PhD in Physics-Aware Graph Neural Networks for Robust Short-Term Visibility Forecasting

Accurate short-term visibility forecasts are essential for safe and efficient transport operations, particularly aviation, where low visibility can reduce airport capacity and cause delays and cancellations. Yet visibility remains difficult to predict. Poor visibility is highly localised and governed by interactions between moisture, temperature, aerosols, surface conditions and atmospheric boundary-layer processes, many occurring at scales that numerical weather prediction (NWP) models cannot fully resolve. Current machine-learning approaches are also largely site-specific or used to post-process NWP, leaving both spatial relationships between observing locations and the physical information learned by ML models under-explored.

This PhD will develop physics-aware spatio-temporal graph neural networks (STGNNs) for 0–6-hour visibility forecasting, initially focusing on low-visibility events (<1000 m) in southeast England. The central hypothesis is that ML can exploit distributed observations to capture local atmospheric processes inadequately represented by NWP, and that connecting learned ML representations with atmospheric physics can improve model understanding, reliability and forecasting skill.

The project comprises six work packages.

WP1 Atmospheric processes, data and benchmarks: characterise low-visibility events using observations and NWP, identify relevant boundary-layer and local processes, examine situations in which NWP struggles, and establish persistence, NWP and ML benchmarks.

WP2 Graph-based forecasting: develop STGNNs that exploit spatial and temporal dependencies between observing sites for continuous visibility prediction, threshold classification, and prediction of event onset and cessation.

WP3 Linking machine learning model representations with atmospheric physics: analyse latent states, learned spatio-temporal relationships and feature importance to determine whether internal model representations correspond to physically meaningful atmospheric regimes and processes, and investigate why forecasts succeed or fail.

WP4 Physics-aware multi-source learning: combine observations and NWP fields, using insights from WP1–3 to investigate physically informed model inputs, graph structures and learning constraints, and quantify the value added by different information sources.

WP5 Robustness and failure diagnosis: develop methods to maintain performance when observations are missing, delayed or degraded, and distinguish data-related failures from unusual or poorly represented meteorological conditions.

WP6 Operational evaluation and translation: evaluate forecasting skill, reliability and failure modes with Met Office scientists using operationally relevant cases, and assess whether the framework can transfer to other local-scale weather phenomena.

The project brings together a distinctive supervisory team. Dr Zhou Zhou will lead the machine-learning and spatio-temporal modelling research. Prof. Georgios Efstathiou will contribute expertise in boundary-layer meteorology, turbulence and high-resolution NWP. Prof. Cyril Morcrette will provide complementary expertise in atmospheric processes, sub-grid representation and parameterisation, strengthening the connection between machine learning and operational weather modelling. Steven Ramsdale (Met Office) will provide operational forecasting expertise, relevant data and guidance on practical forecasting requirements and failure modes.

The student will work across Exeter's data-centric engineering and atmospheric-science communities and engage closely with Met Office scientists. Subject to availability, a Met Office placement or secondment will provide first-hand experience of translating AI research into operational forecasting practice. The project will therefore train the student at the interface of AI, atmospheric physics and operational meteorology.

Location

University of Exeter, Penryn Campus