Accurate prediction of tropical cyclone (TC) intensity remains one of the most formidable challenges in meteorology, critical for mitigating the devastating impacts of these powerful storms on communities worldwide. While forecasting the paths of tropical cyclones has witnessed significant improvements over the past few decades, accurately predicting changes in their intensity has lagged behind. This gap poses immense risks, given how sudden intensification or weakening can drastically alter preparedness and response strategies. Addressing this persistent challenge, a team of researchers led by Professor Wei Zhong at the National University of Defense Technology, China, has introduced a revolutionary framework that applies advanced deep learning techniques to elevate the accuracy and reliability of TC intensity forecast models.
This novel approach, termed TCI–KAN, represents a fusion of deep learning with interpretable neural architectures, specifically leveraging Kolmogorov–Arnold networks (KANs) alongside a dynamic predictor pruning optimization module. The architecture of TCI–KAN breaks away from conventional deep learning systems, enhancing both efficiency and interpretability in capturing complex atmospheric dynamics influencing cyclonic intensification. The framework is structured around three primary modules: a predictor pruning optimization module that intelligently selects the most influential input parameters, a neural network optimization module fine-tuning the model’s learning capability, and a prediction module that generates precise intensity forecasts.
The driving innovation behind TCI–KAN lies in its ability to prune a vast pool of potential predictors down to a concise subset that significantly impacts the prediction of tropical cyclone intensity. From an initial collection of 317 predictors—variables ranging from oceanic thermodynamics to atmospheric conditions—this pruning mechanism distills the inputs to just 15 high-impact features. This reduction not only streamlines computational complexity but also enhances model interpretability, a crucial advantage over typical black-box deep learning methods that often struggle to elucidate their decision-making processes.
Testing the TCI–KAN framework on historical cyclone data revealed breakthrough performance, particularly in six-hour intensity forecasts where it achieved a mean absolute error (MAE) of only 2.85 knots. This result marks a significant leap forward, outperforming current operational forecasts by 31 percent and exceeding the accuracy of both single and hybrid deep learning models by 13 and 6 percent, respectively. Such precision improvements are instrumental in providing coastal regions and emergency planners with reliable, timely warnings that can save lives and reduce economic losses.
Beyond accuracy, TCI–KAN demonstrates remarkable versatility and robustness across different ocean basins and tropical cyclone categories. While exhibiting the highest fidelity in the eastern Pacific—a region characterized by a particular set of environmental influences—the model maintains strong predictive capabilities in other areas, adjusting to varying storm intensities. Notably, the framework’s uncertainty in prediction increases moderately with escalating cyclone intensity, reflecting inherent challenges in modeling extreme atmospheric phenomena but still offering superior confidence compared to existing methods.
At the heart of TCI–KAN’s success is the interpretability offered by Kolmogorov–Arnold networks, which differ from traditional deep neural networks by decomposing complex nonlinear mappings into simpler functions. This mathematical foundation allows researchers to better understand and trust the internal workings of the model—a significant stride in applying artificial intelligence in operational meteorology where transparency is essential. The dynamic predictor pruning module enhances this by continuously optimizing the feature set, ensuring that the model adapts to evolving atmospheric conditions and data availability.
Professor Wei Zhong underscores the broader implications of this research, emphasizing that the integration of data-driven techniques with physical mechanisms heralds a new era in meteorological forecasting. “TCI–KAN not only pushes the boundary of forecasting accuracy but also bridges the gap between interpretable machine learning and the traditionally physical mechanism-based methods,” he stated. This fusion can pave the way toward next-generation forecasting systems that balance empirical data insights with robust atmospheric science principles.
The practical implications for disaster management agencies and meteorological services worldwide are profound. Enhanced six-hour intensity forecasts can enable better allocation of resources, refined evacuation planning, and more targeted warnings that reduce unnecessary economic disruptions. Furthermore, the model’s adaptability across regions suggests it could be globally adopted and tailored to local cyclone characteristics, representing a universal tool in the fight against tropical cyclone hazards.
This research also contributes to the ongoing discourse about the role of artificial intelligence in environmental and geophysical sciences. By demonstrating that deep learning models can be both highly accurate and interpretable, TCI–KAN challenges the assumption that sophisticated AI methods must remain opaque. Instead, it illustrates a path forward where explainability complements performance—an essential balance for operational deployment and scientific advancement alike.
The foundation of this work rests heavily on rigorous mathematical optimization, feature selection techniques, and neural network training algorithms that are intricately designed to capture the dynamic and chaotic nature of tropical cyclones. The pruning optimization reduces input redundancy and noise, focusing computational power and model attention on the most relevant physical indicators, such as sea surface temperatures, wind shear parameters, and moisture content profiles—elements known to critically influence storm evolution.
Developed through meticulous experimentation and validation against historical basin-wide datasets, TCI–KAN’s deployment is timely given the increasing threat of intense storms fueled by climate change. As ocean temperatures rise and more variable atmospheric conditions emerge, predictive tools must evolve in tandem to safeguard vulnerable populations and infrastructure more effectively.
Keyun Li, a master’s student and the first author of the publication, played a pivotal role in designing and testing the TCI–KAN framework under Professor Zhong’s guidance. Their collaborative efforts were supported by the National Natural Science Foundation of China, reflecting a national commitment to advancing meteorological sciences through cutting-edge interdisciplinary research spanning physics, computer science, and atmospheric dynamics.
Published in the reputable journal Atmospheric and Oceanic Science Letters, this study sets a new benchmark for tropical cyclone intensity prediction research. It invites further exploration into the fusion of interpretable AI models with traditional forecasting methods, and is expected to influence future developments in the field, including real-time operational use and expanded applications to other extreme weather phenomena.
As the climate evolves and risks from tropical cyclones intensify, innovations like TCI–KAN represent a beacon of progress. They illustrate how the convergence of data science and atmospheric physics can lead to safer, smarter, and more responsive forecasting systems essential for the resilience of societies worldwide.
Subject of Research: Tropical cyclone intensity prediction using interpretable deep learning networks.
Article Title: Tropical cyclone intensity prediction based on Kolmogorov–Arnold networks with predictor pruning optimization
News Publication Date: 13-Aug-2025
Web References:
https://doi.org/10.1016/j.aosl.2025.100694
Image Credits: Keyun Li
Keywords: Tropical cyclones, Deep learning, Meteorology, Cyclone intensity prediction, Kolmogorov–Arnold networks, Predictor pruning optimization, Interpretability, Atmospheric science
