Can bluebottles be forecast 10 days ahead? We tested it, and no

The hypothesis: spot them early, warn the beach

A bluebottle (Physalia physalis) has no engine. The colony floats on a gas-filled "sail" and drifts with currents and wind. The wind pushes it at an angle, and left- and right-handed colonies veer to opposite sides, like boats on different tacks⁶⁹.

That suggests a neat idea for a forecast service: spot a swarm offshore, compute its drift and warn beaches 5–10 days ahead. Similar drift models have been used for Physalia in the Bay of Biscay and the Mediterranean⁷⁸. We tested whether it works off Thailand.

Data and model

• Strandings: 19 Physalia events in 2024–2026, 18 official DMCR warnings¹ plus one documented stranding at Surin Beach. Sixteen are on the Andaman coast in May–September, two in the Gulf of Thailand in February (Songkhla and Pattani), and one on Phuket in March.
• Currents: the HYCOM 1/12° global analysis, surface layer, every 3 hours². Tidal currents (about a quarter of the variability off Phuket) are removed with a 24-hour running mean.
• Wind: the ERA5 reanalysis at 10 m, hourly, on a 1° grid³⁴.
• Model: each "particle" moves with the current plus a fraction of the wind (the leeway, 1–6 % of wind speed) deflected by up to ±45°, which covers both left- and right-handed colonies, plus a random walk. For every stranding, 500 particles per variant were run 10 days back in time.
• Check: the same run in the open-source OpenDrift model⁵ gave nearly the same cloud centres, within 0.2–3 km after 295 km of travel.

For the 2–4 % leeway range we relied on the literature: a theoretical sail model gives about 2.7 % of wind speed⁶, and 2–4.5 % was tested in the Bay of Biscay⁷. The current data has a gap in July–August 2026; no strandings fall in it.

Where the bluebottles came from

First we ran time backwards from each stranding. On the Andaman coast the picture repeats from case to case: the bluebottles arrive almost due west, from open sea between 7 and 9°N.

• 3 days before stranding: about 75 km offshore on average.
• 5 days before: about 110 km, bearing 265° (almost due west) with a spread of only ±13°.
• 10 days before: 150 to 340 km, depending on the leeway assumed (2–4 %).

A key detail: the wind sets the direction, not the currents. With currents alone, particles move along the coast and get only about 70 km away in 10 days. From 2 % leeway upward the source is always to the west; only the distance changes, almost in proportion to the leeway.

The two Gulf of Thailand strandings came from the east-south-east, along the Pattani–Narathiwat coast from the direction of Kelantan and Terengganu in Malaysia. The March event on Phuket cannot be explained by offshore drift at all: in the dry season the wind there blows offshore, the particle cloud cannot leave the coast, so it was a local event.

Why 10 days is too far

The physics finds the source, but precision drops quickly. We released 500 particles from one point and measured how the cloud spreads: by 5–8 km every day.

At day 5 the spread is 25–40 km, at day 10 it is 45–65 km. Individual beaches can be told apart only 1–2 days ahead, a ~50 km coast segment about 3–5 days ahead.

And this is the best case: we used "perfect" past weather (reanalysis), not a forecast. In a real forecast the wind-direction error grows quickly after 3–5 days.

Forward test: does offshore-to-coast transport help?

Next we tested the idea forwards. Every day of the season we released 3,000 particles in a corridor west of Phuket and Phang Nga and counted how many reached beaches 1–10 days later. We compared this index with what the service already uses, a model of onshore wind and waves. The measure is AUC: 0.5 is guessing, 1 is a perfect split between stranding days and the rest.

• Offshore transport alone does worse than the wind model: AUC 0.71–0.73 against 0.81.
• Adding it to the wind model gains nothing: 0.81 → 0.81.
• For "will anything strand somewhere on the Phuket–Phang Nga coast" transport is slightly better (0.72 vs 0.65), but that is no longer a beach-level forecast.
• Even if bluebottles are certainly in the corridor, only 17–35 % of particles reach the coast within 5 days, scattered over 100+ km of shoreline.

The main practical problem is observers. The source is open sea 100–300 km offshore. The nearest places with regular boats, the Similan and Surin national parks, close for the rainy season (the Similans from 16 May to 15 October¹⁰¹¹), exactly the stranding season. iNaturalist holds only 5 Physalia records for the whole Andaman side of the region, almost all on beaches¹².

What worked: long, steady westerly wind

A side result proved more useful than the idea itself. We compared the trajectories before strandings with trajectories on random days of the same season.

Before strandings the 10-day path is about 290 km with a direction spread of ±13°; on ordinary days it is about 195 km and ±46°. Strandings follow 5–10 days of steady westerly transport.

This is a cheap signal: it needs no observers at sea, only last week’s wind. No special source region is involved; what matters is how long and how steadily the transport points at the coast.

What it means for the forecast

• The drift layer in the forecast stays at a 3-day horizon: beyond that the spread exceeds the scale of a beach.
• Drift parameters are refined from the literature and this test: about 3 % of wind speed (2–4 % ensemble) with a deflection covering left- and right-handed colonies.
• A "steady onshore transport over 7–10 days" feature is being tested on historical data and may join the wind part of the model.
• For the Gulf of Thailand, sightings on the Malaysian coast are useful: one record at Perhentian came 9 days before the Pattani stranding. Two events are too few for conclusions, but it is worth watching.

How the whole risk estimate works is described in sources and methodology. What to do if stung is in first aid.

Limitations

• Few events: 19, only 2 of them in the Gulf. Official warning dates are publication dates, not exact stranding times.
• Data resolution is about 9–18 km for currents and 1° for wind, so circulation in bays, around headlands and sea breezes is not resolved.
• Nobody observed the bluebottles offshore, so the leeway cannot be fitted from data; it comes from the literature and the stranding test.
• We used past weather, not forecasts, so real-world skill can only be lower.

Sources

1. DMCR (Department of Marine and Coastal Resources, Thailand) — marine hazard warnings
2. HYCOM — Hybrid Coordinate Ocean Model, global 1/12° analysis
3. Hersbach H. et al. (2020). The ERA5 global reanalysis. Q. J. R. Meteorol. Soc. 146, 1999–2049
4. Open-Meteo Historical Weather API (ERA5)
5. Dagestad K.-F. et al. (2018). OpenDrift v1.0. Geosci. Model Dev. 11, 1405–1420
6. Lee E. A., Schaeffer A., Groeskamp S. (2021). Drifting dynamics of the bluebottle (Physalia physalis). Ocean Science 17, 1341–1351
7. Ferrer L., Pastor A. (2017). The Portuguese man-of-war: Gone with the wind. Regional Studies in Marine Science 14, 53–62
8. Prieto L. et al. (2015). Portuguese Man-of-War (Physalia physalis) in the Mediterranean: a permanent invasion or a casual appearance? Scientific Reports 5, 11545
9. MIO / OSU Pytheas — Bluebottle Watch: wind-tunnel tests of left- and right-handed Physalia models (2024)
10. Similan Islands — Wikipedia
11. Khaosod English (2025). Thailand closes Similan Islands for 5 months to restore nature
12. iNaturalist — Physalia observations