Climate AI

Interpretable representation learning for tropical cyclone genesis.

Learning from spatio-temporal meteorological tensors before cyclone formation, with a focus on what separates developing and non-developing systems.

Problem

Genesis is a rare transition, not just a label.

Tropical cyclone genesis involves complex environmental and dynamical changes. A useful model should not only predict whether a disturbance develops, but also help reveal the latent structure that separates developing systems from non-developing ones.

The research goal is to build representations of pre-genesis meteorological tensors and inspect how latent directions correspond to intensification, organisation, moisture, circulation, or other physically meaningful patterns.

Data

Spatio-temporal meteorological tensors.

Track source

OWZ events

Storm or disturbance event tracks used to define seed points and temporal windows.

Environment

ERA5 fields

Reanalysis variables extracted around event-centred spatial windows.

Variables

Humidity, temperature, wind

Multi-level atmospheric channels across standard pressure levels.

Tensor

T × C × H × W

Temporal movie-like input for MixtureBetaVAE3D-style modelling.

Pipeline

From OWZ event to pre-genesis tensor.

OWZ Event
Seed Point
ERA5 Match
Tensor Window
Latent Trajectory

Tensor concept

A compact view of the atmospheric movie.

t - 5
t - 4
t - 3
t - 2
t - 1
Placeholder note: Later this section can show real humidity, temperature, and wind fields as animated GIFs or interactive frames.

Scientific questions

What should the model help us ask?

What separates developing systems?

Latent structure may reveal environmental patterns associated with genesis.

When does organisation become visible?

Temporal latent trajectories can show whether development signals emerge before declaration.

Can counterfactuals remain physical?

Generated fields should preserve meteorological plausibility while varying relevant factors.

Next

See the project implementation side.

The projects page includes a placeholder for the tropical cyclone tensor builder and HPC workflow.

View projects