| 14 Jan 2026
SPIN (v1.0): A Spontaneous Synthetic Tropical Cyclone Model Empowered by NeuralGCM for Hazard Assessment
Abstract. A hybrid framework for simulating SPontaneous synthetic tropical cyclones (TCs) with realistic INtensity, hereafter SPIN, is developed for TC risk assessment. A key advantage of SPIN over previous synthetic TC models is that it avoids the assumption of independence between TCs, while enabling two-way interactions between synthetic TCs and their ambient environment. The SPIN model leverages a Neural General Circulation Model (NeuralGCM) to simulate spontaneously generated TC tracks, and then couples a dynamic TC intensity model to estimate their intensity evolutions based on the large-scale environment. SPIN reproduces the observed climatology of TC activity, including interannual variability, seasonal cycle, genesis, tracks, and lifetime maximum intensity distributions. It also faithfully reproduces the observed return periods of landfall intensity across different regions, enabling its future application to TC risk assessment. Beyond individual TC events, SPIN demonstrates improved skills in representing multiple tropical cyclone events (MTCEs), including their interannual variability, peak concurrent TC count per cluster, and the spatial relationship between consecutive TCs. By circumventing the independent TC assumption and allowing for two-way TC-environment interactions, SPIN opens new potential for assessing compound hazards like MTCE and beyond.
This manuscript describes a new approach for generating synthetic tropical cyclone (TC) event sets for risk assessment. Specifically, it leverages the newly developed AI/ML model NeuralGCM and the existing statistical-dynamical TC intensity model FAST, named SPontaneous synthetic TC with realistic INtensity (SPIN). The authors show that SPIN has an advantage over conventional statistical-dynamical downscaling models by enabling two-way interactions — TCs are not just forced by their environments but now feed back to them — and claim that the model has improved skill in predicting multiple TC events (MTCs). The idea of combining an AI/ML weather model like NeuralGCM with FAST is novel; thus, I think the manuscript should be published. However, I have some minor questions on some of the details, especially the discussions around MTCs. Below is a list of my comments/questions.
(1) First, the authors argue that SPIN enables two-way interaction between TC and environment, which in my opinion is only partially true. Storm intensity in SPIN is post-processed using FAST, so it does not really 'feedback' to NeuralGCM's environment. As a result, MTCs in SPIN do not really reflect true storm-to-storm interaction. Please add a couple sentences of this limitation.
(2) Could you elaborate on how the simulations were conducted? Are the 14 ensemble members’ simulations initialized 6 hours apart from each other? Can these simulations be considered SST-forced runs, meaning that after the initial time, the only input from ERA5 is the monthly SST and SIC? And there is a 2.5-month spin-up period, am I correct and is this necessary?
(3) I am not really following the argument here — it basically says that NeuralGCM better captures the ENSO modulation of TCs, but both the JL models and other models (Lin et al. 2024; Lee et al. 2025) show that they can simulate ENSO modulation of TCs as well. A clearer demonstration would be a direct comparison of interannual TC frequency or intensity anomalies conditioned on ENSO phase across models.
Lee, C., S. J. Camargo, C. Francis, C. Karamperidou, and C. M. Patricola-DiRosario, 2025: Climate Change Impact on the ENSO–TC Relationship in CMIP6: Synthetic TC Analysis. J. Climate,  38, 5595–5614, https://doi.org/10.1175/JCLI-D-24-0662.1.
Jonathan Lin, Chia-Ying Lee, Suzana Camargo et al. The Response of Tropical Cyclone Hazard to Natural and Forced Warming Patterns, 21 October 2024, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-5248169/v1]
(4) Also, regarding Line 190, one assumption of the random-seeding approach in the JL model is that the genesis process is simply part of intensification. However, your argument seems to suggest that this assumption may not hold for interannual variability. Could you further discuss this, and whether approaches that use genesis indices would be a better way to handle interannual variability?
(5) Figure 8. The area definition is not precise. It is not just the eastern US — you also include the Gulf of Mexico, which includes the southern US.
(6) How did you handle extratropical transition (ET) storms? If you simply run FAST all the way to the mid-latitudes, you are likely to overestimate storm intensity and introduce a positive bias in the number of MTCs.
(7) Line 300. You may need to check with JL23 for details — I think most existing statistical-dynamical downscaling models can provide date information. It may stop at monthly resolution because the input is monthly data, and thus 'daily' information is simply artificially generated due to the seeding rate. However, they do have date information, and MTCs will exist when the monthly environmental conditions are more favorable than in other months. So in a way, is this not similar to your approach of using monthly SST input.
(8) L330. In SPIN, when you apply NeuralGCM output to FAST, do you use instantaneous output or monthly averaged fields? Also, do you have any idea why JL23-ERA5 performs better than JC23-NeuralGCM?
(9) Can you show me the sample errors of your MTCE analysis, like those in Figure 10?
(10) Figure 11: Do STCs forming in these quadrants have a geographic or seasonal preference?