Research vision

We work on three scientific problems - compound extreme events, abrupt changes and tipping points, and the influence of small-scale processes on large-scale climate - and build a direct path from that science to real applications and industry.

Our work rests on three pillars: bringing together heterogeneous climate data; making AI models explainable and physically grounded; and tackling these three problems with proper mathematical rigour.

(1) Compound Extreme Events and Climate Hazards. Compound extremes — heat and drought arriving together, flooding followed by infrastructure failure, wildfire season extended by anomalous wind patterns - are harder to predict, harder to plan for, and often more damaging than single hazards. They are also where the standard statistical toolkit breaks down, because it was built for individual variables. We work on methods for joint tail-risk estimation and rare-event simulation, drawing on dynamical systems theory and large climate model ensembles, and we use ML to connect coarse model output to the local scales where the actual risk lies. The connection to applications is direct: better compound hazard estimates feed into infrastructure planning, agricultural risk assessment, and adaptation services for communities. These are the outputs that industry partners — insurers, energy companies, water utilities — are already asking for.

(2) Abrupt Changes, Tipping Points, and Regime Transitions. Parts of the Earth system can shift abruptly — Arctic sea ice, the Atlantic overturning circulation, monsoon systems — and once they do, the shift may be very difficult to reverse. These transitions are not well captured by the gradual trend narrative of most climate projections, and detecting their approach is genuinely hard: the signals are weak, the noise is large, and the models disagree. We use operator learning and large-ensemble analysis to look for precursors of regime transitions in high-dimensional model and observational data, and we are honest about the limits of what ML can and cannot resolve here. The practical stakes are high — earlier warning of these shifts has clear value for long-range risk planning, for reinsurance, and for international policy discussions where the question of irreversibility matters.

(3) Cross-Scale Dynamics: How Small-Scale Processes Shape Large-Scale Climate. One of the genuinely open questions in climate science is how small-scale processes — boundary-layer turbulence, ocean eddies, convective organisation — feed upward to affect the large-scale circulation, seasonal predictability, and decadal variability. We study these pathways using ML model hierarchies trained on

CMIP6/7 large ensembles, with particular focus on the Signal-to-Noise Paradox: climate models turn out to be more skilful at predicting the real atmosphere than their own ensemble spread would suggest. Understanding why this happens, and correcting for it, matters for seasonal and decadal forecasts that energy companies, water managers, and agricultural planners actually use. It also connects to a broaderquestion we are working on with David Dritschel — what are the mathematical limits on what AI can ever predict in a chaotic system? — which we think is worth putting in front of Nature or Science.

(4) Data, Explainability, and the Route to Applications. The three priorities above share a set of technical challenges: how do you bring together reanalysis, satellite, in-situ, and model data in a principled way? How do you make sure an ML model is doing what you think it is, and that a practitioner can actually use the output? How do you evaluate whether an AI forecasting system is worth the compute it costs? These questions sit behind all our applied work, and we treat them as research problems in their own right. The tools we develop — physics-informed neural operators, XAI/SHAP attribution for atmospheric fields, the Energy-Adjusted Skill Score (EASS) — are written as open tools that the wider community can use.