Too Rare to Learn: Prescribed Cyclone Tracks Degrade a Bay of Bengal Ocean Emulator

Sumaiya Islam

cs.LGSep 4, 2026

Too Rare to Learn: Prescribed Cyclone Tracks Degrade a Bay of Bengal Ocean Emulator

Authors: Sumaiya Islam

Organizations: Department of Software Engineering University of Dhaka Dhaka, Bangladesh

Abstract

Neural ocean emulators are being proposed for regional forecasting in cyclone-exposed coastal seas, and a natural design choice is to hand the network the cyclone as a prescribed input. We test that choice in the Bay of Bengal and find it harmful. We withhold 15 whole cyclones spanning 65 to 150 kt from GLORYS12 reanalysis and compare two U-Nets that are identical except for four prescribed cyclone-track channels. Across three seeds the ocean-only model beats persistence in every run and the storm-conditioned model loses to it in every run, with the two skill ranges disjoint (p = 3.1e-5, paired across storms). The cause is exposure frequency rather than signal content: the channels are non-zero on only 7.9% of training days, so they are out of distribution the moment they activate. The extra error falls inside the prescribed storm footprint, and replacing the real cyclone map with a no-storm map at inference improves held-out storm forecasts by 7.5 to 16.4% in every seed. The conditioned network has learned a response to a rare signal that is confidently wrong.

Figures & tables

Figure 1: The four prescribed cyclone channels for one day of Amphan, 18 May 2020: smooth track proxies carrying no ocean state. Amphan moved almost due north, hence the faint eastward panel.

Arm	Checkpoint	Seed 1	Seed 2	Seed 3	Storms beaten, by seed (of 15)
Ocean-only	best	$+0.125$	$+0.153$	$+0.161$	12 / 13 / 14
Ocean-only	final	$+0.048$	$+0.153$	$+0.161$	11 / 13 / 14
Storm-conditioned	best	$-0.206$	$-0.194$	$-0.293$	6 / 6 / 4
Storm-conditioned	final	$-0.075$	$-0.078$	$-0.477$	6 / 6 / 0

Table 1: Skill against persistence on 272 windows from 15 held-out cyclones, higher is better.

Figure 2: Per-cyclone skill against persistence, one point per seed, bar at the seed mean, ordered by persistence difficulty.

Figure 3: Left: skill lost to conditioning, by regime, one point per seed. Right: RMSE on held-out cyclone days given the real cyclone map and given a no-storm map. The channels are active on 7.9% of training days.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Name	Year	Month	Season	Max wind (kt)
Fani	2019	Apr	pre	150
Amphan	2020	May	pre	145
Mocha	2023	May	pre	145
Unnamed	1994	Apr	pre	125
Unnamed	1996	Nov	post	115
Unnamed	1997	May	pre	115

Appendix

Table 2: The 15 held-out cyclones, ordered by intensity. Pre-monsoon is March to June, post-monsoon September to December.

Figure 4: The event-disjoint split, drawn from IBTrACS tracks: 227 training cyclones, Phailin held for validation, and the 15 test cyclones. Whole events are withheld with 7-day buffers, so no day from a test storm, or from the week either side of it, appears in training.

	SST RMSE ( ∘ C) $\downarrow$			SSS RMSE (PSU) $\downarrow$
Model	Seed 1	Seed 2	Seed 3	Seed 1	Seed 2	Seed 3
Persistence	0.141			0.226
Ocean-only	0.174	0.163	0.166	0.208	0.183	0.180
Storm-conditioned	0.248	0.244	0.463	0.222	0.225	0.261

Appendix

Table 3: Pooled RMSE on held-out cyclones in physical units, final-epoch checkpoints, per seed. Lower is better; bold beats persistence.

Figure 5: Where the skill comes from. Bars are seed means, points individual seeds. Only ocean-only salinity falls below the persistence line.

Regime	Seed 1	Seed 2	Seed 3
Calm days	$-0.026$	$+0.105$	$+0.599$
Cyclone days	$+0.123$	$+0.232$	$+0.638$

Appendix

Table 4: Ocean-only skill minus storm-conditioned skill, by regime. Positive means conditioning hurts.

Figure 6: Occupancy of the cyclone channels over the 11,195 training days. Left: the fraction of days carrying any signal, by year, against the 7.9% overall rate. Right: training days by the peak wind on that day, on a log scale. The network sees no storm on 10,310 days and a storm above 120 kt on five.

	Inside footprint			Outside footprint
	Seed 1	Seed 2	Seed 3	Seed 1	Seed 2	Seed 3
Ocean-only	0.291	0.276	0.272	0.216	0.191	0.187
Storm-conditioned	0.497	0.545	0.548	0.230	0.222	0.277
Penalty	$+70.6\%$	$+97.8\%$	$+101.5\%$	$+6.2\%$	$+15.8\%$	$+48.5\%$

Appendix

Table 5: Salinity RMSE (PSU) on the 112 held-out cyclone days, inside and outside the prescribed storm footprint, final-epoch checkpoints. Penalty is the storm-conditioned arm relative to the ocean-only arm in the same region.

Figure 7: One day of Cyclone Fani, 28 April 2019, seed 2. Left to right: the observed one-day salinity change (PSU), the ocean-only error, the storm-conditioned error on the same colour scale, and the cyclone channel the conditioned arm was given. The conditioned arm carries a broad error across the southern basin, where the channel is active. Fani is the held-out cyclone whose skill gap is closest to the median among those showing the reported sign pattern, and the day drawn is the one with the strongest prescribed signal; neither was chosen by inspecting the errors.

Baseline	RMSE $\downarrow$	Skill vs persistence
Climatology	0.0929	$-2.842$
Persistence	0.0242	$0.000$
Damped persistence ( $\alpha=0.97$ )	0.0239	$+0.010$

Appendix

Table 6: Classical baselines on the same 272 held-out cyclone windows, standardised state, lower RMSE is better.

Neural Surrogate Modeling

Neural surrogate modeling uses learned models to approximate costly simulations, enabling faster prediction, design, control, and inference.

See all 418 papers

HClimRep-Ocean: A Global Ocean Emulator on an Unstructured Mesh
Sep 23, 2026Kacper Nowak, Aleksei Koldunov, Nikolay Koldunov +7Neural Surrogate Modeling Climate Modeling
Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, Germany · Forschungszentrum Jülich GmbH, Jülich Supercomputing Center, Jülich, Germany · Karlsruhe Institute of Technology, Karlsruhe, Germany +3
Samudra 2: Scaling Ocean Emulators across Resolutions
May 24, 2026Yuan Yuan, Jesse Rusak, Alexander Merose +5Neural Surrogate Modeling Climate Science
Courant Institute School of Mathematics, Computing, and Data Science, New York University, New York, NY, USA · Open Athena AI Foundation, Inc., New York, NY, USA · Program in Atmospheric and Oceanic Sciences, Princeton University, Princeton, NJ, USA
Neptune: An AI model for Global Ocean Subseasonal Prediction
Sep 8, 2026Davide Donno, Italo Epicoco, Massimo Cafaro +8Neural Surrogate Modeling Climate Modeling
Department of Engineering for Innovation, University of Salento, Via per Monteroni, Lecce, Italy · CMCC Foundation - Euro-Mediterranean Center on Climate Change, Italy · Department of Earth and Environmental Engineering, Columbia University, New York, NY, USA +2

Sep 23, 2026J/K move · Enter open · S save

HClimRep-Ocean: A Global Ocean Emulator on an Unstructured Mesh

Kacper Nowak, Aleksei Koldunov, Nikolay Koldunov, Savvas Melidonis+6

Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, Germany · Forschungszentrum Jülich GmbH, Jülich Supercomputing Center, Jülich, Germany · Karlsruhe Institute of Technology, Karlsruhe, Germany+3

Neural Surrogate Modeling Climate Modeling Spatiotemporal Forecasting Climate Emulation

Open paper AskPDF arXiv

Machine-learning (ML) emulators for atmospheric processes have advanced rapidly in recent years, transforming weather forecasting. Although early ML ocean forecasting models now exist, they remain less developed than their atmospheric counterparts. Unlike the atmosphere, much of the ocean's kinetic energy resides in mesoscale eddies whose characteristic spatial scales are approximately an order of magnitude smaller than those of comparable atmospheric features. Moreover, complex coastlines, narrow straits, and ice-covered seas make boundary representation a central challenge that atmospheric models do not face. Consequently, numerical ocean simulations commonly use locally refined or even completely unstructured meshes. However, their data-driven counterparts have so far been built around latitude-longitude grids. We present HClimRep-Ocean, an ocean emulator that operates directly on the native unstructured mesh of FESOM2. The emulator is trained on a 209-year AWI-CM3 control integration and is run without atmospheric forcing, receiving the atmospheric state only at initialisation time, which isolates the predictability carried by the ocean state itself. Skill is strongly field-dependent: for currents, HClimRep-Ocean outperforms every reference at 30 day forecast, whereas for temperature and salinity a damped-anomaly persistence forecast remains the more accurate estimator. This behaviour is physically interpretable: current variability is largely geostrophic and internally generated, whereas sea-surface temperature and salinity fluctuations are driven by atmospheric forcing through weather state. Evaluated independently on the OceanBench benchmark, a reanalysis-trained variant of HClimRep-Ocean achieves the lowest RMSE against GLORYS reanalysis among all assessed systems, confirming the competitiveness of the native-mesh approach.

Abstract

Neural Surrogate Modeling

Explore similar work

HClimRep-Ocean: A Global Ocean Emulator on an Unstructured Mesh

Samudra 2: Scaling Ocean Emulators across Resolutions

Neptune: An AI model for Global Ocean Subseasonal Prediction