Why the blue LED took thirty years to arrive

CHAPTER 1Colour is not colour, colour is energy

At heart, a light-emitting diode is a simple device. Two differently doped semiconductor layers sit on top of each other. Apply a voltage and electrons drift in from one side and holes, that is missing electrons, from the other. When an electron meets a hole it drops into it and gives up exactly the energy it had in excess.

That energy difference is called the band gap. It is not a setting but a property of the material, as fixed as its density or melting point. And because the energy of a light particle directly determines its wavelength, the band gap determines the colour. The relationship fits on one line:

Red at 650 nanometres needs about 1.9 electronvolts. Green at 550 nanometres about 2.25. Blue at 460 nanometres needs 2.7. So anyone who wants to build a blue LED is not looking for a blue material. They are looking for a semiconductor whose band gap is roughly one and a half times that of the red one. Drag the slider to the right and watch how far out blue sits.

Nick Holonyak built the first visible LED at General Electric in 1962: red, made from gallium arsenide phosphide. Yellow and green followed in the sixties and seventies. After that, almost three decades passed without anything that glowed blue and could also be seen in daylight. The reason is that a large band gap is only the first of three conditions.

• The semiconductor needs a direct transition, otherwise the energy turns into heat instead of light.
• It has to be dopable in both directions, because without a p-side and an n-side there is no diode.
• It has to be growable as a usable single crystal, and on a foundation you can actually buy.

For red, gallium arsenide phosphide meets all three almost by itself. For blue, not a single material met all three at once for decades. That, not the band gap, is where the whole thing hung.

CHAPTER 2The first hurdle: direct or indirect

An electron in a crystal has not only an energy but also a momentum, and both have to be conserved in every transition. Plot the energy of the allowed states against momentum and you get the band diagram: at the bottom the valence band with the bound electrons, at the top the conduction band with the free ones.

In a direct semiconductor the minimum of the conduction band sits exactly above the maximum of the valence band. The electron drops straight down, its momentum does not change, and all of the energy goes to a photon. That works, because a photon carries energy but almost no momentum.

In an indirect semiconductor the two points are offset against each other. Now the electron also has to shed its momentum as it falls, and a photon cannot take that off its hands. It needs a phonon as well, a lattice vibration. Having an electron, a hole and a matching phonon meet at the same place at the same time is considerably less likely than the direct case. So something else usually happens instead: the energy seeps into the lattice by detours and becomes heat.

Silicon is indirect. That is why there is no silicon LED, even though silicon is the best mastered semiconductor in the world. Silicon carbide is indirect too, and silicon carbide was for years the only blue LED you could buy. It did work. Its efficiency, however, was in the range of hundredths of a percent; you had to look for it in a darkened room.

CHAPTER 3Three candidates, two dead ends

By the mid-eighties it was clear which materials were even in the running. There were three, and two of them led nowhere.

Silicon carbide: it glows, but the wrong way round

The only candidate that already existed as a commercial device, and at the same time the most hopeless one. An indirect semiconductor does not get better by being made cleaner. The physics simply does not permit a high yield. Silicon carbide LEDs disappeared from the market as soon as there was an alternative.

Zinc selenide: the favourite that died

On paper, zinc selenide was perfect. Direct transition, a band gap of around 2.7 electronvolts, so right in the blue, and almost exactly the same lattice constant as gallium arsenide, for which finished high-quality substrates were available to buy. Most of the field worked on it, with substantial budgets.

The device had just one flaw that nobody could remove: it died in operation. Under current, crystal defects multiplied, grew into dark lines through the active region and ate up the brightness. Lifetimes were in the range of hours. That is enough for a laboratory experiment, but not for a product. The problem was never solved.

Gallium nitride: the material everyone had given up on

Gallium nitride has a direct transition and a band gap of 3.4 electronvolts, which is already in the ultraviolet. Mix in indium and the band gap drops, continuously across the entire visible range. So you can tune the colour. A material could hardly fit this task better.

It just could not be made. Jacques Pankove built the first GaN light-emitting diode at RCA in 1971, but without a real p-n junction, because one half of it refused to be doped. RCA stopped the work in 1979. After that, gallium nitride was considered finished. Anyone who said at a conference in the eighties that they worked on GaN was met with pity.

CHAPTER 4Problem one: there was no foundation

Semiconductor layers are grown epitaxially. You put a substrate into a hot reactor, pass the source materials over it as a gas and let the new crystal grow atomic layer by atomic layer. In doing so, the growing crystal adopts the lattice of the foundation beneath it. If the foundation does not fit, nothing that grows on it fits either.

For gallium nitride there was no matching foundation. GaN substrates would have had to be pulled from a melt, and that would take well over two thousand degrees and a nitrogen pressure of several thousand atmospheres. So sapphire was used instead, because it is available, heat resistant and cheap. The lattice constants of the two materials, however, are about sixteen percent apart, and their thermal expansion differs by roughly a quarter.

The result was layers with a dislocation density on the order of a billion defects per square centimetre. For comparison: in gallium arsenide a device is considered dead at ten thousand dislocations per square centimetre. The early GaN films were rough, cracked and milky, and by far the friendliest word for them was polycrystal.

Hiroshi Amano found the way out in Akasaki's group in 1986, and it is so simple that it sounds implausible. You first coat the sapphire with a very thin layer of aluminium nitride, but at only about five hundred degrees. That is far too cold for orderly crystal growth; what forms is a fine-grained, almost structureless buffer layer. Only then do you heat up to the actual thousand degrees. Out of the buffer, very many very small and above all identically oriented nuclei then crystallise, and they quickly grow together. The real film no longer starts on the foreign sapphire lattice but on a surface that suits it.

Nakamura later rebuilt his own reactor at Nichia and replaced the aluminium nitride with gallium nitride as the buffer material, which was more reproducible. On top of that came his two-flow method: alongside the horizontal gas stream carrying the source materials, a second, inactive stream blows down vertically onto the substrate and presses the reactant gases against the hot surface. Without that trick, the convection rising above the thousand-degree substrate tears the gas away before it can react.

CHAPTER 5Problem two: the diode with only one half

A light-emitting diode needs both sides: surplus electrons on the n-side, missing ones on the p-side. Gallium nitride, however, was always n-conducting, entirely on its own. Build in magnesium or zinc, that is atoms meant to contribute a hole, and the crystal still stayed practically insulating. Resistivities around a million ohm-centimetres; that is not a semiconductor, that is an insulator.

Two effects overlapped, and that kept the matter opaque for a long time. The first is hydrogen. It is unavoidably present in the process, because the nitrogen source is ammonia and the carrier gas is hydrogen. Every magnesium atom built in catches a hydrogen atom, and the pair is electrically dead. The second effect is where the magnesium level sits: about two hundred millielectronvolts above the band edge, so comparatively deep. At room temperature only about one magnesium atom in a hundred gives up its hole at all. You therefore have to dope extremely heavily, which in turn strains the crystal.

The first breakthrough was a side observation. In 1989 Amano and Akasaki noticed that magnesium-doped GaN glowed more brightly under the electron beam of a scanning electron microscope, and that it was conductive afterwards. Out of that observation came the first genuine GaN diode with a p-n junction. The method was no good for manufacturing, though: the electron beam only acts near the surface and has to be moved across the wafer piece by piece.

In 1992 Nakamura worked out what is actually happening there, and that changed everything. It is not a radiation effect, it is the hydrogen. Heat the layer above seven hundred degrees in pure nitrogen and the magnesium-hydrogen pairs break apart, the hydrogen diffuses out of the crystal and the magnesium becomes active. The resistivity falls from around a million to two ohm-centimetres, and the hole density rises to about three times ten to the seventeenth per cubic centimetre. With that, p-GaN was no longer a feat but a furnace programme for entire wafers.

The counter-experiment in the diagram is not made up. The same sample, annealed in ammonia instead of nitrogen, becomes high-resistance again, because the ammonia releases fresh hydrogen as it decomposes. That reversibility made the explanation unambiguous, and a nuisance nobody had understood for decades turned into a process parameter.

CHAPTER 6Assembly: what actually happens inside the finished LED

With a smooth crystal and a working p-layer, the road was clear. In November 1993 Nichia presented the first blue light-emitting diode bright enough for everyday use, at around one candela. Its structure is in principle the one blue LEDs still have today.

At the very bottom the sapphire, on it the buffer layer, above that a thick n-conducting GaN layer as an electron reservoir. In the middle the active region: several layers of indium gallium nitride only a few nanometres thin, between GaN barriers. These quantum wells hold electrons and holes in the same place instead of letting them run through the device, and the indium fraction sets the colour. Above that a thin aluminium gallium nitride barrier that stops electrons from shooting past the active region, since they are considerably more mobile than holes. At the very top the p-layer with its contact.

This efficiency drop at high current has been given a name of its own, droop, and it is still a research topic today. In practice it means an uncomfortable inversion of the usual cost logic: you cannot make an LED lamp cheaper by running fewer chips harder.

CHAPTER 7The real puzzle: why does this work at all?

This is a good place to pause. The active region of this diode sits on a crystal with roughly a billion dislocations per square centimetre. That is one defect in an area of about three hundred by three hundred nanometres. Dislocations are non-radiative recombination centres: a charge carrier that arrives there gives up its energy as heat and is lost to the light. In gallium arsenide, not a single photon would come out at that defect density.

The answer lies in the indium. It does not distribute evenly through the InGaN layer but fluctuates on a scale of a few nanometres. Where there is more indium the band gap is slightly smaller, and a smaller band gap is a well for a charge carrier. Electrons and holes fall into these wells before they can travel far enough to find a dislocation. They then sit tight, meet each other there and recombine radiatively. The crystal is full of defects, but the charge carriers never reach them.

In hindsight that is the real point of the story. The decisive property of indium gallium nitride was not the tunable band gap; that much had been expected. It was the insensitivity to crystal defects, and nobody had predicted that. Had it been predictable, work on zinc selenide would probably have stopped much earlier.

CHAPTER 8What came of it

Blue was never the goal, blue was the missing piece. With red, green and blue you can mix any colour, and above all you can produce white light. In practice, though, this is not done with three LEDs but with a single blue one and a phosphor on top of it, usually cerium-doped yttrium aluminium garnet. The phosphor swallows part of the blue light and re-emits it as broad yellow. Blue and yellow together read as white to the eye.

The numbers behind it explain why this so quickly became a rebuild of the world's lighting. An incandescent bulb delivers about fifteen lumens per watt, a fluorescent tube seventy, a white LED in a product today one hundred and fifty to two hundred. The press release for the 2014 Nobel Prize states the order of magnitude at stake: around a quarter of the world's electricity consumption goes into lighting.

The same material family brought the violet laser diode at 405 nanometres in 1996, and with it Blu-ray, because a shorter wavelength means a smaller laser spot and more data on the same disc. The white backlight in every phone and laptop is a blue LED with a phosphor too.

In 2014 the Nobel Prize in Physics went to Isamu Akasaki, Hiroshi Amano and Shuji Nakamura, with a citation unusually close to application for this prize: for the invention of efficient blue light-emitting diodes which has enabled bright and energy-saving white light sources.

CHAPTER 9What the story leaves behind

Three things about this story carry over, and none of them has anything to do with semiconductors.

First, the problem was not where everyone was looking. The field optimised zinc selenide for years because it was the logically obvious material, and failed on a defect that could not be optimised away. The task was solved with the material that was considered finished.

Second, the two breakthroughs were not a new theory but process control. A temperature ramp and a furnace programme. What was missing was not the knowledge of what to build, but of how to make it.

Third, both key observations were side findings. A sample that mirrored instead of clouding over. A crystal that grew brighter under the microscope. Both could have been filed away as measurement errors. The difference was that somebody asked why.

Sources and original papers