“Sooner than expected” · Science

Florence, sixty years after the flood: the forecast is ready, the defences are not

In 1966 nobody warned Florence. Today's models, AI ones included, would see the flood days ahead. But the flood basins are half built and phone alerts for heavy rain are still on trial.

Line drawing: Ponte Vecchio with its shops and the corridor above the arches, the river risen almost to the top of the arches, a raining cloud and, in the water, a staff gauge with black blocks.
In 1966 the flood destroyed the Nave di Rosano river gauge, which is why that night's flow is still an estimate.

Listen to the episode

Podcast episode in Italian · 10 min. The voice is mine, recreated with artificial intelligence.

In short

  • On 4 November 1966 the Arno reached Florence carrying about 4,000 cubic metres a second, with no warning to the public. The basin authority counts 35 dead, Italy's National Research Council 47 across Tuscany.
  • A reconstruction with today's models says the flood would have been visible 3-4 days ahead. AI models such as the European Centre's AIFS add up to a day on rainfall and cost far less to run.
  • The weak point lies elsewhere: of the four flood basins at Figline only one was finished by 23 October, and IT-alert for heavy rain is still on trial.

The night of 3 to 4 November 1966

The four days before 3 November 1966 were dry. Then it rained without a break for two days across the whole Arno basin, and at Badia Agnano, in the Ambra valley, the rain gauge recorded 437.2 millimetres in 48 hours, a figure never seen again in Tuscany.1

The first flood peak reached the edge of Florence at 5:30 on 4 November, carrying about 4,000 cubic metres a second. The Nave di Rosano river gauge was destroyed, so every figure for that night remains an estimate. According to the basin authority, about 230 million cubic metres of water poured over Florence and the plain.1

The basin authority's most reliable estimate puts the dead at 35, 17 in Florence and 18 in the province, most of them elderly. Among them was a little girl from Sesto Fiorentino.1 The CNR-IRPI, part of Italy's National Research Council, also counts indirect deaths and reaches 47 in Tuscany, and more than 130 across Italy for the storms of those days.2

Nobody warned the public. On the evening of 3 November the mayor, Piero Bargellini, was meeting several councillors at the Hotel Minerva over a crisis in the city government. Forecasts were unreliable, there were no sensors sending river levels remotely, and the last time the Arno had flooded the city was in 1844.1

437.2 mmrain in 48 hours at Badia Agnano, never matched since in TuscanyArno basin authority
4,000 m³/sestimated Arno flow at the edge of Florence at 5:30 on 4 NovemberArno basin authority
230 million m³water that flooded Florence and the plainArno basin authority
35-47deaths, according to the different reconstructionsBasin authority, CNR-IRPI

On 4 November 2026 the Florence State Archive opens the exhibition "Ri-emergere 1966-2026", which also covers the roughly 45,000 documents damaged by the flood.3 The anniversary brings a practical question with it: if the same rain fell tonight, would we see it coming?

What a forecast would see today, with and without AI

The answer already exists. The University of Florence, the LaMMA consortium and the basin authority re-ran the November 1966 forecast with today's tools. The global model would have shown the storm system 3-4 days ahead, the limited-area model would have produced rainfall only slightly below what was measured, and the Region would certainly have issued an orange alert, probably a red one. The hydrological model would have estimated the flood peak well, perhaps with some delay.1

Today reasonably reliable forecasts for the Arno are possible 36-48 hours ahead, helped by sensors installed since the 1990s that send rainfall, levels and flows in real time.1 So the 1966 flood would be seen coming even without artificial intelligence. That does not make AI useless, but it helps show where it is useful.

In November 2023 Google DeepMind presented GraphCast, a model that learns from historical atmospheric data instead of solving the equations of physics. Across 1,380 checks it beat Europe's best deterministic model 90% of the time, and it produces a 10-day forecast in under a minute.4

The biggest step came from the European Centre for Medium-Range Weather Forecasts (ECMWF), which on 25 February 2025 made AIFS, its machine-learning model, operational alongside the physics-based one. According to ECMWF it uses about a thousand times less energy per forecast, a rare case of AI weighing less than the method it works beside.5 The review published in May 2026 says that for most variables AIFS errors are 5-15% lower out to 15 days, and that for rainfall the gain is worth up to a day of extra lead time.6

The limits matter with Florence in mind. AIFS works on a 31-kilometre grid, the physics model on a 9-kilometre one, and ECMWF says the latter is still needed for detail. AIFS tracks tropical cyclones better but underestimates their intensity.6 A study published in PNAS in 2025 showed that an AI model trained without the strongest cyclones cannot forecast them when they come: it learns from examples and struggles with what it has never seen.7 The 437 millimetres at Badia Agnano are that kind of event.

Who forecasts the rain, and how

SystemRun byStatusResult
GraphCastGoogle DeepMindResearch, since 2023Better than the European model on 90% of 1,380 checks
AIFSECMWFOperational since February 2025Errors 5-15% lower, up to one extra day on rainfall, 31 km grid
IFSECMWFOperational physics model9 km grid, still needed for detail
FloodCastItaliaMeteo, CIMA and partnersPilot serviceForecast flows for more than 500 river sections, machine-learning early warnings
Google DeepMind 2023, ECMWF 2025 and 2026, ItaliaMeteo October 2026.

In Italy this work falls to ItaliaMeteo, the national weather agency. Together with other weather services it takes part in Anemoi, ECMWF's framework for testing forecasts made entirely with AI, and with the Bruno Kessler Foundation it is developing AI forecasting chains for heavy rain over the next 12-24 hours.8 In mid-October, with the CIMA Foundation, it presented FloodCast, the pilot service of the European Destination Earth initiative that turns forecast rainfall into river flows at more than 500 sections.9 Carlo Cacciamani, then the agency's director, described the limit that remains: "The atmosphere is a chaotic system: we can reduce uncertainty, but not eliminate it."10

The last stretch, from alert to people

In Italy yellow, orange and red alerts are issued by the Regions.10 In Tuscany the bulletins are prepared by the functional centre and LaMMA for the regional civil protection service, which alerts municipalities and emergency services.1 The system works, but it travels through bulletins, websites, press releases and mayors.

IT-alert, the system that sends a message to every switched-on phone in the cells of an area, has been operational since 13 February 2024 only for nuclear accidents, industrial accidents, the collapse of large dams and volcanic activity at Campi Flegrei, Vesuvius and Vulcano. For heavy rain the trial was extended, and as of 23 October 2026 the official site still lists it that way: only test messages are sent.11

Valencia showed what happens when the forecast is there and the message arrives late. On 29 October 2024 the Spanish weather agency raised the alert to red for the northern interior of the province at 7:36. The Poyo ravine overflowed at Chiva at 11:30 and at Torrent at 18:30. The phone alert arrived at 20:12.12 The judge investigating how the emergency was handled counts 233 dead.13

The defences: what has been done and what is missing

Where the water goes is decided by engineering works, and that is where Florence has built up its longest delay. The Arno plan of the De Marchi Commission, led by Giulio Supino and dated 1974, called for 23 reservoirs holding 240 million cubic metres. In 1999 the basin authority approved a plan for flood basins covering more than 20,000 hectares, then estimated at 1.6 billion euros.14

The largest work built since then is Lake Bilancino, on the Sieve: the project got its decisive push after the flood, work began in 1985 and the lake was only full in January 2001. It holds 69 million cubic metres, supplies drinking water to more than a million people and since October 2017 has also been managed to limit floods.15

In the city, the riverbed sills under Ponte Vecchio and Ponte Santa Trinita, the floor of the river beneath the arches, were lowered in 1980.16 On 30 September 2026 the Region and the city presented the last works on the riverside: about 15 million euros of EU recovery funds and removable barriers on the Lungarno Acciaiuoli. The system is designed to let up to 3,500 cubic metres a second pass under Ponte Vecchio.17 In 1966 around 4,000 arrived.

The rest has to be held back upstream. The Figline system plans four basins for about 25 million cubic metres and 206 million euros. Pizziconi, 3.4 million cubic metres, was completed with its intake works opened in September 2025. Restone, 5.5 million, is due by the end of 2026. Leccio, 8 million, is still being designed.18 Work on Prulli, the largest, started on 30 March 2026 and will take three years. Once complete, the system should lower the flood peak by about 50 centimetres.19

Water in 1966 and room to hold it back (million m³)

Water that flooded Florence and the plain in 1966230
Reservoirs in the 1974 Supino plan240 planned
Lake Bilancino69
Figline basins when complete25
Pizziconi, the first basin finished3.4
Arno basin authority, Publiacqua, Tuscan Region. Striped: the water that overflowed in 1966, for comparison. Protecting the city only requires taking the top off the flood peak. Bilancino serves mainly for drinking water.

The basin authority acknowledges that in the centre the risk has fallen, but on the outskirts it has risen, because a great deal was built near the river after 1966.1 The international scientific committee appointed for the flood's fiftieth anniversary was blunter: it wrote that Florence remains at risk of flooding and that the risk grows every day.16

What I think

My answer to the opening question is yes: today a flood like 1966 would be seen coming days ahead, and AI models would see it a little sooner still, at lower cost. But the most useful finding in this research is not about AI. According to the reconstruction made in Florence with today's tools, physics models were already enough, which means the weak point moved some time ago.

It has moved to the engineering: sixty years on, only one of the four Figline basins was finished as of 23 October. And to the last stretch, the one between the bulletin and someone asleep in a basement flat on the plain between Sesto and Scandicci, where in 1966 the water stayed for days.1 A day of warning is only worth something if someone uses it to close a road or move an elderly neighbour.

The best protections can go unnoticed: when they work, they prevent the worst from happening.

Tenth law of innovation, from Un perché alla volta (translated)

Flood basins are like that: they cost hundreds of millions, they take up fields and, on the day they work, nobody notices. That is why they get postponed. AI models have the opposite problem: they make the news and give the impression that the job is done.

If I had to pick one thing to speed up, I would choose IT-alert for heavy rain, before any new model. The technology exists, and Valencia shows what the hours lost between a red alert and a warning to the public cost. What convinces me about FloodCast is that it goes all the way to the river, the figure a mayor can actually decide on.

What people who live near a river can do

  1. BeforeFind out whether your home can floodThe maps are on ISPRA's IdroGEO platform and in your municipality's civil protection plan, with escape routes and safe areas.
  2. Every dayFollow your Region's alertIn Tuscany it is at regione.toscana.it/allertameteo, with yellow, orange and red levels.
  3. During an alertDo not sleep in basementsAvoid areas that flood and check that your children's school has been told.
  4. When the water comesGo to the upper floorsNever into a cellar or garage to save the car. Avoid lifts, underpasses, embankments and bridges.

This is the advice of the Italian Civil Protection's "Io non rischio" campaign, which also recommends telling your municipality if someone in the family is elderly or disabled and would need help.20

Notes and sources

  1. Arno basin authority, 10 questions on the past, 10 on the present ↗ (in Italian), 2016. Answers on rainfall, flow, volumes, deaths, the missing warning, today's forecast and current risk. ↩
  2. CNR-IRPI, November 1966: the victims of the flood in Tuscany ↗ (in Italian), POLARIS archive. ↩
  3. ANSA, 1966 flood, major exhibition at the Florence State Archive for the 60th anniversary ↗ (in Italian), 10 June 2026. ↩
  4. Google DeepMind, GraphCast: AI model for faster and more accurate global weather forecasting ↗, 14 November 2023. The study was published in Science. ↩
  5. ECMWF, ECMWF's AI forecasts become operational ↗, 25 February 2025. ↩
  6. ECMWF, 2025 forecast evaluation met with positive feedback ↗, 8 May 2026. ↩
  7. Y. Q. Sun, P. Hassanzadeh and others, Can AI weather models predict out-of-distribution gray swan tropical cyclones? ↗, PNAS, 2025. ↩
  8. ItaliaMeteo, Coordination and innovation: ItaliaMeteo's work on Italian weather modelling ↗ (in Italian), January 2026. ↩
  9. ItaliaMeteo, The national agency's work in Destination Earth on flood forecasting in Italy ↗ (in Italian), 22 October 2026. ↩
  10. ItaliaMeteo, When weather turns extreme: can we forecast it? ↗ (in Italian), 29 December 2025. ↩
  11. Italian Civil Protection Department, IT-alert, what it is ↗ (in Italian), accessed 23 October 2026. ↩
  12. S. Pitarch, elDiario.es, Timeline of the alert that came too late ↗ (in Spanish), 30 October 2024. ↩
  13. Valencia Plaza, Valencia DANA death toll rises to 233 ↗ (in Spanish), 28 September 2026. ↩
  14. Arno basin authority, Flood management after 1966 ↗ (in Italian). ↩
  15. Publiacqua, Bilancino, the history of the reservoir that supplies Florence ↗ (in Italian). ↩
  16. G. Seminara, Accademia dei Lincei, The Arno has given warning, Florence must be protected urgently ↗ (in Italian), 19 November 2019. It quotes the 2016 report of the international committee, "Saving a World Treasure". ↩
  17. ANSA, Florence safer 60 years after the flood, riverside barriers unveiled ↗ (in Italian), 30 September 2026. ↩
  18. Tuscan Region, Toscana Notizie, Arno, the Pizziconi flood basin now has its "tap" ↗ (in Italian), 26 September 2025. ↩
  19. Intoscana, Tuscan Region portal, Figline flood basin, work starts at Prulli ↗ (in Italian), 30 March 2026. ↩
  20. Italian Civil Protection Department, Io non rischio, Flood: what to do ↗ (in Italian). ↩