TC‐GEN: Data‐Driven Tropical Cyclone Downscaling Using Machine Learning‐Based High‐Resolution Weather Model.

Synthetic downscaling of tropical cyclones (TCs) is critically important to estimate the long‐term hazard of rare high‐impact storm events. Existing downscaling approaches rely on statistical or statistical‐deterministic models that are capable of generating large samples of synthetic storms with characteristics similar to observed storms. However, these models do not capture the complex two‐way interactions between a storm and its environment. In addition, these approaches either necessitate a separate TC size model to simulate storm size or involve post‐processing to capture the asymmetries in the simulated surface wind. In this study, we present an innovative data‐driven approach for TC synthetic downscaling. Using a machine learning‐based high‐resolution global weather model (ML‐GWM), our approach can simulate the full life cycle of a storm with asymmetric surface wind that accounts for the two‐way interactions between the storm and its environment. This approach consists of multiple components: a data‐driven model for generating synthetic TC seeds, a blending method that seamlessly integrates storm seeds into the surrounding while maintaining the seed structure, and a model based on a recurrent neural network to correct for biases in storm intensity. Compared to observations and synthetic storms simulated using existing statistical‐deterministic and statistical downscaling approaches, our method shows the ability to effectively capture many aspects of TC statistics, including track density, landfall frequency, landfall intensity, and outermost wind extent. Leveraging the computational efficiency of ML‐GWM, our approach shows substantial potential for TC regional hazard and risk assessment. Plain Language Summary: Tropical cyclones (TCs) cause significant destruction each year. It is crucial to accurately assess the risks they present, but this is challenging due to a scarcity of historical data. A commonly used approach involves creating a large number of synthetic TCs that share key characteristics with real storms. However, traditional synthetic TC generation approaches do not capture the complex interactions between storms and their larger‐scale environment. Furthermore, these approaches do not adequately represent the asymmetric structure of TCs. Recently, advances in machine learning‐based global weather model (ML‐GWM) have provided highly accurate and efficient high‐resolution global weather forecasts that surpass conventional numerical weather forecasting. Here, we introduce a novel synthetic TC generation approach, which we call the synthetic TC‐GENerative Model (or "TC‐GEN"), leveraging the state‐of‐the‐art ML‐GWM. We show that TC‐GEN can generate a large number of synthetic storms that allow the interaction between the storm and its environment. We evaluate the performance of TC‐GEN in various aspects, including several landfall characteristics, which are of the most importance for local TC risk analysis. Our study also serves as a compelling example of the transformative impact of machine learning and data science in revolutionizing climate studies during the era of artificial intelligence. Key Points: We introduce a novel approach to tropical cyclone (TC) downscaling using a machine learning based global weather model, named "TC‐GEN"We generate and integrate synthetic TC seeds into the surrounding environment using a data‐driven approachTC‐GEN establishes a framework that opens the possibility of modeling the two‐way interactions between storms and the environment [ABSTRACT FROM AUTHOR]