How We Learned to Predict an Unpredictable Sky

D-Day — decision under a closing window

On 3 June 1944, the forecast over the English Channel was bad enough to delay the largest amphibious invasion in history. James Stagg could not offer Eisenhower certainty. He could only argue that a narrow improvement on 6 June might be enough to act.

Best opening. It begins with a real decision whose cost of waiting and cost of acting were both visible.

The Royal Charter — a storm no one could see as a system

In 1859, the Royal Charter gale wrecked a steam clipper approaching Liverpool and killed hundreds more around Britain. Pressure and wind had been measured at many ports, but nobody could yet turn scattered observations into a moving storm that another port could act on.

Best historical opening. It starts with the missing shared present that modern forecasting had to invent.

The 1987 Great Storm — a forecast can be wrong in two ways

In October 1987, a violent storm struck southern England. The failure was not simply that forecasters lacked the exact starting atmosphere. A destructive sting-jet process was also unresolved at the model scale.

Best conceptual opening. It immediately separates uncertainty about today from a missing mechanism in the model.

Why was an uncertain forecast enough for D-Day?

On 3 June 1944, wind, cloud and sea conditions over the Channel threatened the planned 5 June invasion. British and American forecasting teams disagreed about what the North Atlantic would do. James Stagg did not tell Eisenhower that 6 June would be good. He argued that a short, marginal improvement might be sufficient, while waiting could close the invasion window altogether. The opening is a decision under asymmetric risk, not a claim that meteorologists knew the future.

Met Office — D-Day archive

If a forecast earns its value by changing a decision before the event, how did anyone first learn to see a storm that had not reached them yet?

Before prediction, how did humans construct one shared sky?

The 1859 Royal Charter gale exposed the earlier problem. Pressure, wind and cloud were measured at many harbours, but each observation remained local. After the wreck, Robert FitzRoy used telegraphed reports and storm charts to trace a system moving across Britain. By 1861, the Meteorological Office could issue port warnings with cones and drums. Forecasting began when scattered measurements became a shared present that another place could act on. This is the first reason countries fund national meteorological services: no ship owner, farmer or town can buy a coherent national observation network for itself. The modern investment logic is measurable. A 2026 WMO assessment in Belize estimated BZD 3.30 to 6.80 in social and economic benefit for each BZD invested in its national meteorological service, but only because forecasts were connected to emergency coordination and response.

Met Office — Royal Charter archive; WMO — Belize early-warning benefit analysis

Once a country has paid to build this common picture of now, can the map be moved forward into tomorrow rather than merely described?

How does a numerical forecast turn today into tomorrow?

Lewis Fry Richardson's 1922 proposal supplied the modern answer. Divide the atmosphere into grid cells. In each, track pressure, temperature, humidity and three-dimensional wind. Advance fluid dynamics, thermodynamics, radiation and moisture in small time steps. Pressure gradients accelerate air, Earth rotates it, water vapour condenses and releases heat, and land and ocean exchange energy. Processes smaller than the grid, such as turbulence and cloud microphysics, have to be approximated rather than directly resolved. Richardson's six-hour hand forecast took more than six weeks; the logic had to wait for computers. The output is then sold or provided as an operational input: a flight dispatcher changes a route, a grid operator schedules reserve, and an emergency manager decides whether preparation is worth its cost.

Met Office — 100 years of numerical forecasting; ECMWF — Modelling and Prediction; Met Office — economic valuation study

If the atmosphere follows equations, why cannot faster computation deliver one exact future that every operator can act on?

Why does an exact equation still fail to give one exact future?

In 1961, Edward Lorenz reran part of a weather calculation using values rounded from a printout. The run initially matched, then diverged into a visibly different pattern. The point was not that the atmosphere ignores physics. It was that tiny unknown differences in the starting state can grow until a later forecast disconnects from the real world. This is why every operational model needs a prior task called data assimilation: it combines incomplete observations from satellites, aircraft, balloons, ships, radar and stations with a short prior forecast to estimate the most likely atmosphere now. That analysis is powerful, but it is never the complete atmosphere. A country that underinvests in observations is therefore not merely buying a less pretty map; it is starting every downstream warning and commercial forecast from a weaker estimate.

MIT — Lorenz forecasting history; ECMWF — Data Assimilation; WMO Commons — observation backbone

If there is no single perfectly known starting atmosphere, how can a forecast show uncertainty in a form that an operator can price and act on?

What does an ensemble actually add?

An ensemble starts many forecasts from slightly different but physically plausible analyses, and can vary parts of the model as well. If the members remain close, confidence rises. If their storm tracks, rainfall totals or wind thresholds spread apart, that spread is useful information. ECMWF made an operational 33-member ensemble in 1992. This changes a real commercial decision. In ECMWF's energy example, deterministic runs suggested no extra LNG cargo was needed; looking across the ensemble's high-demand paths showed that an additional cargo would be needed in most members. The useful product is not the average temperature. It is a range against the cost of buying reserve supply versus the cost of being short.

ECMWF — 25 years of ensemble forecasting; ECMWF — How ensemble forecasts enhance the energy sector

Does a family of forecasts protect us from every way a model can be wrong, or can all the members inherit the same missing piece of reality?

What did the Great Storm of 1987 reveal that chaos alone could not?

The October 1987 storm in southern England distinguishes two failures. Lorenz's problem is uncertainty about the real initial state. In 1987, a second problem mattered: operational models used grid boxes roughly 150 kilometres wide, and a damaging sting-jet process was not represented at that resolution. An ensemble can explore several plausible beginnings. It cannot automatically discover a consequential mechanism that every member fails to resolve. Agreement between forecast members is therefore not the same thing as completeness. And a forecast does not create value simply by being accurate: warnings have to reach people, agencies must coordinate, and someone must still choose whether to close, evacuate, pre-position or do nothing. WMO cites an estimate that universal early-warning access could avoid about US$13 billion of asset losses a year, but that is potential avoided loss, not a promise from a model score. The last-mile decision is part of the forecast system.

Met Office — Great Storm case study; WMO — Triple dividends of early-warning systems

If forecast value depends on this whole chain from observation to action, what does faster AI actually improve, and what part does it leave untouched?

What exactly does AI change, and what does it inherit?

AIFS begins from the same analysed atmosphere as a physics-based system. The conventional IFS repeatedly advances physical equations and approximations over a grid. AIFS encodes that global state with graph neural networks, processes it with a transformer, and decodes a learned later state. Its speed comes from learning this state transition rather than explicitly solving every equation at every step. In 2026, ECMWF upgraded both IFS and AIFS, while AIFS also operates in ensemble form. Faster learned transitions can make more possible futures cheaper, which matters when an aviation service turns high-resolution wind and temperature forecasts into route-specific fuel decisions. They cannot invent absent observations, guarantee rare local processes, or turn a shared model error into independent evidence.

ECMWF — AIFS FAQ; ECMWF — May 2026 IFS/AIFS update; Met Office — AVTECH flight optimisation

So after every technical advance, the same final question remains: what do we do before certainty arrives?

What did humanity actually learn to do with an unpredictable sky?

Return to D-Day. Stagg did not possess a correct future. He had a public observation system, competing interpretations, a narrow range of possible weather, and a decision that could not wait. That is the mature meaning of a forecast. It turns weather from a private surprise into a shared risk that a society can choose to prepare for, price, reroute, postpone, insure against or accept. The emotional payoff is not that uncertainty disappears. It is that uncertainty no longer has to arrive as helplessness.

Met Office — D-Day archive; WMO — Triple dividends of early-warning systems

The future question is therefore larger than whether AI wins a benchmark. Which places will sustain the observations, models, warnings and last-mile institutions that let a probability become protection before the next window closes?

D-Day, June 1944

Stagg had to translate disagreement between British and American weather teams into a recommendation before the invasion window disappeared. The story establishes that forecasts are judged by a decision they change, not by whether they sound certain.

The Royal Charter gale and FitzRoy's storm warnings

The 1859 disaster exposed the difference between local observations and a shared atmospheric present. Telegraph reports, synoptic charts, and port signals made it possible to warn people about a storm they could not yet feel.

Richardson's forecast factory

In 1922, Lewis Fry Richardson described the numerical logic of a modern forecast and attempted a six-hour calculation that took more than six weeks. His imaginary room of 64,000 calculators shows why physics had to wait for computation.

Lorenz's rerun

A small rounding difference in a 1961 model rerun grew into a different weather pattern. Deterministic equations did not produce one usable exact future because the starting atmosphere was never known exactly.

The Great Storm of 1987

The event tests a different limit. Even a well-observed initial state and a family of model runs can share a blind spot when a consequential process is not resolved or parameterised at the relevant scale.

AIFS and faster ensembles

Operational AI forecasts learn a state transition from analysed weather data instead of explicitly stepping through every equation. Their speed can make more plausible futures affordable, but their inputs, verification, and uncertainty problem remain.