Projects

Six globes showing ATLAS ensemble forecasts of atmospheric moisture

ATLAS

Probabilistic medium-range weather forecasting: latent rollouts of probabilistic transformers trained with spatial and spectral CRPS, with over 30x faster inference than diffusion and stochastic-interpolant models without loss of forecast skill.

An aircraft split into its CAD surface patches

CANTO

CAD-native transformer operators: aerodynamic prediction directly from CAD.

False-colour map of TESSERA pixel embeddings over fields and towns

Tessera v2

Scaling pixel-wise Earth observation foundation models.

Zonal jet time series from numerical integration and the Stochastic Latent Transformer

Stochastic Latent Transformer

Probabilistic emulator of stochastically forced zonal jets, about five orders of magnitude faster than direct numerical simulation.

Beta-plane jet time series across four beta values, numerical integration versus neural network emulation

Emulation across parameter regimes

Extending deep learning emulation across parameter regimes to assess stochastically driven spontaneous transition events (ICLR 2024 AI4DifferentialEquations workshop).

Reconstruction error of autoencoders versus PCA as the latent size grows

Data-driven exploration of mid-latitude weather

MRes thesis: an autoencoder compresses chaotic quasi-geostrophic dynamics 40x onto an inertial manifold, quantifying latent variability from the stochastic parameterisation.

Flood risk in Shipston-on-Stour

LSTM rainfall-runoff model for flood-risk assessment.

Informal-settlement exposure

Image segmentation and change detection in satellite imagery to quantify settlement vulnerability and damage after extreme weather in the Caribbean.

NOvA neutrino classification

BSc thesis: DANN-based CNN for neutrino event classification.