Upstream of Semiconductors — Wafers, Photoresist, and Specialty Gases

How sand becomes silicon that is 99.999999999% pure, gets pulled out of a melt as one giant single crystal, and is coated with a film that dissolves where light hits it. From segregation in the Czochralski process to acid blur in chemically amplified resist — and why a handful of Japanese firms still own this layer.

The story starts with sand

Articles about semiconductors almost always open with a process node — 3nm, 2nm. But what actually arrives at the loading dock of a fab is not a circuit or a transistor. It is a mirror-polished disc, a bottle shipped refrigerated, and a cylinder of gas.

If lithography and deposition are the cooking technique, this article is about the ingredients. No knife skill saves a bad ingredient, and chips are unusually sensitive to this: a shift of one part per billion in an impurity can move the yield several hundred process steps downstream.

Ingredients have one more awkward property here. The growing regions are absurdly concentrated. Trace the upstream of any leading-edge fab on earth and you land on a handful of companies and a handful of plants. This article walks the three pillars — silicon wafers, photoresist, specialty gases — starting from sand.

Purity measured in nines

Silicon is the second most abundant element in the crust, and the raw material — quartz sand — is cheap. Wafers are expensive because purity is expensive.

Sand (SiO₂) is reduced with carbon in an electric arc furnace to give metallurgical-grade silicon, roughly 98–99% pure. The standard route from there to electronic grade is the Siemens process: react the silicon with hydrogen chloride to turn it into trichlorosilane (SiHCl₃), a liquid; distill that liquid; then decompose the purified gas onto a hot silicon rod so that only silicon deposits.

The distillation step is the trick. Pulling impurities out of a solid is hard, but once you convert the material into a liquid whose impurities have different boiling points, you have moved the problem onto ground the chemical industry has worked for a century. This is a large part of why people say the semiconductor industry is a child of the chemical industry. The resulting polysilicon is often described as "eleven nines" — 99.999999999% pure.

The thing worth internalizing is that purification is multiplicative, not additive. Each pass cuts the remaining impurity by some factor. You pay a cost that grows with the number of stages, and purity improves by orders of magnitude.

Czochralski growth — pulling a crystal out of molten metal

Purity alone still isn't a wafer. The next requirement is a single crystal: one enormous crystal whose atoms are aligned end to end without interruption. In polycrystalline material, grain boundaries scatter carriers and the result is useless as a transistor.

That crystal comes from the Czochralski (CZ) process. Jan Czochralski found the method for metals in 1916; Bell Labs adapted it to silicon around 1950. The recipe is almost disappointingly intuitive. Melt polysilicon in a quartz crucible above 1414 °C (silicon's melting point), touch a small, correctly oriented seed crystal to the surface, then rotate and slowly pull it upward. The moment the melt touches the seed it freezes onto it, inheriting its atomic arrangement. Pull slowly and the crystal grows fat; pull fast and it necks down. It feels a lot like drawing taffy.

There is one classic trick here. At the very start of the pull, the operator deliberately forms a thin neck about 3mm across. Seed crystals always contain dislocations — misalignments in the atomic lattice — but dislocations tend to migrate out to the side surface of a crystal, so if you make the cross-section extremely narrow they all escape. This "Dash necking," found by William Dash in the late 1950s, is what makes dislocation-free silicon possible. It also means a few hundred kilograms of crystal hang from a neck a few millimetres wide, which is mechanically ridiculous.

And as the pull continues, the crystal quietly acquires a personality that varies along its length. That is where lot-to-lot variation actually comes from.

The top and the tail are not the same

CZ growth has an unavoidable quirk: segregation. When the melt contains impurities or dopants — elements deliberately added to create electrical conduction — the portion that freezes and the portion that stays molten do not end up at the same concentration. The ratio is the segregation coefficient $k$.

$C_s$ is the concentration in the solid, $C_l$ the concentration in the liquid. Strip the notation away and it is a ratio which says what fraction of the impurity sitting in the melt gets carried into the solid at the instant it freezes. Boron sits near $k \approx 0.8$, phosphorus near $k \approx 0.35$. A $k$ below 1 means the freezing solid leaves impurities behind — so the remaining melt keeps getting richer.

Accumulate that effect over a whole pull and you get:

Here $g$ is the fraction of the ingot that has already solidified (0 at the start of the pull, 1 at the end) and $C_0$ is the concentration originally loaded into the melt. Put in words, all it says is that the start of the ingot is lean, the end is rich, and the very end climbs steeply. The factor $(1-g)$ is the fraction of melt still unfrozen, so as that pool shrinks the impurities left behind get crowded into less and less liquid — the same reason a sauce gets stronger as you reduce it.

Numbers make the shape concrete. For phosphorus ($k \approx 0.35$), taking the starting concentration as 1, you are at roughly 1.6× when half the ingot has frozen and roughly 4.5× at 90%. The first half is gentle; the last tenth runs away. That is because the exponent $k-1 = -0.65$ is negative, so the expression diverges as $g$ approaches 1.

On the floor this shows up as resistivity differing between the top and the tail of the ingot. That is why wafer specifications quote a resistivity range, and why the head and tail that fall outside it are scrapped or sold as a different grade. A good share of "same lot, different behaviour" originates right here.

The crucible leaves its own signature. Quartz (SiO₂) dissolves slowly into the melt, so CZ silicon always contains oxygen. That is not purely bad news: during thermal processing the oxygen forms fine precipitates that act as a trash can for metallic impurities — gettering — while also mechanically strengthening the wafer. For applications that dislike oxygen, such as high-voltage power devices, the crucible-free float zone (FZ) method is used instead.

Becoming a disc

The finished ingot is a cylinder roughly 300mm across and two metres long. It is ground to an exact diameter, sliced with a diamond wire saw — the width of the wire is lost as kerf dust — then lapped, etched, and polished by CMP to a mirror finish. High-performance logic usually adds an epitaxial layer on top.

Purity is only part of the spec. There is flatness (a scanner's depth of focus is only tens of nanometres, so a bowed wafer defocuses at the edge), particle counts at tens of nanometres and above, and edge geometry. Being able to hit all of that repeatably — the yield of the material itself — is the real barrier to entry. See Yield and DFM for how yield compounds.

Photoresist — translating light into "dissolves / doesn't"

However sharp an image the scanner projects, it means nothing if the film receiving it is dull. Even the 13.5nm light of EUV lithography ultimately lands as a chemical reaction in this layer.

Resist is a resin backbone, a photoactive compound, and a solvent to make it coatable. The classic i-line (365nm) formulation was novolac resin plus DNQ. Novolac dissolves happily in alkaline developer (2.38% TMAH is the standard), but DNQ acts as a weight that inhibits dissolution. Expose it to light and the DNQ converts to a carboxylic acid, which now promotes dissolution instead. The weight turns into a float, so exposed and unexposed regions differ in dissolution rate by orders of magnitude. That gap is contrast.

One photon, hundreds of reactions

As wavelengths shortened to 248nm, 193nm, and 13.5nm, a hard problem emerged: there aren't enough photons. Scanner throughput is set by delivered energy, so an insensitive resist directly cuts factory output.

The answer, devised in the early 1980s by Ito, Willson and Fréchet at IBM, was chemically amplified resist (CAR), and the idea was catalysis. Exposure makes a photoacid generator (PAG) release a single acid molecule. The wafer is then heated — the post-exposure bake (PEB) — and the mobile acid strips protecting groups off the resin one after another. The acid is not consumed, so one acid molecule catalyses hundreds of the same reaction. Deprotected resin dissolves in developer. One photon becomes hundreds of chemical events, which is why sensitivity jumped by orders of magnitude. Everything from ArF onward, including mainstream EUV resist, descends from this idea.

The price of amplification — acid blurs

Catalysis has a side effect. During PEB the acid diffuses thermally, so it walks into regions that were never exposed and deprotects the resin there too. The diffusion length is a few nanometres, which is fatal when line widths are in the low tens. The image you worked so hard to project sharply gets smeared at the chemistry stage. The countermeasure is a quencher — a base that neutralizes acid — which kills the thinned-out acid at the fringes and lets only the concentrated core survive. It reshapes the exposure distribution into something spikier.

Shot noise is a bill from physics

Resist developers talk about a three-way trade-off: resolution, line-edge roughness (LER), and sensitivity cannot all be optimized at once. At the deepest level, statistics is what enforces it. Exposure is really "how many photons landed at each point of the film," and arrivals are random, so a location receiving $N$ photons on average carries an unavoidable spread:

Read in words, it says 100 photons on average means about 10% fluctuation; 10,000 means about 1%. Since the denominator is $\sqrt{N}$, halving the spread costs four times the photons. Raise sensitivity so you can cut the dose, and $N$ falls, and the edges of your lines get rougher.

EUV makes this structurally worse. A 13.5nm photon carries about 92 eV, while 193nm ArF light carries about 6.4 eV — so for the same delivered energy, EUV supplies roughly 14× fewer photons. EUV roughness is therefore a bill from physics that recipe tuning alone cannot pay, which is why tin-based metal-oxide resists, with their much higher absorption, have been pushed into production.

Specialty gases — the invisible material

By volume the fab consumes mostly bulk gases such as nitrogen and argon, but the cost and difficulty concentrate in specialty (electronic) gases.

• Etch: CF₄, C₄F₈, SF₆ for oxides; Cl₂, HBr, BCl₃ for metals and polysilicon. The selectivity between the film you want to remove and the one you don't is largely a gas choice
• Deposition: SiH₄ (silane), NH₃, WF₆, TEOS. Most are pyrophoric, corrosive, or both
• Dopants: B₂H₆, PH₃, AsH₃ — among the most acutely toxic materials in the plant, dangerous at ppm levels
• Lithography: the ArF excimer laser gas mix is mostly neon as a buffer, with only small amounts of Ar and F₂

That last one exposed how thin the supply chain is. Neon is a byproduct of cryogenic air separation, and for historical reasons the large air-separation capacity of the former Soviet bloc — Ukraine in particular — was a major source of semiconductor-grade neon. The 2022 invasion cut that off and prices spiked. A gas that sits inside the laser and doesn't even react turned out to be a bottleneck, and essentially nobody had it on their risk register.

Gases live or die on purity too: an inadequate purge during a cylinder changeover can let in enough air to write off a lot. There is an environmental bill as well — the NF₃ used for chamber cleans and the SF₆ used in etch are greenhouse gases with 100-year warming potentials more than ten thousand times that of CO₂, so abatement scrubbers are mandatory rather than optional.

Why Japanese firms hold such large shares

All three areas are dominated by Japanese suppliers. Shin-Etsu Chemical and SUMCO together account for a substantial share of the world's silicon wafers; leading-edge photoresist is a list of JSR, Tokyo Ohka Kogyo, Shin-Etsu, Sumitomo Chemical and Fujifilm. High-purity hydrofluoric acid and CMP slurry look much the same. Japan is routinely described as having lost the semiconductor business, so why did the upstream survive? The reasons are structural.

They came out of chemicals and photography. Photoresist is a cousin of photographic emulsion. Decades of organic synthesis and precision coating built up in silver-halide film, dyes and pigments transferred sideways. These firms entered as chemical companies, not device companies, and that framing mattered.

Purity is a process, not a recipe, so it resists copying. Controlling impurities at ppb does not follow from knowing a molecular formula. Piping materials, cleaning sequences, containers, transport — it is a mass of tacit knowledge that never appears in a patent.

Qualification locks customers in. Switching resist or gas vendors means re-establishing process conditions, re-running reliability tests, and proving yield in volume. At leading-edge nodes that can take on the order of a year, so an incumbent material rarely gets displaced on price alone.

The market is too small to be attractive. Materials revenue is modest next to tools or chips, while the specifications are the harshest in the industry. The return on entering is poor, so each niche settles into an oligopoly.

They were insulated from the device industry's defeat. The materials makers' customers were fabs worldwide, so when domestic device makers lost the DRAM war it did not end the materials business.

None of this is unambiguously good news. An oligopoly is the same thing as a single point of failure. In the 2011 Tōhoku earthquake, Shin-Etsu's Shirakawa plant — a major source of 300mm wafers — went down, and fabs worldwide discovered how thin their buffers were. When the Japanese government tightened export controls on photoresist, hydrogen fluoride and related materials bound for Korea in 2019 (relaxed in 2023), it demonstrated that this structure can be used as a diplomatic instrument, and in response accelerated Korean localization and supplier diversification. Owning a chokepoint is a risk for the holder, not only for the held.

How this shows up on the job

Who touches it, and when. When a new node is brought up, process integration engineers and materials procurement at the device maker settle the POR (Process of Record) for resist and gases. Once in volume production, qualifying second sources for cost and business-continuity reasons becomes steady-state work.

The parameters you actually read. For wafers: resistivity range, interstitial oxygen (Oi), flatness (SFQR), particle counts, edge exclusion. For resist: dose, PEB temperature and time, develop time, film thickness, and LER/LWR measured on CD-SEM. Process window is characterized with a FEM (Focus-Exposure Matrix) — step focus and dose on a grid and look at the area over which dimensions stay in spec. For gases: purity grade (6N and similar), mass-flow-controller setpoints, and cylinder changeover purge procedures.

Traps that turn into incidents.

• Resistivity differs between ingot head and tail. Misdiagnosing within-lot spread as a tool problem means you can tune the tool forever without converging
• A few degrees of PEB temperature moves critical dimensions. Hotplate non-uniformity maps straight onto CD variation across the wafer, so suspect the bake before you suspect the resist
• Resist ships refrigerated and has a shelf life. PAG and quencher degrade over time; "same part number, different sensitivity" is usually a lot-history problem
• A second source with the same spec sheet is not the same material. Trace-metal profiles that nobody wrote down will move your yield — which is exactly why qualification takes a year
• An inadequate purge at cylinder change admits moisture, and it surfaces not as an incident but as a slow yield decline, which makes it hard to catch
• Nobody has audited tier-2 and tier-3 suppliers. As neon showed, the risk hides in layers whose names you do not know

Questions you get asked. "CZ wafers always contain oxygen — why is that tolerated, even welcomed?" (Line of reasoning: precipitates provide gettering and mechanical strengthening, but you don't want it in the device layer, so you separate it out with epi or a denuded zone.) "What sets the resolution limit of a chemically amplified resist?" (Line of reasoning: acid diffusion length and photon shot noise; you fight the first with quencher and PEB conditions, the second with dose and absorption.) Both probe the same instinct — that the limiter can sit in the material rather than in the tool.

Wrapping up

• A wafer's price is not the price of sand; it is the price of purity and single-crystallinity, delivered by distillation and by the pull
• Segregation in CZ growth makes the head and tail of one ingot chemically different. That is the physical origin of within-lot variation
• Chemical amplification bought orders of magnitude in sensitivity and introduced a new resolution limit in exchange: acid blur
• Japanese dominance upstream is a product of chemical-industry ancestry, tacit process knowledge, sticky qualification, and an unattractively small market — and it is simultaneously a single point of failure for the world

For what happens to these materials once they reach the fab, see How a chip gets made.