Interpretable AI network pinpoints what drives 24-hour tropical cyclone intensity change

Tropical cyclone intensity can shift dramatically within a day, and capturing that evolution remains one of the hardest problems in operational forecasting because the underlying physical processes are strongly nonlinear. Deep learning models have improved intensity forecast accuracy, but their black-box character has limited the confidence forecasters can place in their outputs. A research team led by Professor Wei Zhong at the National University of Defense Technology in China has now proposed a forecasting framework built on Kolmogorov–Arnold Networks (KANs) that pairs high prediction accuracy with robust interpretability, converting data-driven outputs into analytical symbolic expressions that can be read against physical theory.

The framework, described in Atmospheric and Oceanic Science Letters, consists of three linked modules. An attribution score-guided recursive predictor pruning module identifies and retains the most influential input variables, a neural network optimization and prediction module carries out the intensity forecast, and a symbolic regression-based predictor contribution analysis module translates the trained network into symbolic expressions and quantifies how much each dominant predictor contributes to the predicted 24-hour intensity change. According to the team, this design aims not only to prioritize forecast accuracy but also to bridge the gap between data-driven forecasting and physical interpretability.

In evaluation, the pruning module effectively selected high-impact predictors, enhancing both model robustness and accuracy. On an independent 24-hour test set, the framework achieved a mean absolute error of 7.87 knots. The contribution analysis module then transformed the KAN model into symbolic expressions, allowing the researchers to quantify the contributions of the dominant predictors rather than merely ranking their importance.

The analysis yielded a physically coherent hierarchy of influences. Among climatology and persistence predictors, initial tropical cyclone intensity exerted the strongest negative contribution to 24-hour intensity change. Among dynamic predictors, the 850-hPa maximum symmetric tangential wind showed the largest positive contribution. On the thermodynamic side, the maximum potential intensity dominated, indicating that a more favorable future thermodynamic environment promotes cyclone intensification. Taken together, the relationships derived from the symbolic expression largely align with existing tropical cyclone intensification theory.

“Although the forecasting framework is data-driven, it can effectively capture the complex physical processes underlying TC intensity evolution, establishing a solid bridge between data-driven forecasting and physical interpretability,” said Professor Wei Zhong, one of the corresponding authors of the study.

As a next step, the researchers plan to extend the symbolic expressions to incorporate quadratic, exponential, and interaction terms, with the goal of disentangling linear and nonlinear contributions and deepening the understanding of the nonlinear dynamic processes that govern tropical cyclone intensity evolution.

Subject of Research: Interpretable deep learning network quantifies dominant predictors of 24-hour tropical cyclone intensity change

Article Title: Interpretable deep learning network quantifies dominant predictors of 24-hour tropical cyclone intensity change

Article References: Interpretable deep learning network quantifies dominant predictors of 24-hour tropical cyclone intensity change. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: Interpretable, deep, learning, network, quantifies, dominant, predictors, hour, tropical, cyclone, intensity, change