Storm Lab

An interactive laboratory for severe weather. Set the season and draw fronts on a map, shape a sounding by hand or load a real one, then watch a cloud model build the storm that environment produces, down to the tornado vortex beneath it. Everything runs on your device.

The four views

Map

The United States as the atmosphere stands now, from the latest hourly model analysis with the Weather Prediction Center's fronts, highs and lows. Or build a severe weather day yourself: the month and hour, Gulf moisture, the elevated mixed layer, the jet stream, and lows, highs, cold fronts, warm fronts, drylines, outflow boundaries and jet streaks you place and reshape. Overlays show CAPE, inhibition, shear, helicity, the significant tornado parameter and the expected storm type. Play it forward and storms form, move and decay where the air allows.

Sounding

A skew-T log-p diagram and hodograph you can drag, or build from controls. Every change is analyzed at once, with the storm type the environment favors and the reasons. Load idealized setups, soundings from notable tornado days, any weather balloon launch in the United States archive, or a GFS and HRRR model sounding for any point and hour since March 2021.

Storm

A three-dimensional cloud model integrates the sounding forward for 150 minutes. Watch simulated reflectivity and velocity, updraft, rotation and cold pool fields, a vertical cross section, or the cloud in three dimensions. Set the speed and scrub back through every simulated minute.

Tornado

An axisymmetric vortex model started from the simulated storm's low-level rotation or from the sounding. Change the parent circulation, updraft, eddy viscosity and cloud base, and follow the vortex from a one-celled core through vortex breakdown to a two-celled tornado, with its winds, EF-scale equivalent, pressure deficit and condensation funnel.

The Guide menu holds a tour of each view, a step-by-step walkthrough that builds a severe weather day on the map, and the units. Every index and control has an information button with a plain definition. On the radar, any spot on the map and every thunderstorm warning, watch and mesoscale discussion shows the sounding for that place and time, with a link that opens it here.

Sounding analysis

The analysis follows the Storm Prediction Center's mesoanalysis and SHARPpy definitions, so a value can be checked against an SPC sounding.

Thermodynamics

• Saturation vapor pressure over liquid water and the lifted condensation level from Bolton (1980); equivalent potential temperature from Bolton's equation 43.
• Parcels rise dry-adiabatically to the LCL and pseudoadiabatically above it, integrated with a second-order Runge-Kutta step of at most 10 hPa. Every buoyancy comparison uses virtual temperature for both parcel and environment (Doswell and Rasmussen 1994).
• CAPE and CIN are integrated in height, with each layer split exactly at the LCL and at every zero crossing of buoyancy, so coarse and fine soundings give the same result. The level of free convection is the base of the positive layer holding the most energy, not the first crossing, which keeps a thin buoyant sliver just above cloud base from filing a capping inversion under CAPE. CIN is the negative area below that LFC.
• Surface-based parcels start from the lowest level. Mixed-layer parcels take the pressure-weighted mean potential temperature and mixing ratio of the lowest 100 hPa. Most unstable parcels start from the level of highest equivalent potential temperature in the lowest 300 hPa.
• Downdraft CAPE starts from the level of lowest equivalent potential temperature in the lowest 400 hPa, at its wet-bulb temperature, and descends pseudoadiabatically to the ground.
• Precipitable water is integrated to 300 hPa. The lifted index uses the surface parcel at 500 hPa.

Kinematics

• Storm motion is the Bunkers et al. (2000) internal dynamics method: the 0-6 km non-pressure-weighted mean wind, deviated 7.5 m/s to either side of the shear between the 0-0.5 km and 5.5-6 km mean winds.
• Storm-relative helicity is the layer sum used by SHARPpy, on winds sampled every 50 m, relative to the right mover.
• The effective inflow layer is the lowest contiguous layer, searched in 100 m steps through the lowest 4 km, whose parcels have at least 100 J/kg of CAPE and no worse than −250 J/kg of CIN (Thompson et al. 2007). Effective bulk wind difference runs from its base halfway up to the most unstable equilibrium level, and at least 500 m.
• The bulk Richardson number divides surface-based CAPE by half the squared difference of the density-weighted 0-6 km and 0-500 m mean winds (Weisman and Klemp 1982). Corfidi vectors use the 1.5-9 km mean wind and the strongest wind below 1.5 km.

Composite parameters

• Effective-layer significant tornado parameter (Thompson et al. 2012): (MLCAPE/1500) × LCL term × (ESRH/150) × shear term × CIN term, with the LCL term 1 below 1000 m and 0 above 2000 m, the shear term 0 below 12.5 m/s and 1.5 above 30 m/s, and the CIN term 1 above −50 and 0 below −200 J/kg. Zero unless the effective layer reaches the ground.
• Fixed-layer version (Thompson et al. 2003) with surface-based CAPE and LCL, 0-1 km helicity and 0-6 km shear.
• Supercell composite: (MUCAPE/1000) × (ESRH/50) × effective shear term (0 below 10 m/s, capped at 1 from 20 m/s) × CIN term (−40/MUCIN below −40 J/kg).
• Significant hail parameter as defined by SPC, with its low-CAPE, lapse-rate and freezing-level reductions.

The outlook

The expected storm type is a rule set whose thresholds are the published ones, and every conclusion prints the numbers behind it. Convection initiates when a parcel (mixed-layer, or most unstable when the effective layer is elevated) has a level of free convection, at least 50 J/kg of CAPE, an equilibrium level more than 3 km above the LFC, and inhibition no stronger than the lift supplies: 5 J/kg with no lift, 35 with heating, 150 moderate, 300 strong. Deep shear under 10 m/s gives pulse storms (multicells with 2000 J/kg and strong lift or downdraft CAPE), 10-18 m/s multicells or lines, and above 18 m/s supercells, unless strong lift runs along a boundary within 40° of the shear, which gives a line (Dial, Racy and Thompson 2010). A line or multicell with 0-2.5 km shear of at least 14 m/s, 2000 J/kg of CAPE and 900 J/kg of downdraft CAPE becomes a bow echo (Weisman and Rotunno 2004). Supercell tornado risk follows the effective STP; line tornado risk follows 0-3 km shear of 15 m/s or more (Schaumann and Przybylinski 2012). Supercell type follows the 9-11 km storm-relative wind and precipitable water (Rasmussen and Straka 1998).

The cloud model

The storm view runs a compressible, nonhydrostatic cloud model of the class used for idealized storm research since Klemp and Wilhelmson (1978), and structured like the research model CM1 (Bryan and Fritsch 2002). It runs on the graphics processor through WebGPU, and on the processor in browsers without WebGPU.

Equations

Prognostic variables are the three velocity components, the Exner function perturbation π′, the potential temperature perturbation θ′, and the mixing ratios of water vapor, cloud water, rain, cloud ice, snow and hail. The base state, written with a subscript 0, is horizontally uniform, taken from the sounding and hydrostatically balanced. In flux form with a divergence correction:

∂u/∂t = −ADV(u) − cp θv0 ∂π′/∂x + D(u)
∂w/∂t = −ADV(w) − cp θv0 ∂π′/∂z + B + D(w)
B = g [θ′/θ0 + 0.608 (qv − qv0) − qc − qr − qi − qs − qh]
∂π′/∂t = −(cs² / (cp ρ0 θv0²)) ∇·(ρ0 θv0 v)
∂θ′/∂t = −ADV(θ′) − w dθ0/dz + (Lv Pv + Ls Ps + Lf Pf) / (cp π) + D(θ′)
∂u1/∂t += −(Cd / Δz) (|V1| u1 − |V0| u0)

where ADV is advection, D the filter and damping terms, c_s the speed of sound in the base state, and P_v, P_s and P_f the net rates of condensation, deposition and freezing, each less its reverse. The last line is the drag on the lowest level, for u and likewise for v, with V₁ the ground-relative wind there and V₀ the environment's; see the lower boundary below.

The pressure equation omits the small advection and diabatic terms of π′, as most split-explicit cloud models do. There is no Coriolis force, no radiation, no surface heat or moisture flux, and no explicit subgrid turbulence scheme. The flow is effectively an implicit large-eddy simulation: grid-scale dissipation comes from the odd-order upwind advection and a sixth-order horizontal filter.

Lower boundary

The ground is rigid, and it slows the wind in the lowest layer through a bulk surface stress ρ C_d |V| V. The drag coefficient comes from the logarithmic profile of a neutral surface layer, C_d = (κ / ln(z₁/z₀))², with z₁ the height of the lowest wind level and a roughness length z₀ of 10 cm for farmland: 0.0026 on the 500 m grid. With no boundary layer scheme to spread the stress upward, it acts on the lowest level alone, and advection and the storm's own circulation carry its effect higher.

Applied as it stands, the stress would also slow the environment itself, and within two hours the model would be simulating a different sounding from the one chosen. So a constant forcing balances the drag on the environment wind, as the large-scale pressure gradient does in a real boundary layer (the balance Dawson et al. 2019 and Davies-Jones 2021 discuss): an undisturbed atmosphere keeps its wind profile for the whole run, and only the storm's departures from it feel the ground. That is what lets friction generate horizontal vorticity near the ground in outflow and inflow, which recent simulations find matters to low-level mesocyclones (Roberts et al. 2016).

Grid and numerics

Microphysics

A single-moment bulk scheme with five condensate categories, each with a prognostic mixing ratio advected by the fifth-order scheme: cloud water and cloud ice, which move with the air, and rain, snow and hail, which fall. Every particle category has an exponential size distribution with a fixed intercept.

The ice processes follow Lin, Farley and Orville (1983) with the rates of Rutledge and Hobbs (1984): conversion of cloud ice to snow and of snow to hail; riming of snow and hail by cloud water; collection of rain by snow and hail, and of ice by rain; freezing of raindrops by contact with ice and by heterogeneous (Bigg) nucleation; wet and dry growth of hail, where a hailstone collecting more liquid than the latent heat it can shed will freeze grows wet and sheds the rest as rain (Musil 1970); sublimation of snow and hail in air below ice saturation; and above 0 °C, melting of cloud ice, snow and hail, including the heat collected cloud water and rain bring. Latent heats of vaporization, sublimation and fusion are applied to every change of phase, so moist enthalpy is conserved; no process removes more of a species than exists.

Rain evaporation is still scaled by 0.2. Warm-rain evaporation is known to overcool outflow compared with ice schemes (Gilmore, Straka and Rasmussen 2004), and melting hail now adds its own cooling below the freezing level. Observed supercell outflow is typically 2 to 6 K colder than its surroundings (Markowski et al. 2002); the strongest simulated storms run colder than that (see the limitations). Snow and hail fall at most 0.8 of a level per large step, a limit the fall speeds here rarely reach.

Initialization and lift

• The sounding is interpolated to the model levels; vapor is the saturation mixing ratio at the dew point, capped at 30 g/kg; the Exner function is integrated hydrostatically upward from the surface pressure using virtual potential temperature.
• The grid moves with the expected storm: the Bunkers right mover for supercells, the Corfidi downwind or upwind vector for lines, the 0-6 km mean wind otherwise. After 20 minutes a tracker locates the main updraft (the strongest 2-5 km updraft helicity above 150 m²/s², or the strongest 5 km updraft) every 5 minutes and changes the frame velocity by at most 3 m/s to keep it near the center. The equations are Galilean invariant, so the frame changes only what stays in the domain.
• Convection starts from a cos² warm bubble 10 km in radius and 1.4 km in vertical radius, centered 1.4 km up, or from a line thermal 6 km in half-width across the domain at the boundary's angle to the shear. The bubble's excess is 1.5, 2.5, 3 or 3.5 K for no lift, heating, moderate and strong lift, with ±0.1 K of deterministic noise below 1 km near the trigger only.
• When the analysis says the lift overcomes the inhibition, the model also nudges vertical velocity toward a target (Naylor and Gilmore 2012) in a column over the trigger, from the ground to 500 m above the level of free convection (at least 2.8 km), on a 60 s time scale: 6 m/s for 15 minutes with heating (not applied when the mixed-layer LFC is below 1.2 km, where the bubble suffices), 8 m/s for 30 minutes with moderate lift, and 10 m/s for the whole run with strong lift. Lines are nudged for 15 minutes at most, since forcing held in place would anchor a line that should advance on its cold pool.
• Sustained strong lift stands for the mesoscale and synoptic ascent that a single, horizontally uniform sounding cannot contain. Observed soundings often carry their instability high in the column over a warm layer near 3 to 5 km, and storms in that air depend on continued ascent to survive. The tornado-day soundings load with the lift that acted that day.

What the storm view shows

• Reflectivity is the equivalent reflectivity of rain, snow and hail for their exponential size distributions (Smith, Myers and Orville 1975; Smith 1984), shown at about 1 km above ground: rain as Z = 3.63×10⁹ (ρ₀ q_r)^1.75 mm⁶ m⁻³; frozen snow and hail scaled by the dielectric factor of ice (0.224 of water's, times their density relative to solid ice squared); and above 0 °C snow and hail scatter as the water coating them, which is what makes a bright band where snow melts and pushes wet hail cores past 65 dBZ.
• Velocity is the ground-relative wind at about 1 km projected toward a virtual radar fixed on the ground 55 km west and 45 km south of the storm's starting point, shown wherever there is echo.
• Updraft is vertical velocity at 5 km. Rotation is vertical vorticity at about 1 km. Updraft helicity is the sum of positive w times vertical vorticity from 2 to 5 km. Cold pool is the potential temperature perturbation at the lowest level, 250 m above ground.
• The cross section runs along the storm's motion through the tracked updraft. The 3D view raymarches cloud and rain water through the model volume; its lighting and texture are illustration, its shape is the model's.
• Tornadoes are inferred, not simulated: a 1 km grid resolves the mesocyclone and not the tornado inside it. The inference asks for the four things tornadogenesis is understood to need, each measured over a 2.5 km circle around the strongest low-level rotation, which is about the smallest circle this grid represents honestly.
  1. A low-level mesocyclone: vertical vorticity at 1 km of at least 0.008 s⁻¹. Detection algorithms use about 0.01 s⁻¹ on finer data (Trapp et al. 2005), and a 1 km grid smooths the peak.
  2. Rotation already reaching the ground, measured as circulation about that circle at the lowest level rather than as a peak value: at least 1.5×10⁴ m²/s, the order observed in tornadic storms at this scale (Marquis et al. 2012). Circulation about a circle survives smoothing; a peak vorticity does not.
  3. Stretching to concentrate it: an updraft of at least 10 m/s somewhere below 3 km over the circle, a mean ∂w/∂z near 5×10⁻³ s⁻¹, which amplifies rotation by e in about three minutes. It is the low-level updraft that matters (Markowski and Richardson 2014; Coffer and Parker 2017), and it is read over a depth rather than at one height because hail can hold the updraft down at a single level inside the precipitation.
  4. Outflow the updraft can still lift: mean potential temperature perturbation over the circle no colder than −4 K. Tornadic rear-flank downdrafts are within a few kelvin of the inflow, nontornadic ones colder (Markowski et al. 2002).
  The environment must support tornadoes as well: an effective STP of at least 0.5, or some STP with a cloud base below 1300 m (Thompson et al. 2003, 2012). That is what keeps a rotating summer storm in still air from being called tornadic. A tornado is rated likely when the rotation, the circulation, the updraft and the STP are all well past their thresholds. The circulation measured here starts the tornado model.
• A simulated storm that produces no tornado signal on a day a tornado really happened is an ordinary outcome, not a failure. One run is one realization of a chaotic flow; tornadoes often fail to form in environments that look identical to ones that produce them; the grid cannot resolve what actually makes the vortex; and the model has no terrain, no boundaries and no neighboring storms, which are often what tipped the real day one way.

The tornado model

An axisymmetric model of the tornado vortex chamber configuration (Ward 1972) and its numerical counterparts (Rotunno 1977; Fiedler and Rotunno 1986; Nolan and Farrell 1999). It solves the dry Boussinesq equations for the streamfunction ψ, azimuthal vorticity η and angular momentum Γ = rv:

u = −(1/r) ∂ψ/∂z,   w = (1/r) ∂ψ/∂r,   ψrr − ψr/r + ψzz = −r η
DΓ/Dt = ν (Γrr − Γr/r + Γzz)
Dη/Dt − uη/r = (1/r³) ∂Γ²/∂z − ∂b/∂r + ν (ηrr + ηr/r − η/r² + ηzz)

• Domain 2 km in radius and 2 km deep on a 101 × 101 point grid (20 m spacing). The outer wall holds the parent circulation; the floor is no-slip, with wall vorticity from Thom's condition; the lid is free-slip.
• The storm's updraft is a buoyancy field between 700 m and 1.9 km over a Gaussian 240 m radius on the axis, scaled so a parcel would reach the chosen updraft speed. Turbulence is a constant eddy viscosity, 3 to 60 m²/s.
• Heun (second-order Runge-Kutta) time stepping with a time step from advective and diffusive limits, third-order upwind advection, and successive over-relaxation for ψ.
• The pressure deficit is cyclostrophic, integrated inward at 150 m. The condensation funnel is drawn where the pressure perturbation at height z is at most −ρg(LCL − z), where surface air would cool to saturation.
• The EF equivalent applies the Enhanced Fujita wind ranges to the strongest horizontal wind in the lowest 50 m, averaged over 30 model seconds because the vortex pulses. Survey ratings come from damage, not wind.
• Started from the sounding, the circulation is 8,000 + 120 × 0-1 km helicity m²/s, the updraft 30 + 0.008 × MLCAPE m/s (at most 75), and the viscosity 6 m²/s. Multiple suction vortices are a three-dimensional instability and do not appear in an axisymmetric model.

The map

The weather now

The map opens on the current analysis. Every hour the Rapid Refresh (RAP), the National Centers for Environmental Prediction's hourly model analysis, is sampled onto a 72 × 45 grid, about 80 km apart: at each point the surface pressure, temperature, dew point and wind, and temperature, dew point, height and wind on 37 pressure levels from 1000 to 100 hPa, with winds turned from the model's grid to true north. Each point's sounding is analyzed by the same code as the sounding editor, and a point you inspect is interpolated between its four neighbors.

• Highs, lows, cold, warm, stationary and occluded fronts and troughs are the Weather Prediction Center's coded surface analysis, issued every three hours.
• Jet streaks are local maxima of the 250 hPa wind of at least 70 kt on the RAP's 32 km grid, traced along the flow while the wind stays within a quarter of its peak.
• Drylines are the sharpest eastward rise of the 2 m dew point over the Plains and the Southwest, at least 5 °C per 100 km, with the moist side above 10 °C, the dry side at least 8 °C drier and no more than 2 °C colder (a colder dry side is a front), followed for at least 200 km.
• Lift: within 90 km of a cold front or 70 km of a dryline in the afternoon, moderate; strong where the air at 700 hPa is also rising. Within 150 km of a warm front, 80 km of a stationary or occluded front or 60 km of a trough, weak, or moderate with rising air aloft. Ascent at 700 hPa of 0.3 Pa/s is moderate lift and 1 Pa/s strong. Where the RAP's composite reflectivity reaches 35 dBZ, storms are under way and the lift is at least moderate. Daytime heating is weak lift everywhere else.
• Outflow boundaries are not in either analysis and do not appear on the live map.

A map you build

A built map is parametric, not simulated. A shallow-water or quasi-geostrophic model could move lows and fronts, but it carries no moisture, no capping inversion and no wind profile, which are what decide the storms. Instead each point of the grid is built the way a forecaster reasons about a surface map:

• Temperature from latitude, season, elevation and a diurnal cycle; dew point from the reach of Gulf and Atlantic moisture.
• A cap of 6 to 10 K where the elevated mixed layer plume lies, which migrates north with the season, weakened by afternoon heating and summer moisture and rebuilt by a nocturnal inversion unless the wind keeps the boundary layer mixed.
• Features change the air around them: cooler, drier air and a stronger cap behind cold fronts; the cool side of warm fronts backing the wind; dry air and a deep mixed layer west of drylines; rain-cooled air behind outflow boundaries; winds circulating around lows and highs, a warm, moist sector southeast of a low, subsidence under a high; and a jet streak adding shear under its core and lift in its left exit and right entrance.
• Lift comes from the nearest boundary, its strength and its angle to the shear. Each point then gets a full sounding analyzed by the same code as the sounding editor.

Storms on the map are a population model after the stochastic seeding idea of Bengtsson et al. (2011): cells whose outlook initiates seed storms at a rate set by storm type and lift, spaced as storms space themselves, carrying the motion and hazards of the air beneath them over lifetimes typical of their type, and merging along boundaries into lines. On the live map the storms start from the analysis at its valid time and move on through it; the environment stays as analyzed. Their radar signatures are drawn procedurally and are not the real radar. To see a storm simulated from the equations, inspect a point and choose Simulate a storm here.

Real soundings

• Radiosonde launches come from the Iowa Environmental Mesonet archive, which holds every United States upper-air site from its first launch. Levels below ground are dropped, and missing temperatures, dew points and winds are filled from neighboring levels in log pressure, with dew points above the last report held at least 25 °C below the temperature.
• Model soundings come from Open-Meteo: GFS blended with HRRR over the United States, on 23 pressure levels from 1000 to 50 hPa plus 2 m and 10 m values, for any hour since March 2021 and up to a week ahead. Each is the model's grid cell nearest the point: about 3 km across where the HRRR covers it, 13 km elsewhere.
• Every real sounding says when its data were last updated: a model sounding names the model runs behind it, when they were published and when the next is due; a balloon sounding gives its launch time and the next routine launch.
• For a warning with a stated storm motion, the sounding is taken 40 km ahead and to the right of the storm, in its inflow, rather than inside the warning's rain-cooled outline. Otherwise it is taken at the product's center at the hour it was issued.
• A sounding is horizontally uniform in the model: fronts, drylines and other gradients near the storm are not represented, so the lift setting stands in for them.

How it compares with real storms

• Sounding parameters for nine observed tornado-day soundings match an independent calculation checked against SPC output.
• Idealized setups reproduce the classic Weisman and Klemp (1982, 1984) regimes: pulse storms in weak shear, mirror-image splitting storms with a straight hodograph, a dominant right mover with a clockwise-curved one, and lines under linear forcing.
• In 150-minute runs with the lift that acted each day, the soundings near El Reno 2013, Plainfield 1990, Moore 2013, Bridge Creek 1999, Greenfield 2024, Enderlin 2025 and Menasha 2026 produce supercells that last the whole run, with hail cores of 71 to 79 dBZ; Jarrell 1997, in weak shear under extreme instability, a slow, long-lived storm. Five of the ten raise the tornado signal at some point: El Reno, Moore, Bridge Creek, Mayfield and Plainfield, as do the idealized tornadic and classic supercells. Greenfield, Enderlin and Menasha do not, the first two because they spread outflow colder than a tornadic rear-flank downdraft and the third because its low-level updraft stays under the threshold, and Jarrell never builds a mesocyclone in its nearly shearless environment. The elevated, inversion-topped Nashville sounding of the Mayfield evening produces no storm at all, as its outlook states, and the summer pulse storms rotate without a signal because their environment cannot support a tornado.
• The live map's soundings match surface observations and weather balloons closely: temperatures and dew points within about 1 °C at airports, and winds aloft within a few degrees of balloon measurements.

Limitations

• Resolution. Features narrower than about six grid lengths (6 km) are smoothed. Mesocyclones are resolved; tornadoes, near-ground vortex sheets and fine outflow structure are not. The effective STP and the rotation criteria stand in for what a 100 m grid would simulate.
• Microphysics. One moment per category with fixed size-distribution intercepts, so particle sizes cannot sort independently of mass: hail cores are reasonable, but size sorting, differential reflectivity and hail size are not simulated. Without sorting, the rain below a storm holds too many small drops and evaporates too readily (Dawson et al. 2010), which is why rain evaporation is scaled. Over the raining part of a supercell the outflow comes out 1 to 7 K colder than the inflow, close to observations; the coldest few cells reach 8 to 16 K, and the storms with the coldest outflow, bow echoes and storms with a high cloud base, are the ones where real outflow is strong too.
• Boundary layer. Drag acts on the lowest level only, with no turbulence scheme to spread it through a boundary layer, and there are no surface heat or moisture fluxes and no radiation: the boundary layer does not evolve through the day, and rain-cooled air is not warmed again by the ground.
• No Coriolis force, no terrain, and a horizontally uniform environment with forced lift in place of fronts and troughs.
• Domain. 128 km with periodic, damped edges: long-lived storms, fast storms that outrun the tracker and large systems feel the edges after about two hours.
• The tornado model is axisymmetric and dry with constant viscosity, forced by a prescribed updraft rather than coupled to the storm.
• The map. On the live map the environment is a real analysis, but on an 80 km grid that blurs fronts and drylines and misses outflow boundaries, and it does not change as storms play forward. A built map comes from rules. On either, the storms are a population model rather than the cloud model, and nothing here is a forecast.

Common questions

Is this a forecast?

No. It simulates environments you choose, build or load. For real weather, see the forecast and warnings from the National Weather Service.

Why did no storm form?

Usually the cap. A rising parcel must push through a layer of warmer air before it can rise freely, and the diagram shades that inhibition in blue. Add lift, lower the cap, add moisture or heat the surface.

Why did a real tornado day give no tornado?

Because one run is one realization. Tornadoes often fail to form in environments that look like the ones that produce them, and the difference is usually something this model does not have: a boundary left by earlier storms, terrain, a neighboring cell, or motions far smaller than a 1 km grid can carry. The page marks a tornado only when the simulated storm has a low-level mesocyclone, rotation already at the ground, an updraft strong enough to stretch it and outflow warm enough to be lifted. When one of those is missing, the honest answer is that this storm, in this run, did not produce one.

Why did a real tornado day give a weak storm?

Check the lift. The storms that day were kept going by a front, a dryline or a trough aloft that no single sounding contains. Set the lift to strong and simulate again.

Is the sounding different a few miles away?

Usually a little. A model sounding comes from the model grid cell nearest the point, about 3 km across where the HRRR covers the United States, so two points 5 miles apart usually get different cells whose profiles differ slightly, most near the ground and near fronts, drylines and storms. A balloon sounding is one launch, the same for every point near its site. On the live map, points are interpolated from cells about 80 km apart.

What device does it need?

Any current browser. With WebGPU the model runs on the graphics processor in a background thread, leaving part of each step idle so the page stays responsive; a slower device automatically uses the 1.5 km grid, about seven times faster. Without WebGPU the processor runs the 1.5 km grid.

References

• Bengtsson, L., M. Steinheimer, P. Bechtold and J.-F. Geleyn, 2011: A stochastic parametrization for deep convection using cellular automata. Q. J. R. Meteorol. Soc.
• Bolton, D., 1980: The computation of equivalent potential temperature. Mon. Wea. Rev.
• Bryan, G. H., and J. M. Fritsch, 2002: A benchmark simulation for moist nonhydrostatic numerical models. Mon. Wea. Rev.
• Bunkers, M. J., and coauthors, 2000: Predicting supercell motion using a new hodograph technique. Wea. Forecasting.
• Corfidi, S. F., 2003: Cold pools and MCS propagation. Wea. Forecasting.
• Dial, G. L., J. P. Racy and R. L. Thompson, 2010: Short-term convective mode evolution along synoptic boundaries. Wea. Forecasting.
• Doswell, C. A., and E. N. Rasmussen, 1994: The effect of neglecting the virtual temperature correction on CAPE calculations. Wea. Forecasting.
• Fiedler, B. H., and R. Rotunno, 1986: A theory for the maximum windspeeds in tornado-like vortices. J. Atmos. Sci.
• Gilmore, M. S., J. M. Straka and E. N. Rasmussen, 2004: Precipitation and evolution sensitivity in simulated deep convective storms. Mon. Wea. Rev.
• Dawson, D. T., B. Roberts and M. Xue, 2019: A method to control the environmental wind profile in idealized simulations of deep convection with surface friction. Mon. Wea. Rev.
• Davies-Jones, R., 2021: Invented forces in supercell models. J. Atmos. Sci.
• Coffer, B. E., and M. D. Parker, 2017: Simulated supercells in nontornadic and tornadic VORTEX2 environments. Mon. Wea. Rev.
• Dawson, D. T., M. Xue, J. A. Milbrandt and M. K. Yau, 2010: Comparison of evaporation and cold pool development between single-moment and multimoment bulk microphysics schemes. Mon. Wea. Rev.
• Kessler, E., 1969: On the distribution and continuity of water substance in atmospheric circulations. Meteor. Monogr.
• Klemp, J. B., and R. B. Wilhelmson, 1978: The simulation of three-dimensional convective storm dynamics. J. Atmos. Sci.
• Lin, Y.-L., R. D. Farley and H. D. Orville, 1983: Bulk parameterization of the snow field in a cloud model. J. Climate Appl. Meteor.
• Marquis, J., Y. Richardson, P. Markowski, D. Dowell and J. Wurman, 2012: Tornado maintenance investigated with high-resolution dual-Doppler and EnKF analysis. Mon. Wea. Rev.
• Markowski, P. M., and Y. P. Richardson, 2014: The influence of environmental low-level shear and cold pools on tornadogenesis. J. Atmos. Sci.
• Markowski, P. M., J. M. Straka and E. N. Rasmussen, 2002: Direct surface thermodynamic observations within the rear-flank downdrafts of nontornadic and tornadic supercells. Mon. Wea. Rev.
• Naylor, J., and M. S. Gilmore, 2012: Convective initiation in an idealized cloud model using an updraft nudging technique. Mon. Wea. Rev.
• Nolan, D. S., and B. F. Farrell, 1999: The structure and dynamics of tornado-like vortices. J. Atmos. Sci.
• Rasmussen, E. N., and J. M. Straka, 1998: Variations in supercell morphology. Mon. Wea. Rev.
• Roberts, B., M. Xue, A. D. Schenkman and D. T. Dawson, 2016: The role of surface drag in tornadogenesis within an idealized supercell simulation. J. Atmos. Sci.
• Rotunno, R., 1977: Numerical simulation of a laboratory vortex. J. Atmos. Sci.
• Schaumann, J. S., and R. W. Przybylinski, 2012: Operational application of 0-3 km bulk shear vectors in assessing QLCS mesovortex and tornado potential. 26th Conf. on Severe Local Storms.
• Musil, D. J., 1970: Computer modeling of hailstone growth in feeder clouds. J. Atmos. Sci.
• Rutledge, S. A., and P. V. Hobbs, 1984: The mesoscale and microscale structure and organization of clouds and precipitation in midlatitude cyclones. Part XII. J. Atmos. Sci.
• Smith, P. L., 1984: Equivalent radar reflectivity factors for snow and ice particles. J. Climate Appl. Meteor.
• Smith, P. L., C. G. Myers and H. D. Orville, 1975: Radar reflectivity factor calculations in numerical cloud models using bulk parameterization of precipitation. J. Appl. Meteor.
• Tao, W.-K., J. Simpson and M. McCumber, 1989: An ice-water saturation adjustment. Mon. Wea. Rev.
• Thompson, R. L., and coauthors, 2003: Close proximity soundings within supercell environments obtained from the Rapid Update Cycle. Wea. Forecasting.
• Thompson, R. L., C. M. Mead and R. Edwards, 2007: Effective storm-relative helicity and bulk shear in supercell thunderstorm environments. Wea. Forecasting.
• Thompson, R. L., and coauthors, 2012: Convective modes for significant severe thunderstorms in the contiguous United States. Part II. Wea. Forecasting.
• Trapp, R. J., G. J. Stumpf and K. L. Manross, 2005: A reassessment of the percentage of tornadic mesocyclones. Wea. Forecasting.
• Ward, N. B., 1972: The exploration of certain features of tornado dynamics using a laboratory model. J. Atmos. Sci.
• Weisman, M. L., and J. B. Klemp, 1982: The dependence of numerically simulated convective storms on vertical wind shear and buoyancy. Mon. Wea. Rev.
• Weisman, M. L., and R. Rotunno, 2004: "A theory for strong long-lived squall lines" revisited. J. Atmos. Sci.
• Wicker, L. J., and W. C. Skamarock, 2002: Time-splitting methods for elastic models using forward time schemes. Mon. Wea. Rev.

How tornadoes form · Tornado profiles · Live radar