operational boundary conditions. Second, scenario generation will create climate and urban growth stress-tests by downscaling projections and integrating plausible land-use changes, with uncertainty explicitly represented to evaluate robustness rather than single-point forecasts. Third, a preference-learning reinforcement learning agent will propose intervention portfolios. Stakeholder preference rankings derived through inverse reinforcement learning and constraints related to compliance, safety, cost, and equity will guide policy learning, ensuring auditable and human-aligned decisions.

Deliverables include validated surrogate models for rapid risk evaluation, scenario libraries for stress-testing, a human-in-the-loop optimisation engine for invertion decisions with auditable decision logs, and reusable software modules suitable for utility integration.

Research objectives
Build and validate physics-informed graph surrogate models for sewer network states and overflow outcomes. Develop climate and urban growth scenario generation with uncertainty. Implement preference-based reinforcement learning with explicit human and regulatory constraints. Evaluate policies on UK and Australian case studies using performance and equity metrics.

Key research questions

• How can stakeholder preferences and regulatory constraints be encoded to produce auditable, equity-aligned policies?
• How does climate and urban uncertainty alter robust intervention prioritisation?
• What performance and equity gains are achievable relative to hydraulic-only and purely data-driven baselines?

Contact
Questions about this project should be directed to Dr Jawad Fayaz at J.Fayaz@exeter.ac.uk

Advert information