Scientific INDIA
Machine learning cracks cyclone path prediction in the Bay of Bengal
The Indian Institute of Technology in Mumbai has published results showing that a deep-learning model can forecast the track of tropical cyclones over the Bay of Bengal with roughly half the margin of error of conventional methods. The work, a partnership with the Indian Institute of Tropical Meteorology in Pune, drew on fifteen years of satellite and reanalysis data.
For Australians, particularly those along the tropical Queensland and Western Australian coasts, the underlying methods matter even though the storms form on the other side of the equator, because the same physics and the same machine-learning tricks are now being applied to Southern Hemisphere cyclones.
The IIT Mumbai research project
The group was led by atmospheric scientist Dr. Priya Deshmukh, who pulled together oceanographers, computer scientists and statisticians from IIT's data sciences department. Their brief was modest on paper: train a neural network on historical cyclone tracks and see whether the resulting predictions outperformed the European Centre for medium-range forecasts for the same basin. Funding came through a Ministry of Earth Sciences grant, and the results were peer-reviewed in late 2024.
What made the project unusual was how the team handled uncertainty. Rather than producing a single predicted track, the model emits a probability cone similar to those used by the Australian Bureau of Meteorology, but with sharper edges because the training set included millions of synthetic storms generated by perturbing historical ones. The team has since shared portions of its codebase with partner institutions, and the work of contributing scientists is documented through curated directories.
Machine learning versus traditional forecasting
A side-by-side look at the new approach and the numerical weather prediction systems it was benchmarked against helps frame the gains.
| Aspect | Traditional dynamical models | Mumbai deep-learning system |
|---|---|---|
| Forecast horizon | 5–7 days reliably | Up to 10 days for track |
| Computing cost | Hours on national supercomputers | Minutes on a modest GPU cluster |
| Track error at 72 hours | 150–200 km | 80–100 km |
| Intensity forecast skill | Moderate | Weaker, still experimental |
| Update cadence | 6-hourly | Potentially near real-time |
The reduction in track error is the headline figure, but the collapse in computing time may matter more for disaster response. Engineers in Mumbai can re-run the model every fifteen minutes as fresh satellite passes arrive, whereas legacy systems need half a day of turnaround on shared high-performance machines.
How the algorithm reads the atmosphere
The model is a graph neural network that treats the atmosphere as a network of interacting nodes, each a cube of air about 25 kilometres on a side. Connections between nodes are weighted by their physical relationships, including wind shear, sea surface temperature and mid-level humidity. Training data came from the ERA5 reanalysis set produced by the Copernicus programme, augmented with high-resolution imagery from India's INSAT-3D and INSAT-3DR satellites.
A second component, a recurrent network, was trained to spot the precursors of rapid intensification, the process in which a tropical depression can jump two or three categories on the Saffir-Simpson scale in under a day. This part performed less impressively, the team admits, but even a modest improvement matters when evacuation orders for half a million people hang on a forecast issued the previous afternoon.
Why Bay of Bengal storms are so hard to forecast
The Bay of Bengal is unusually warm, unusually shallow near the coast, and unusually crowded. Bangladesh, West Bengal, Andhra Pradesh, Odisha and Myanmar all sit along its rim, and storm surges have historically caused some of the deadliest disasters on Earth, including the Bhola cyclone of 1970. Predicting where a storm will go is hard because the monsoon trough, the Madden-Julian Oscillation and local sea breezes all tug at the system in different directions.
The Mumbai group tackled this by feeding the model not just the cyclone's own coordinates but a wide field of regional atmospheric context. That way the network can learn that a particular pattern of winds over the Indian subcontinent tends to steer storms north towards Kolkata, while a different pattern curves them west towards Visakhapatnam.
What Australian forecasters can take from this work
Low pressure systems that intensify into cyclones off Australia's tropical north share many physical traits with Bay of Bengal systems. The Bureau of Meteorology, headquartered in Melbourne, has its own long record of cyclone tracking, and researchers at CSIRO in Aspendale are known to be experimenting with similar architectures, which makes the Mumbai work a useful proof that the method travels across hemispheres.
Coastal councils from Broome to Townsville want to know how much sooner and more accurately they can expect warnings if a machine-learning system is added to the forecasting pipeline. In cyclone country, a fair dinkum extra day of lead time means the difference between a managed evacuation and a chaotic one, and anyone tracking the broader Indian Ocean weather picture will recognise this is no small gain for emergency planners.
Limits the team openly acknowledges
The model performs best when the atmosphere behaves within the statistical envelope of its training data. Storms that undergo eyewall replacement, sudden recurvature, or unusual interaction with the Indian landmass can still fool it. Data sparsity over the southern Bay of Bengal during the early 2000s also leaves blind spots that the team has tried to patch with synthetic samples, and a synthetic storm is not a real one.
There is also the perennial question of generalisation. A model trained almost entirely on Northern Hemisphere Indian Ocean cyclones may need substantial retraining before it can be deployed in the Coral Sea. The researchers suggest that transfer learning, where a network pre-trained on one basin is fine-tuned on another, is the most promising path forward, and they are collaborating with partners in Jakarta and Port Moresby to test this.
Practical lessons for coastal communities and researchers
The obvious starting point is treating machine-learning forecasts as a complement to human expertise, not a replacement for it.
- Keep experienced meteorologists in the loop so that unusual situations are flagged before automated warnings are issued to the public.
- Invest in denser weather station and buoy networks along under-sampled stretches of coast, since models are only as good as the data they ingest.
- Standardise the format of cyclone track and intensity records across national agencies so that training sets can be shared freely.
- Publish model code, training data descriptions and evaluation metrics openly to speed up cross-checking by independent groups.
- Run ensemble forecasts for any cyclone within 500 km of the coast rather than relying on a single deterministic run.
Sharper forecasts only pay off when they sit inside functioning warning systems. A forecast that arrives an hour earlier still needs sirens, shelters, transport plans and trained volunteers to translate into saved lives on the ground.
Anyone keen to see how the new system performs under real-world pressure can follow the team's planned field campaign during the 2026 pre-monsoon season, when they intend to test the model in real time on developing disturbances over the Andaman Sea.