How do tornadoes form?

A tornado forms when a thunderstorm’s rotating updraft and rotation that already exists at the ground are organized into a single vortex and stretched until it reaches tornado strength at the surface. That sentence hides where the difficulty is: the National Severe Storms Laboratory still opens its own answer to this question with “the truth is that we don’t fully understand.”

The supercell guide describes the storm this happens inside. What follows is the process, and which of its stages a radar can observe.

The four ingredients

The Storm Prediction Center’s tornado FAQ names four: moisture, instability, lift, and wind shear. “Those are the four needed ingredients.” It qualifies them at once: “how much is enough” varies from situation to situation and “sometimes is unknown.” The first three build a thunderstorm; shear lets it rotate.

Stage 1: the updraft starts rotating

The current synthesis is Fischer, Dahl, Coffer, Houser, Markowski, Parker, Weiss and Schueth (2024), in the Bulletin of the American Meteorological Society: they take the three-stage model of Davies-Jones (2015), split its third step in two, and describe the four stages this page follows. Stage 1 they call “far and away the most well-understood component.” Shear means the air carries horizontal vorticity — spin about a horizontal axis. The updraft tilts that spin upright, and when its axis aligns with the storm-relative inflow, which Davies-Jones (1984) named streamwise vorticity, the updraft rotates as a whole. That is a mesocyclone; the mesocyclone guide covers the mechanism and its velocity signature.

Rotation lowers pressure at its core, and that gradient drives upward acceleration near and below cloud base, where buoyancy alone would lift nothing: the mesocyclone supplies the pull Stage 3 needs. What it cannot do is make rotation at the ground, because tilting requires air to rise away from the surface as it turns. The common image of rotation aloft descending is among the misconceptions Fischer and colleagues wrote to correct.

Stage 2: rotation at the ground is a separate problem

Something else has to put vertical rotation near the surface, beneath the mesocyclone. In idealized simulations the most effective route is the baroclinic downdraft mechanism of Davies-Jones and Brooks (1993): air moving horizontally through a downdraft crosses a buoyancy gradient, which generates horizontal vorticity, and that vorticity is tilted upright as the parcel nears the ground. Given how supercell updrafts and downdrafts are arranged, what accumulates near the updraft is predominantly cyclonic.

Where that buoyancy gradient lives has been argued over for forty years. The classical answer is forward-flank baroclinity, from Klemp and Rotunno (1983) and Rotunno and Klemp (1985). The newer answer is environmental: Coffer, Parker, Peters and Wade (2023, Monthly Weather Review) found low-level mesocyclones drawing almost their entire inflow from the lowest 500 to 1,000 meters of undisturbed air. Fischer and colleagues treat these as complements, and how much the storm-generated part adds is unresolved.

The rear-flank downdraft is the other half, and the reason two storms in one environment differ. Markowski, Straka and Rasmussen (2002) found tornado likelihood rising as rear-flank air became less negatively buoyant, and Fischer and colleagues count that negative buoyancy against tornadogenesis. See the rear-flank downdraft guide.

Stages 3 and 4: organization, stretching, transition

What Stage 2 produces is not yet a vortex: in high-resolution simulations, amorphous elongated patches of rotation — vortex patches — with no closed circulation. Stage 3 is two things happening together: the patches roll up and merge into one symmetric vortex, and that vortex is stretched by the upward acceleration into the low-level mesocyclone.

Stage 4 is friction. In the lowest few meters, drag slows the rotating air while the inward pressure-gradient force does not, so the balance breaks, air converges on the core and erupts upward through the corner flow. The vortex then sustains itself by tilting horizontal vorticity abruptly upward right at the ground — the in-and-up mechanism — and no longer needs the Stage 2 downdraft. Fischer and colleagues call tornadogenesis complete there. A correction follows: a tornado does not touch down. The funnel marks only where the pressure drop makes condensation visible, so it appears to reach down while the vortex spins up from the surface — which is why the funnel cloud guide treats a funnel as evidence, not definition.

Tornadoes without a supercell

Wakimoto and Wilson (1989, Monthly Weather Review) analyzed 27 visual vortices in Colorado during the 1987 CINDE field project. They formed as shear instabilities along radar-detected convergence lines, generally before any precipitation echo existed, and strengthened to tornadic intensity when a rapidly developing storm’s updraft moved over them — stretching, with no mesocyclone anywhere. The authors estimated damage capability as high as F2. Storm chasers call these landspouts; the SPC FAQ calls that “storm-chaser slang for a non-supercell tornado” and notes they “are tornadoes by definition.” How much of the supercell model transfers to QLCS tornadoes is, Fischer and colleagues say, still unclear.

Why it is still an open problem

VORTEX — Verification of the Origin of Rotation in Tornadoes Experiment — ran in the springs of 1994 and 1995 on the Plains, led by Erik Rasmussen of NSSL. VORTEX2 ran in May and June of 2009 and 2010; per the SPC FAQ the 2009 season sampled exactly one tornado, near LaGrange, Wyoming, on 5 June. VORTEX-SE moved the question to the Southeast under the VORTEX-USA program Congress authorized in 2015, its cases running 2016 to 2020 by NSSL’s listing. TORUS then targeted the storm’s internal boundaries and found the streamwise vorticity current of Orf et al. (2017) does exist in nature but is “by no means ubiquitous.” Thirty years of that has bought a far better description of the sequence and very little ability to say, in advance, which storm will produce a tornado.

What radar sees of this, and what it does not

An operational WSR-88D sees Stage 1 well and the rest barely at all.

Stage 1 is visible. A mesocyclone is kilometers across, deep and persistent, which is the scale a radar resolves. It appears in Doppler velocity as inbound beside outbound returns stacked across several elevation cuts — see how to read radar velocity.

Stages 2 and 3 are not. Vortex patches are meters to hundreds of meters across, in the lowest tens of meters of the atmosphere. The simulation Fischer and colleagues illustrate resolves them on an 80 meter grid, and the observations supporting them come from mobile Ka-band radars parked a few kilometers from the storm. An operational beam is about a degree wide and rises with range, so at any useful distance that layer is below the lowest tilt and the structures are far smaller than one sample volume.

Time is the other wall. A severe-weather volume scan takes four to six minutes, and even with three supplemental low-level scans the lowest tilt updates every 75 to 90 seconds — the NEXRAD guide has the numbers. Stages 3 and 4 can finish inside one gap.

What mobile radars did settle matters. Fischer and colleagues composite nine rapid-scan Doppler observations through tornadogenesis: tornadic rotation tends to appear at or very near the surface and build upward, or to intensify at all heights within about a minute. It does not descend.

Why warnings rest on rotation

A warning therefore cannot wait for tornadogenesis to be observed. It rests on rotation the radar can see plus a supportive environment — what the “radar indicated” tag on a National Weather Service warning means, and why warnings go out on storms that produce nothing. As the SPC FAQ puts it: “Many supercells never spawn a tornado!” Tornado watch vs. warning covers the two products, and tornadoes on radar the signatures. Act on National Weather Service warnings.

WeatherStory Radar reads full-resolution Level II velocity and reflectivity, every gate and every elevation the radar actually swept, straight from NOAA’s public archive.