The Thermometer Shop

Every daily max and min ever written into a weather ledger was read off a particular instrument, hanging in a particular shelter, on a particular schedule. Would a different thermometer have written down a different day?

Fill a cart with two to four instrument-and-shelter combinations — their specifications are real and cited, from 1780s liquid-in-glass to reference-network platinum — then run one identical simulated day through all of them and read the receipt.

  • How a mechanical min/max register, a spot-sampling logger, and an averaging logger each turn a day of air into two numbers — and whether they agree.
  • What a shelter does to the air an instrument actually sits in.
  • How each item's meteorological average — the (Tmax+Tmin)/2 midpoint — compares with the 24-hour average of the same simulated air — a number no instrument on the shelf records.

What this is actually computing

The weather in this lab is synthetic. Unlike the other labs on this site, nothing here is a measurement: the day your thermometers experience is generated by a small physical model (sun geometry from your chosen latitude and date, cloud cover you set, seeded random turbulence), built to be plausible — not to reproduce any real place's climate. The named locations set only the sun's path and a generic seasonal baseline for that latitude.

Read more about the simulation

What is real: the instrument and shelter behaviour. Response times, tolerances, reading resolutions, sampling schedules and screen radiation errors are taken from published values and ranges. Where a cited work documents outcomes rather than mechanisms — the wall exposure's coupling constants, the 1780s instrument's exact figures, the friction of a mechanical index — the numbers are this lab's estimates, chosen so the resulting behaviour sits inside the ranges those works document, and labelled as such where they appear. Sources are cited on each product card and at the end of the page.

One thing this lab is not: a ranking. Every instrument in the shop measures the same atmosphere differently — by design, by era, by shelter, by schedule — and running them through one identical day is a way to understand each instrument on its own terms. Whether any of the differences you find are large enough to matter, and for what, stays your call.

Why invent weather at all? Because no real site ever hung an 18th-century thermometer at a poleward wall, a screened Victorian pair, and an aspirated platinum sensor in identical air. A simulation is the only way to hold the day fixed and vary nothing but the instrument. It also means the simulation genuinely knows one number no instrument records — the average of the air itself over all 86,400 one-second steps of the day — and can print it on your receipt at the end.

How the simulated day is generated

A single energy balance drives the air temperature: absorbed sunlight in; thermal radiation out at the layer's own T⁴ against a cloud-dependent sky (the Swinbank 1963 clear-sky downwelling, evaluated at the seasonal baseline and enhanced under cloud — so clear nights can cool toward a radiative equilibrium well below the day's baseline, while overcast nights barely can); and relaxation toward a latitude-based seasonal baseline standing in for the boundary-layer mixing and advection the model doesn't resolve. That relaxation weakens on calm nights — stable air decouples from the layers above, which is what lets a still, clear night dig a deep minimum that a windy night cannot. Sunlight is split into an always-present diffuse part and a direct beam (the Meinel clear-sky airmass formula) that switches off whenever a simulated cloud is over the sun's disc — a random on/off gate whose average matches your cloud slider, passing faster in wind. The diffuse part is sized so the gate-averaged total reproduces the published Kasten & Czeplak (1980) cloud relation; instant by instant, though, a half-cloudy sky flickers the way real broken sky does, and every flicker reaches the air itself: each cloud gap is a small heating spike the thermometers then chase, each covering a dip. (One admitted crudity: to keep the day's total anchored to the published relation, this model's sun-out level runs above the clear-sky value for the whole gap — real records exceed clear-sky only briefly, near cloud edges.) The same gated sunshine drives the wall exposure's direct beam, the screens' radiation errors, and the damping of daytime turbulence while the sun is hidden.

On top of that sits small, fast turbulence (a mean-reverting random wiggle, stronger in daytime convection) — the tenths-of-a-degree texture real air has, the part of the signal that response times and sampling schedules act on. All randomness is seeded: the "weather number" printed on your receipt replays the identical day, so shared links reproduce exactly.

What the shelters do

An instrument never measures "the air" — it measures the air inside its shelter. Each shelter here is a lag (its own thermal response time) plus the radiation and ventilation behaviour documented for its class: the naturally-ventilated louvred screen carries an insolation-scaled daytime excess and nighttime deficit, both largest in calm and shrinking with wind; the fan-aspirated shield is designed to hold both near zero and to be indifferent to the outside wind; and the wall exposure couples the sensor to a building's hours-long thermal memory, restricts its view of the night sky, and receives whatever direct sun the wall's orientation actually admits — computed from the sun's real position at your latitude and date, not assumed. Whether a poleward-facing wall ever sees the sun depends on where and when you shop.

The wind regime you pick feeds all of this: naturally-ventilated lags stretch and radiation errors grow as the air stills, gusts make ventilation flicker, and the whole scaling is clamped to the ranges the cited intercomparisons cover — no regime extrapolates beyond them.

What the instruments do

Each instrument is a thermal lag (how long its sensing mass takes to follow the shelter air), a reading resolution (the finest step it is read to), and a registration scheme: mechanical min/max indexes that ride the liquid and need a small push to move, read and reset once daily; a logger keeping the highest and lowest instantaneous sample; or a logger keeping the highest and lowest block average.

One more thing is modelled because it is on every card: tolerance. A tolerance of ±0.3 °C means any individual unit may read up to 0.3 °C high or low across its scale. At checkout, each unit you buy draws its own fixed scale error at random within its tolerance — seeded separately from the weather, so re-running an order or re-rolling the weather number keeps the same physical units, and a shared link delivers the same units to whoever opens it. You know the tolerance you paid for, not the error you got. Buy two of the same item and you get two different units.

What "the average of the air itself" means here

The average of every one of the day's 86,400 one-second air-temperature samples. Historically, (Tmax+Tmin)/2 and this integrated average have both served as definitions of "the day's temperature" — which one a record uses is a convention, and how much the two differ (and when, and which way) is one of the things this lab leaves to you. No thermometer in the shop measures the integrated average. The simulation knows it only because the simulation generated the air; that is the one honest advantage synthetic weather buys, and the receipt's final line uses it.

What is deliberately left out

Weather systems (the wind regime and cloud fraction you pick hold for the whole day — no fronts, no diurnal wind cycle); observation time of day (every register here is read and reset at local midnight, so all items cover the identical 24 hours — the time-of-observation problem in real records is its own large subject); precipitation, frost, and dew on sensors; instrument aging and drift; screen paint weathering; site moves; and the assumption that your equipment survives whatever you order — a hurricane here changes the ventilation physics, not the mounting bolts. Each of these matters in real records; none is simulated, and nothing on the receipt reflects them.

About the three shelters
Stevenson screen
Louvred wooden box on legs, naturally ventilated — adopted worldwide from the late 1800s to give thermometers a standardised, shaded, ventilated home. Documented behaviour: a sun-scaled daytime excess and a small nighttime deficit relative to the passing air, both largest in calm and shrinking as the wind you choose picks up. (WMO CIMO Guide (WMO-No. 8); van der Meulen & Brandsma (2008))
Mechanically aspirated shield
A fan pulls a constant airstream over the sensor, a design intended to minimise radiation error, hold the exposure lag to about a minute, and make the shelter indifferent to the outside wind. Documented behaviour in field intercomparisons matches that intent. This lab assumes the fan never loses power — in any weather you pick. (Diamond et al. (2013), USCRN; WMO CIMO Guide)
Poleward-facing wall / window recess
How temperatures were taken before screens: the thermometer hangs at the building’s shaded side — a north-facing wall or window in the northern hemisphere, south-facing in the southern. Documented behaviour: the sensor part-couples to the building’s thermal mass, sees less sky at night, and receives whatever direct sun that wall’s orientation actually admits at your latitude and season — computed here from the sun’s real path, not assumed. Wind flushes the recess: its lags and excesses shrink as the wind you choose picks up. (Parker (1994))
Questions you might take shopping (written before we computed any answers)
  • Buy two units of the same item, same shelter. Do their receipts agree? Within what?
  • Run the same cart on a clear day and an overcast one. Which parts of the receipt move — Tmax, Tmin, both, neither?
  • Hold the instrument fixed and vary only the shelter, then hold the shelter fixed and vary only the instrument. Which choice moved your receipt more?
  • Rank your items by how close each met average lands to the true mean. Does the ranking follow instrument age? Shelter? Sampling schedule? Does it hold on a different day?
  • Take the same cart to a different latitude, or the opposite season, or the other hemisphere — does whatever pattern you found travel?
  • Re-roll the weather with everything else fixed. How much of what you concluded survives a different draw of the same sky?
Cart is empty

Your cart

Nothing in the cart yet — add 2 to 4 items from the catalogue.

Data sources & methodology
The synthetic weather (methodology, not a data source)
The day itself is generated, not measured: a single-layer energy balance relaxed toward a latitude-based seasonal baseline, with seeded stochastic cloud gating and turbulence. Its magnitudes are tuned to plausible screen-level diurnal behaviour, and it is not a reconstruction of any real location or date. Radiation pieces use the two published relations below; everything else in the weather generator is this lab's own simplification, stated in "How the simulated day is generated" above. The day's short-term texture (the size of its 5-minute wiggles on clear, broken, and overcast days) was tuned against US Climate Reference Network 5-minute records from 2024 at two near-sea-level stations — one Pacific, one Atlantic — with observed days classified by their measured solar radiation, then checked against the same stations' 2023 records as a held-out test (nothing was adjusted against 2023). In both years the model's flicker sits within the observed day-to-day spread, though real broken-cloud spikes run sharper than the model's at the convective-cumulus site and softer at the marine-cloud one, a cloud-type distinction one cloud slider cannot express.
Cloud transmission — Kasten & Czeplak (1980)
"Solar and terrestrial radiation dependent on the amount and type of cloud," Solar Energy, 24(2), 177–189. doi:10.1016/0038-092X(80)90391-6 — global-irradiance-vs-cloud-fraction relation; here it fixes the gate-averaged total, with the stochastic gate supplying the instant-by-instant broken-cloud flicker around it.
Clear-sky downwelling longwave — Swinbank (1963)
"Long-wave radiation from clear skies," Q. J. Royal Meteorological Society, 89(381), 339–348. doi:10.1002/qj.49708938105 — the clear-sky sky-radiation formula the night-time cooling balance runs against; the cloud enhancement applied to it is this lab's simplification within the standard family of such corrections.
Clear-sky direct beam — Meinel & Meinel (1976)
Applied Solar Energy: An Introduction, Addison-Wesley — airmass attenuation formula for the direct beam (the gated component of the sunlight, and what strikes the wall exposure).
Instrument response times, tolerances, and Tmax/Tmin definitions — WMO CIMO Guide
WMO-No. 8, Guide to Instruments and Methods of Observation, Vol. I, Ch. 2 (Measurement of temperature). library.wmo.int — thermometer time constants, tolerances, screen behaviour, and the 1-minute-mean definition of daily extremes.
Historical exposures — Parker (1994)
"Effects of changing exposure of thermometers at land stations," Int. J. Climatology, 14(1), 1–31. doi:10.1002/joc.3370140102 — documented behaviour of pre-screen wall/window and stand exposures; basis for the wall-exposure model's mechanisms and magnitudes.
Screen vs. aspirated behaviour — van der Meulen & Brandsma (2008)
"Thermometer screen intercomparison in De Bilt (The Netherlands), Part I: Understanding the weather-dependent temperature differences," Int. J. Climatology, 28(3), 371–387. doi:10.1002/joc.1531 — field magnitudes of naturally-ventilated screen radiation errors.
Reference-network practice — Diamond et al. (2013)
"U.S. Climate Reference Network after one decade of operations," Bull. Amer. Meteor. Soc., 94(4), 485–498. doi:10.1175/BAMS-D-12-00170.1 — aspirated PRT design and 5-minute averaging.
PRT tolerance classes — IEC 60751
Industrial platinum resistance thermometers and platinum temperature sensors, IEC 60751 — Class A/AA tolerance values used on the platinum cards.
Product illustrations
Original schematic drawings made for this site — illustrative, not to scale, and not depictions of specific museum instruments.
Early instrument history — Middleton (1966); Austin & McConnell (1980)
W.E.K. Middleton, A History of the Thermometer and Its Use in Meteorology, Johns Hopkins Press; J. Austin & A. McConnell, "James Six F.R.S. — two hundred years of the Six's self-registering thermometer," Notes and Records of the Royal Society, 35 (1980) — Six-pattern design, era manufacture and calibration practice behind the two older liquid-in-glass cards.