Andrew Zammit Mangion: Physics-based machine learning models for spatio-temporal forecasting

Atmosphere/ML Workshop 2026 Atmosphere/ML Workshop 2026 Abstracts
1

Andrew Zammit Mangion: Physics-based machine learning models for spatio-temporal forecasting

Andrew Zammit-Mangion

  1. Forecasting Earth system processes, such as ocean temperature and rainfall, requires models that can learn how spatial patterns evolve over time. Existing statistical models for these processes can be accurate over short time periods, but they often need frequent re-fitting when the dynamics change. This is computationally expensive and leads to the use of sub-optimal approximations.

  2. Here, we develop a hierarchical statistical forecasting model based on the underlying physics, with a deep convolutional neural network used to learn changing spatial dynamics from the process’ recent behaviour. After training, the model produces fast probabilistic forecasts using an ensemble Kalman filter, without repeated parameter estimation/calibration. We demonstrate the use of our model on forecasting daily sea-surface temperature data from the North Atlantic Ocean and weather radar nowcasting in Sydney.

  3. We show that our approach produces accurate and well-calibrated forecasts in various settings, and that the neural network provides a realistic, interpretable, and computationally efficient representation of time-varying dynamics. A key result is model transferability: a model trained on ocean temperature data is seen to produce high quality short-term weather radar nowcasts. This demonstrates how machine learning can be embedded within a probabilistic Earth system forecasting framework, supporting fast, uncertainty-aware prediction across different environmental applications.

  4. This talk will interest researchers in machine learning for Earth system science, environmental forecasting, spatio-temporal modelling, data assimilation, ocean prediction, and weather nowcasting.

  5. Keywords: Machine learning; Earth system forecasting; probabilistic forecasting; spatio-temporal modelling; convolutional neural networks; data assimilation; sea-surface temperature; weather radar nowcasting.