Weather and Climate Modeling

Zusammenfassung

In 1922 an English Quaker named Lewis Fry Richardson published a book proposing that weather could be calculated from the laws of physics. His one trial forecast — computed by hand — took him weeks and was spectacularly wrong, predicting a pressure surge that never happened. He imagined a “forecast factory” of 64,000 human computers working in concert. The machine he needed arrived in 1950, when Jule Charney and John von Neumann ran the first numerical weather forecast on the ENIAC. From that beginning grew two of the largest and longest-running computations humanity performs: the daily weather forecast and the century-scale climate projection. Both turned out to depend on a discovery that haunts the whole enterprise — chaos.

Richardson’s Dream

Lewis Fry Richardson (1881–1953) was the first to state weather forecasting as a problem in numerical mathematics. In Weather Prediction by Numerical Process (1922), he divided the atmosphere into a grid of cells and proposed solving the governing physical equations — the conservation of mass, momentum, and energy — step by step over time. The idea was exactly right. The execution, in an age before computers, was hopeless: his single retrospective six-hour forecast took him six weeks of hand calculation and produced an absurd result (a predicted surface-pressure change of 145 hectopascals, owing to subtle errors in the initial data and numerics).

Undaunted, Richardson imagined a “forecast factory”: a great hall of 64,000 human “computers”, each handling one patch of the globe, coordinated by a conductor with colored lights — a vision of massively parallel computation decades before the hardware existed. His failure was not of concept but of arithmetic speed, and a stability condition discovered later (the CFL condition, 1928) explained why his time steps had been mathematically doomed.

The ENIAC Forecast, 1950

The machine Richardson needed was the stored-program computer. At the Institute for Advanced Study, John von Neumann — who saw weather as an ideal showcase for computing and a target for eventual control — assembled a meteorology group led by the brilliant Jule Charney. Charney’s key insight was to simplify: rather than Richardson’s full equations, he used a “filtered” model that suppressed the fast noise that had wrecked the 1922 attempt.

In 1950, Charney, Ragnar Fjørtoft, and von Neumann (with Arnt Eliassen’s help) ran the first successful numerical weather forecast on ENIAC at Aberdeen Proving Ground. Each 24-hour forecast required about 250,000 operations; the machine’s run time was roughly on par with the weather itself — about 24 hours, including punch-card handling and breakdowns, to forecast 24 hours ahead. It barely kept pace with reality, but it proved Richardson right. Charney sent Richardson the resulting paper, and Richardson, near the end of his life, wrote back congratulating him on the “remarkable progress.”

From Forecasting to Climate

Numerical weather prediction matured fast. By 1955 the U.S. Joint Numerical Weather Prediction Unit was issuing operational computer forecasts, and the European Centre for Medium-Range Weather Forecasts (ECMWF), founded in 1975, became the gold standard for global forecasting.

The deeper leap was from weather (what happens next week) to climate (the statistics of the atmosphere over decades). In 1956 Norman Phillips built the first general circulation model (GCM) — a numerical experiment that reproduced the large-scale circulation of the atmosphere, the jet streams and pressure belts, from physics alone. It is often called the first true climate model and a founding moment of computational science.

The figure who turned GCMs into instruments for understanding global warming was Syukuro Manabe. Working at the U.S. Geophysical Fluid Dynamics Laboratory from the 1960s, he built models coupling atmospheric radiation to the vertical movement of air and, later, the atmosphere to the ocean. As early as 1967 his models quantified how much surface temperature would rise if atmospheric CO₂ doubled. For “the physical modelling of Earth’s climate,” Manabe shared the Nobel Prize in Physics in 2021 with Klaus Hasselmann — the first Nobel awarded for climate science, and a striking recognition of computer simulation as a mode of physics.

The 1979 Charney Report (“Carbon Dioxide and Climate”), led by the same Jule Charney, used these early models to estimate that doubling CO₂ would warm the planet by 1.5–4.5 °C — a range that, remarkably, current far more sophisticated models have only modestly narrowed.

The Discovery of Chaos

Weather modeling also produced one of the twentieth century’s great scientific ideas. In 1961, meteorologist Edward Lorenz was rerunning a simulation on a small computer and, to save time, typed in a midpoint value rounded to three decimals instead of the machine’s six. The forecast diverged wildly from the original. The cause was not a bug but a property of the equations: sensitive dependence on initial conditions. His 1963 paper, Deterministic Nonperiodic Flow, founded modern chaos theory, and the phenomenon acquired its enduring name — the butterfly effect — from a talk title asking whether a butterfly’s wings in Brazil could set off a tornado in Texas.

Chaos sets a hard, permanent limit on weather prediction: no matter how good the model or the data, deterministic forecasts lose all skill after about two weeks. The field’s response was to stop pretending to certainty. Modern forecasting is ensemble forecasting: run the model dozens of times from slightly perturbed starting points and report the spread — the basis of every “70% chance of rain.” Crucially, chaos limits weather but not climate: you cannot say whether it will rain on a given day next year, but you can predict the statistics of next year’s climate, just as you can know a casino’s odds without predicting any single spin.

The Computational Scale

Weather and climate codes are among the most demanding workloads on Earth, and have driven supercomputing for seventy years. Operational forecast centers and climate institutes are perennial occupants of the TOP500. The core challenge is resolution versus time: halving the grid spacing multiplies the cost roughly sixteen-fold (three space dimensions plus a shorter required time step), so every gain in detail is paid for in exaflops. Today the frontier is shifting again, as machine-learning models trained on decades of reanalysis data (such as ECMWF’s and others’ neural forecasters) match or beat traditional physics-based models at a tiny fraction of the running cost — the latest turn in a problem Richardson posed a century ago.

⚠️ Dead End: Weather Control

Von Neumann pursued numerical weather prediction partly because he believed understanding the atmosphere was the first step to controlling it — nudging a chaotic system with small, well-placed interventions to steer storms or engineer climate, a prospect he saw as a strategic weapon. The U.S. ran serious cloud-seeding programs (Project Cirrus, Project Stormfury) into the 1980s, and the military’s Operation Popeye seeded clouds over Vietnam. None achieved reliable, verifiable control. The same chaos that limits prediction makes deliberate steering nearly impossible to confirm: when you cannot run the counterfactual, you cannot prove your intervention did anything. Weather control quietly faded as a science even as climate change — humanity’s vast, unintentional experiment on the atmosphere — became the defining application of the very models built to forecast it.