A single forecast run, sometimes called the "deterministic" run, starts from one best estimate of today's atmosphere and projects it forward once. The problem is that the atmosphere is chaotic: tiny, unavoidable errors in that starting snapshot, far smaller than any instrument could ever resolve, grow over time and can completely change the outcome by day seven or ten. This is the famous "butterfly effect." A single run therefore shows only one possible future among many equally plausible ones. An ensemble forecast solves this by running the same model dozens of times, each time with a slightly different, equally realistic starting point, to sample the range of futures the atmosphere could actually take.
How ensembles are created
Two kinds of perturbations build an ensemble. First, the initial conditions are nudged within the known uncertainty of the observations feeding the model, producing many slightly different but equally valid starting atmospheres. Second, stochastic physics perturbations vary how the model represents processes too small to be calculated directly, such as individual clouds or turbulence, since these are approximated rather than solved exactly. ECMWF's ensemble, called ENS, runs 50 perturbed members plus one unperturbed control run out to 15 days. NOAA's GEFS, the GFS ensemble, runs around 30 members. ECMWF's AI-based AIFS is cheap enough to compute that its own ensemble can be generated quickly on far less hardware, making large ensembles more accessible than ever. Some forecasters go further and build multi-model ensembles, blending members from several different national centres' models into a single, broader pool.
How to interpret an ensemble
The classic visualization is the "spaghetti plot": one line per member, all overlaid for the same variable, like 500 hPa height or 850 hPa temperature. When the lines bunch tightly together, the members agree and confidence is high; when they fan out widely, the atmosphere's future is genuinely uncertain at that point, regardless of how convincing any single line looks. The ensemble mean, the average of all members, is smoother than any individual run and is often a good single number to quote, but it can blur two distinct, equally likely scenarios into one misleading middle value, so it should never be read alone. Percentages are usually the most useful summary: if 35 of 50 members show measurable rain at a location, that is roughly a 70% probability of rain there, not a certainty and not a coin flip. As a rule of thumb, ensemble spread is naturally narrow one or two days out and widens steadily through the week, so treat day-one ensemble output as a sanity check and day-eight or nine output as a guide to the range of possibilities rather than a single answer. On ngmeteo.com's ECMWF, GFS and AIFS ensemble pages, click any location to see exactly this fan of member forecasts for temperature and precipitation, which is the most honest picture of what a forecast model actually knows about the future.