Tropical cyclone (TC) intensity evolves rapidly and is governed by nonlinear processes, making it a central challenge in operational forecasting. Deep learning models have effectively improved TC forecast accuracy, but their "black-box" nature severely limits their credibility.
To address this issue, Professor Wei Zhong's research team at the National University of Defense Technology, China, proposed a TC intensity forecasting framework based on Kolmogorov–Arnold Networks (KANs) that integrates high prediction accuracy and robust interpretability. It comprises an attribution score-guided recursive predictor pruning module; a neural network optimization and prediction module; and a symbolic regression-based predictor contribution analysis module (see the framework diagram). This work not only prioritizes TC intensity forecast accuracy but also attempts to convert data-driven model outputs into analytical symbolic expressions, bridging the gap between data-driven forecasting and physical interpretability. The research results were recently published in Atmospheric and Oceanic Science Letters .
Results show that the pruning module effectively selects high-impact predictors, thus enhancing model robustness and accuracy, with a mean absolute error of 7.87 kt on the 24-hour independent test set. The contribution analysis module transforms the KAN model into symbolic expressions and quantifies the contributions of dominant predictors. The analysis reveals that initial TC intensity exerts the strongest negative contribution to 24‑h TC intensity change among climatology and persistence predictors; 850‑hPa maximum symmetric tangential wind shows the largest positive dynamic contribution; and the maximum potential intensity dominates the thermodynamic contribution, indicating that a more favorable future thermodynamic environment promotes TC intensification. Therefore, the relationships derived from the symbolic expression largely align with the existing TC 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." Says Prof. Wei Zhong, one of the corresponding authors of the study.
In the future, the team will extend the symbolic expressions to incorporate quadratic, exponential, and interaction terms to disentangle linear and nonlinear contributions and deepen the understanding of the nonlinear dynamic processes governing TC intensity evolution.
