TCG: Physics-Guided ML for Tropical Cyclone Genesis Prediction
physics-guided machine learning for tropical cyclone genesis prediction, localization, and interpretation in the North Atlantic
Physics-Guided ML · Scientific ML · Weather & Climate AI · Interpretability
Graduate research at the Climate Risk & Extreme Dynamics Lab, Georgia Tech, with Aren Russ, Thomas Ehrmann, and Prof. Ali Sarhadi (Russ et al., 2026).
Physics-Informed Machine Learning for Sub-Seasonal Prediction of Tropical Cyclone Genesis — under review at Journal of Geophysical Research – Machine Learning and Computation.
Overview
Tropical cyclone genesis (TCG) prediction asks where and when a storm will form — a rare-event, spatiotemporal forecasting problem where skillful early alerts are decision-relevant. This work builds a physics-guided ML system over the North Atlantic basin that predicts, localizes, and explains genesis, combining physically motivated environmental predictors with modern sequence and segmentation architectures.
Key Components
- Physics-guided features: nine ERA5 channels with physical meaning for genesis — potential intensity, specific humidity, vertical shear, relative vorticity, saturation deficit, surface radiation, cloud cover, and ocean heat content
- Spatiotemporal architectures: PatchTimesformer and hierarchical Timesformer variants for classification, with a UNet stage for genesis localization
- Interpretability: SHAP and Integrated Gradients saliency to attribute predictions to environmental drivers
- Causal inference: PCMCI causal graphs over predictor time series, connecting learned importance to physical mechanisms
- Leakage-free evaluation: a strict four-way temporal protocol (train 1980–2017 / calibration 2018–19 / development 2020–21 / test 2022–23) replacing the original random split, whose overlapping windows inflated F1 from ~0.14 (strict temporal) to ~0.90
- Lead-time forecasting: causal τ-day-ahead prediction from 14 past days of ERA5 analysis, isolating intrinsic precursor skill from operational NWP transfer error
Resources
- Code: TCG (training, evaluation, SHAP/IG saliency, PCMCI, Slurm pipelines)
References
2026
- JGR-MLCPhysics-Informed Machine Learning for Sub-Seasonal Prediction of Tropical Cyclone GenesisJournal of Geophysical Research – Machine Learning and Computation, 2026Under review