What color is fire, measured: the blue flame is not a temperature reading

Ask what colour fire is and you get two answers, both of them confident. The first is orange, obviously, look at a candle. The second is a correction, usually delivered by whoever remembers more chemistry: blue is the hot one, and if your gas hob burns yellow something is wrong with it.

Both answers are pointing at something real and both are describing the wrong quantity. We built a flame colour pipeline to settle it: Planck’s law and real emission line data on one end, the CIE 1931 observer and sRGB on the other. Two results came out that we had not expected. Your screen has never shown you a candle flame, and it never can. And the blue in a gas flame is more saturated than an infinitely hot glowing object, which rules out temperature as its explanation entirely.

Fire that glows because it is hot

A wood fire, a candle, an oil lamp and the yellow tip of a badly adjusted burner all shine for the same reason. Combustion is making soot, the soot particles get hot, and hot solids radiate. That is thermal radiation, which means Planck’s law applies and the colour depends only on temperature. We mapped this curve in detail when we built the colour temperature chart, so here we can jump straight to flames.

Every row below is Planck’s law integrated against the CIE observer at 1 nm and pushed through the sRGB matrix, brightest channel normalised to 255. The last column is the part nobody mentions.

The blue channel of sRGB goes negative below 1,900 K. Not dim, not dark. Negative, meaning the colour sits outside the triangle your monitor can reach and there is no pixel value that corresponds to it. Every flame temperature a human being encounters in ordinary life is under that line:

• A candle flame peaks at about 1,400 °C, which is 1,673 K. NIST measured candles carefully enough to publish a properties table and that is the figure in it.
• The bright yellow body of the flame, the part you actually look at, is cooler still, around 1,470 K.
• A wood fire glowing dull red is nearer 1,100 K, and a dying ember 900 K or less.

So the screen error column is not a rounding note. At 1,470 K the best the display can do with straight clipping is 6.24 dE00 away from the truth, which is about 2.7 just noticeable differences. At 900 K it is 11.58 dE00. The ember and the charcoal both come out as flat #FF0000, the same pixel, despite being 150 K and a visible amount of colour apart.

The best candle a screen can manage

Clipping is the crude fix. We also ran the proper version: search all 16.7 million sRGB codes for the one closest in CIEDE2000 to the true 1,470 K colour at matched lightness. The winner is #DA6200, and it is still 3.89 dE00 short. For the 1,673 K hottest point the best is #C65F00 at 2.06 dE00. For a 900 K ember, #E24400 at 8.13 dE00.

There is no sRGB code within one JND of a candle flame. That is the same shape of result we found measuring the visible spectrum, where none of the 321 wavelengths fit inside sRGB either, and neon, where the problem was brightness rather than hue. Fire joins the list. Firelight is a genuinely unrenderable colour, and the warm orange every game and film and website uses for flame is a compromise nobody bothers to label as one.

Out of curiosity we measured how far off the usual suspects are from a real 1,470 K candle:

• darkorange #FF8C00 is the closest CSS keyword at 9.04 dE00.
• Crayola’s Flame #E25822, a colour named after this exact thing, is 10.87 dE00 off.
• orangered 12.46, coral 13.92, orange 17.82, red 18.77, tomato 19.01.
• gold is 35.37 dE00 away and plain yellow 46.39, so the cartoon flame palette is not even in the right neighbourhood.

If you want a defensible fire orange, #DA6200 is the measured answer for a candle and #FF8400 is the hottest thermal flame sRGB can represent honestly. More on where to use them in our piece on shades of orange.

The temperature curve runs out of blue

Now the interesting half. If blue means hot, there should be a temperature that produces blue. Push Planck’s law upward and watch what happens.

At 5,772 K, the sun’s surface, you get a near white #FFF1EA. Past that it starts to cool in the colour sense, and by 25,000 K you have the pale steel #A5BEFF of a young B type star. Keep going. At 100,000 K it is #98B3FF. The spectrum approaches the Rayleigh-Jeans limit, where intensity simply falls as the fourth power of wavelength, and the chromaticity stops moving. The limit is:

• #94B1FF, at x = 0.2399, y = 0.2340.
• Chroma C* 56.1. That is the most saturated colour any glowing object of any temperature can ever be on the blue side.

Look back at the thermal table with that number in hand. A candle flame has chroma 153.1. A dull red ember, 218.4. The bluest possible glowing thing in the universe is less colourful than an ember dying in a grate. Heat produces strong colour in one direction only.

The rate of change tells the same story. Doubling temperature moves the colour a lot at the bottom and almost nothing at the top:

• 1,000 K to 2,000 K: 16.65 dE00.
• 4,000 K to 8,000 K: 29.19 dE00, the steepest stretch.
• 16,000 K to 32,000 K: 3.08 dE00.
• 32,000 K to 64,000 K: 1.12 dE00, under half a JND.
• And 25,000 K all the way to infinity: 2.77 dE00.

Doubling the temperature of a star above 30,000 K is a change you cannot see. The entire remaining colour budget from a hot blue star to arbitrarily hot is a little over one JND. Thermal radiation has a blue ceiling and it is a low one.

So where does the blue in a gas flame come from

Not from heat. A clean premixed hydrocarbon flame has almost no soot in it, so there is nothing solid to glow. The light is chemiluminescence: molecular fragments produced in an excited state by the reaction itself, radiating on the way down. Gaydon and Wolfhard catalogued this in the book that is still the reference. The visible players are:

• CH*, a band around 431 nm. This is the blue.
• C2*, the Swan bands at roughly 474, 517 and 564 nm, which add the green you can see in a rich flame.
• OH* at 306 nm, which is the strongest band of the lot in many flames and is entirely in the ultraviolet.

We built a spectrum from those bands with plausible weights and put it through the same pipeline. A clean premixed cone computes to #0067FF at chroma C* 161.4.

That is 2.88 times the chroma of an infinitely hot blackbody. The nearest point anywhere on the temperature curve, searched from 500 K to a million K, is 12.8 dE00 away. There is no temperature that looks like a gas flame. The colour is simply not on the curve, and no amount of extra heat would put it there.

What we think the real rule is

Blue does correlate with hot, and the correlation is not an accident. It is just indirect. Blue means the fuel and air were mixed before they burned, which means the combustion is complete, which means for a given fuel you are getting all the available heat instead of throwing carbon away as soot. Yellow means the opposite. So a blue hob flame really is hotter than a yellow one, on the same hob.

What does not follow is that blue is a reading of temperature across different fires. Compare two rows. A Bunsen burner’s blue cone runs around 1,850 K. An oxy-acetylene torch reaches 3,480 K, nearly twice as hot, and its thermal component computes to #FFC68A, a warm cream. The hotter flame is the yellower one by a factor of two in temperature.

The cleanest demonstration is hydrogen. A hydrogen and oxygen flame reaches about 2,800 °C, hotter than almost anything in a workshop, and it is nearly invisible. No carbon means no CH*, no C2* and no soot. What is left is OH* at 306 nm, and the CIE observer is not defined below 360 nm because the eye response there is zero. The hottest flame in the room can be the one you cannot see at all, which is exactly why hydrogen fires are treated as a safety problem rather than a lighting one.

The honest answer to “what colour is the hottest fire” is that the question mixes up two unrelated mechanisms. Among glowing things, hotter runs orange to white to pale blue and then stops. Among gas flames, colour is a fingerprint of chemistry and carries no temperature information at all. Blue being the popular answer is a coincidence of the two scales overlapping near a kitchen hob.

A slightly rich flame is pink

Real flames are mixtures. The premixed cone sits inside a sooty envelope, and the ratio shifts as you turn the air collar. We titrated it: hold the blue chemiluminescent spectrum and the 1,700 K soot glow together in proportions measured by how much of the visible light each contributes, then track where the colour goes.

It does not slide along the temperature curve. It bulges away from it. At 70 percent soot the mixture computes to #FF96F0, a clear pink, 24.61 dE00 off any blackbody. Blue at 431 nm plus orange soot is a wide mixture, and wide mixtures of the two ends of the spectrum land in the extra-spectral wedge where magenta lives, the region no single wavelength and no temperature can reach.

This is not a modelling artefact. It is the purplish fringe you can see on a gas hob turned slightly rich, and on the boundary layer of a wood fire where flame front meets smoke. The reason the effect is so strong for such a small visible contribution is that CH* sits at 431 nm, where the eye response is 0.021. A band that supplies a tenth of the light you see from there is supplying an enormous amount of radiant power, and chromaticity answers to power, not to brightness.

We should flag one honest caveat. The exact figures depend on the band weights we chose, and a wide random sweep of plausible weightings finds 63 percent of them more saturated than any possible blackbody, not all of them. Mixtures dominated by the green Swan bands can pass close to the temperature curve. The conclusion that survives is about the blue specifically: CH* at 431 nm is unreachable by heat, and a flame whose colour is set by it is not reporting a temperature.

Flame tests, and the elements whose colours you cannot see

Drop a metal salt into a flame and the atoms emit their own lines. This is the school experiment and it is also every firework. The lines themselves are in the NIST atomic spectra database, so the only question worth asking is what the eye does with them. Here is every common flame test line with the CIE photopic response at that wavelength, and the luminous efficacy that follows.

Two things in that table are worth stopping on.

Barium sits almost exactly on the peak of human vision. Its 553.55 nm line lands 1.45 nm from the 555 nm maximum of the luminosity function, at V = 0.99946. A watt of barium green delivers 682.6 lumens against a theoretical ceiling of 683. No other element in common use comes close. Barium is, in the strict photometric sense, the most efficient colour in chemistry.

Potassium’s famous lilac is made of leftovers. Potassium’s strongest emission by a wide margin is the resonance doublet at 766.5 and 769.9 nm. The eye response there is 0.000034. Sodium at 589 nm is 22,523 times more visible per watt. Potassium’s own violet lines at 404.4 and 404.7 nm are 18 times more visible than its main output, and those weak violet lines are what you are actually looking at in a potassium flame test. The bright part is invisible and the visible part is faint.

It gets more extreme down the group. Caesium’s resonance lines at 852 and 894 nm sit at V = 0.0000005, which makes them roughly 1.7 million times less visible per watt than sodium. Caesium is a spectacular emitter that emits almost nothing you can see.

And sodium is the opposite problem. At 589 nm it is close enough to the sensitivity peak, at V = 0.766, that a trace contaminant swamps everything else in the flame. Anyone who has tried to run a flame test on glassware that once held table salt knows this as a practical nuisance. The table explains why it is unavoidable: sodium is thousands of times more efficient at converting watts into apparent brightness than most of what you are trying to detect.

One correction while we are here, because it appears in a lot of school notes. The blue in copper flame colourant packets is not the copper lines listed above, which are green. It comes from copper halide bands, mostly CuCl, in the 428 to 452 nm region. Burn copper sulphate and you get green. Burn copper chloride and you get blue-green. The chloride is doing the work.

Every fire colour there has ever been fits on a line

Here is a number that surprised us. Sweep the whole in-gamut Planckian locus, from the 1,900 K floor up to a million K, at max channel normalisation, and count the distinct sRGB hex codes you touch.

498. Four hundred and ninety eight codes out of 16,777,216, which is 0.003 percent of the space. Restrict it to temperatures a real flame can reach and stay in gamut, 1,900 to 3,500 K, and it collapses to 201.

Thermal colour is one dimensional. Temperature is a single number, so the colours it produces trace a single curve, and a curve through a three dimensional space has no volume. This is why firelight looks like firelight whether it is a match or a furnace, and why warm and cool feel like one axis rather than two. The whole aesthetic vocabulary of glow, from candlelight to golden hour to tungsten, is 498 pixels wide.

Gas flames escape the line, which is the other half of why they look strange. A blue hob flame is not a warmer or cooler version of anything. It is off the curve, in a part of colour space that heat cannot visit.

A fire is a heater with a light leak

One last measurement, because it reframes the whole subject. For each temperature, what fraction of the radiated power actually falls in the visible band, and what is the luminous efficacy of that radiation?

• Glowing ember, 900 K: 0.0002 percent of its radiation is visible light. Efficacy rounds to zero lumens per watt.
• Candle, bright zone at 1,470 K: 0.138 percent visible, 0.066 lm/W.
• Candle at its hottest, 1,673 K: 0.446 percent, 0.295 lm/W.
• Tungsten filament, 2,700 K: 8.4 percent, 12.4 lm/W. This is the bulb everyone called inefficient.
• Oxy-acetylene, 3,480 K: 20.4 percent, 36.7 lm/W.
• The sun’s surface, 5,772 K: 46.5 percent, 92.0 lm/W, which is a decent LED.

A candle is roughly 190 times worse than the incandescent bulb that got banned, measured per watt of radiation. The sun is 1,400 times better than the candle. Our piece on the colour of the sun works through why the solar spectrum sits where it does; the short version is that 5,800 K happens to put the Planck peak inside the visible band, and evolution then built eyes around the result.

So when people say fire is a poor light source they are understating it by three orders of magnitude. Fire is a heater that leaks a rounding error of light, and for most of human history that rounding error was the only light there was after sunset. It also explains the colour. A source that inefficient is necessarily far down the low temperature end of the curve, which is the orange end, which is out of gamut.

The short version

• Sooty fire glows thermally and its colour is pure temperature. Below 1,900 K that colour is outside sRGB, and every flame in your house is below 1,900 K.
• The best sRGB stand-in for a candle is #DA6200, still 3.89 dE00 short. The nearest CSS keyword, darkorange, is 9.04 dE00 out.
• Heat cannot make anything bluer than #94B1FF at chroma 56.1, and a gas flame’s blue is 2.88 times that, sitting 12.8 dE00 off the temperature curve.
• Blue therefore is not a temperature reading. It is a soot reading, and soot happens to correlate with wasted heat.
• Barium is 1.45 nm from the peak of human vision at 682.6 lm/W. Potassium’s brightest lines are 22,523 times less visible than sodium’s, and caesium’s are 1.7 million times less.

If you want to test your own eye against the warm end of the space, that is where our game is quietly at its hardest. Deep oranges and browns sit in a region where discrimination is poor and vocabulary is poorer. Try the Name That Colour mode and watch how quickly you run out of words between #DA6200 and #FF8400, or play the standard colour memory game and see how you do when two rounds in a row land in firelight. Our full set of colour games has a few other ways in.

A note on method

Everything above comes from one pipeline. Planck’s law for the thermal spectra, Gaussian bands at published centre wavelengths for the chemiluminescent and atomic emission, the CIE 1931 two degree observer at 1 nm from the CVRL tables, direct integration for tristimulus values, the sRGB matrix and transfer function for encoding, and CIEDE2000 for every distance. Chroma and all colour differences are computed with Y normalised to 1, so lightness is held constant and what we are comparing is hue and saturation only. That is the right comparison for a self luminous source, since how bright a flame looks depends on how much of it there is rather than on what it is made of.

Three judgement calls worth stating. We treat soot as a grey body rather than fitting a wavelength dependent emissivity, which shifts the warm end slightly bluer than reality and therefore makes our out of gamut figures conservative. The blue flame spectrum uses assumed relative band intensities, which is why we ran the sensitivity sweep and reported it rather than quoting a single number as if it were measured. And the temperature figures are literature values, not our own thermometry: the 1,400 °C candle maximum is NIST’s, sourced in turn to Gaydon and Wolfhard. What we measured is the colorimetry, which is the part that gets stated wrongly. One bookkeeping note: our root find puts the sRGB blue channel through zero at 1,900.1 K, and our colour temperature chart reports the same boundary as 1,901 K from a one kelvin walk up the locus. Same number, different search.