AI-aided simulation of tropical cyclone genesis under climate change

It is essential to quantify the frequency and geometric distribution of tropical cyclone genesis (TCG) for assessing tropical cyclone (TC) activities and associated hazards, especially in the context of climate change. Despite the fruitful achievements in TCG studies typically via empirical-indices based methods and statistical techniques, it is still challenging to estimate TCGs accurately and flexibly. This study presents a deep learning (DL) model, namely TCGNet, which is demonstrated to be able to generate reliable and well-generalized TCG simulations under climate change. The primary characteristic of TCGNet is the integration of channel and spatial attention mechanisms into a traditional conventional neural network framework, which enables the model to automatically capture global and local information, effectively address arbitrary environmental factors at different altitudes, and consequently eliminate the need for manual selection of ambient parameters. Results through comparison demonstrate that TCGNet surpasses traditional methods in terms of prediction accuracy and generalization performance. We then utilize TCGNet to assess the behavior of future TCGs in four carbon emission scenarios, with each scenario corresponding to five specific climate models. Our findings suggest that despite the increase in carbon emission and global mean temperature, the spatial distribution of future TCGs does not exhibit significant shifts. However, there is a noticeable decline in annual TCG numbers, and the regions with high TCG occurrence rates reduce as the carbon emission rises.