Mix warm humid air, cool dry air and wind shear over the plains, then release the fronts to see which recipes build a thunderstorm and which drop a tornado.
Finishing an event adds 25 rows, one every half hour of the twelve hours, with Trial, Hour, the three ingredients you set, then Temperature, Pressure, Wind Speed and Rain at one fixed spot, the Outcome, and an EF Rating for the trials that produced a tornado. Rows build up across trials so several recipes can be compared; releasing an identical recipe again adds nothing. A line graph of Pressure against Hour, grouped by Trial, shows every storm's dip and sorts the outcomes by how deep it goes. Plotting Wind Shear against Dew Point across trials, coloured by Outcome, finds the corner of the input space where tornadoes live.
Model details
Three ingredients drive everything. Moisture comes from the warm humid air front and is reported as a dew point from 45 to 72 degrees Fahrenheit. Instability comes from the contrast between that warm moist air below and the cool dry air arriving over the top, which you set from 65 down to 35 degrees. Wind shear is the change in wind speed between the ground and the top of the storm, 0 to 45 mph.
Moisture is a gate rather than a contributor. Below about a 50 degree dew point there is not enough water vapour to condense, and the sky stays clear whatever the other two controls say. This is why starting with shear alone gets you nothing at all.
Instability is weighted toward the cool air rather than the moisture, because it is the temperature difference across the layer that decides whether a parcel of air that starts rising keeps rising. Moisture still contributes, since the heat released as water vapour condenses is part of what drives the updraft.
One thing this sim simplifies, and it is worth being straight about because it is the question everybody asks: it is not really that cold air has to sit on top of warm air. Temperature falls with height almost everywhere, almost always, so that arrangement is the ordinary state of the atmosphere and nothing special. What matters is how FAST it falls. A parcel of air that starts rising cools as it goes, and it keeps rising only while it stays warmer than whatever is around it. So the air above has to get colder with height more steeply than the rising parcel itself cools. Forecasters measure exactly that and call it the lapse rate. Dragging the cool air lower in this sim is a way of steepening it.
Two things follow that surprise people. A layer of WARM air sitting on top of the moist air, which sounds like the opposite of what a storm wants, is actually one of the ingredients of the biggest Plains tornadoes. It is called a cap, and by holding everything down it stops a crowd of weak storms from using up the moist air, so when one finally breaks through it rises very fast. The middle layer of the cut-away is that cap. And tornadoes do form in air that is barely unstable at all. Squall-line tornadoes and the ones that come with hurricanes often have very little of the rising-air ingredient and a great deal of wind shear instead. The three ingredients trade off against each other rather than each having a fixed threshold.
Shear only matters once there is a storm for it to act on. It turns a thunderstorm into a rotating supercell, and at high values together with high instability it puts a funnel on the ground. Following NOAA, the mechanism is a horizontal tube of rolling air that the storm's updraft tilts upright into a rotating column called a mesocyclone.
Tornado strength varies with the recipe, and a rating on the Enhanced Fujita scale is exported with each trial. Worth being straight about how that scale really works: EF is a DAMAGE scale. A real rating is assigned after the event by surveying what the tornado did to buildings and trees, and the wind speeds attached to each rating are the speeds that damage is taken to imply. Nobody puts an instrument inside a tornado. This sim shows no damage at all, so it works the scale backwards, from its own wind speed to the rating that speed corresponds to. The six recipes that reach a tornado here spread from EF1 to EF4, so there is a strongest one to find, and the funnel changes shape as well as size: a weak one is a thin rope, and a strong one stands as a wedge nearly as wide at the ground as it is at the cloud. Worth knowing that width and strength are only correlated in real life, not the same thing. Some narrow tornadoes have done violent damage. The correlation is strong enough to be worth showing and too loose to be a rule.
The storm always arrives from the right and leaves to the left, and that is not arbitrary. The camera faces west with south on the right, which is where the warm humid air comes from, and a Plains supercell travels with the wind a few kilometres up, which in a setup like this blows from the southwest toward the northeast. From where you are standing that is right to left, closing on you and then pulling away to the north. The direction is the same every run on purpose, for the same reason the view does not move: every student's screen matches, so a teacher can say “watch the left of the picture” and be right for all of them.
How FAR the storm travels is up to you, though. Storms move with the mean wind, and the wind is what you set on the shear rail, so more shear means a faster storm: it starts further out and ends up further away in the same twelve hours. Its closest approach stays at the same hour either way, which keeps the readings and the picture in step. One real consequence follows: a fast-moving supercell leaves a longer damage track, which is where long-track tornadoes come from.
The funnel is drawn, not solved. There is no fluid dynamics here: the storm's shape is art driven by a small deterministic model of the three ingredients. The same three settings always give the same storm, so the table shows cause and effect rather than noise. The probabilistic side of real forecasting belongs to how you read a table of many trials, not to the engine.
Two things are deliberately not to scale. Twelve hours of an afternoon and evening run in about thirty seconds, and the storm's whole life from first cloud to funnel is compressed to fit inside that. And the two air masses are drawn with roughly three times the contrast they really have. In life a warm humid air mass announces itself as a milky horizon and a dew point reading, not as a coloured bank you can point at, but a control you are told to drag has to be visible where the instruction says it is.
Everything else is close to real scale: a cloud base at about 2 kilometres, a storm about 5 kilometres away, and a funnel a few hundred metres across. That is what the fence posts, the power poles and the grain elevator are there for. You are standing close enough that the storm's anvil and its huge cauliflower tower are above the top of the view, which is what anyone under a supercell actually sees.
The cut-away on the left is a diagram, not a window: its vertical scale is honest to 3 kilometres, but the rolling tube of air drawn in it is far larger than life. A real one is a few hundred metres across and completely invisible, since air is transparent. It is drawn at all because it is the thing the storm is about to stand upright, and you cannot watch that happen to something you cannot see.
The sound is optional and off until you switch it on with the speaker button in the toolbar. Nothing is a recording: the wind, the rain and the tornado's roar are built out of filtered noise whose loudness and tone are driven by the same numbers the weather station is reporting, so what you hear is what the table records. The thunder is the interesting one. Each bang is scheduled for the time sound actually takes to cover the distance to the bolt, about three seconds every kilometre, so early in the event you see the flash a long time before you hear anything and by the time the storm is overhead they arrive together. Counting the gap is how you work out how far away a storm is, and here it falls out of the model rather than being asserted.
The only mark the storm leaves behind is the crop in the foreground, laid flat by the outflow and staying flat. A flattened gap where the funnel crossed the tree line was built and then cut, and the reason is worth knowing: twenty metres of tree five kilometres away is about four pixels on your screen, so flattening it moves three of them and nobody can see it. Damage on the ground is worse still. A track 500 metres wide at that distance takes up roughly a ten-thousandth of the height of the view, because from an eye a metre and a half up everything that far away is edge-on. You cannot see a tornado's track from the ground at all, and that is exactly why the people whose job is to rate tornadoes survey them from aircraft.
No buildings, animals or people are ever in the tornado's path, and nothing is shown being wrecked or thrown.