What is a photomask? The glass plate that prints a chip

Someone has drawn a chip on a screen, and a factory has to print that drawing onto thousands of silicon discs without anyone touching a pen. The thing that carries the drawing into the factory is a square of glass.

A photomask is a flat plate of very pure glass with a thin, patterned film on one side that blocks light. A lamp shines through the clear parts and copies the pattern onto a light-sensitive coating on a wafer. One plate carries one layer of the chip, so a finished chip needs a stack of them, called a mask set. In a chip factory the same plate is also called a reticle; the two words used to mean different things, and that story comes up below. The photograph at the top of this page is one such plate, tilted against a dark room: a clear glass square with a dark, dense pattern of lines and blocks in the middle and an empty glass border around it. The dark parts are chrome. Everything clear is where light will get through.

The plate in your hand

The common plate for chip work is 6 inches square and a quarter of an inch thick, 152 mm × 152 mm × 6.35 mm. Mask shops call that size a 6025, after the 6 inches and the 0.25 inch, and it is a standard size from SEMI, the industry body that writes such specifications. Older tools take 5-inch plates, which are thinner, about 0.09 inch.

The glass is fused silica, which mask people call quartz: pure silicon dioxide with no additives. It lets deep ultraviolet light through, it is polished very flat, and it barely grows when it warms up. That last property matters: every plate in a set prints on top of the prints made before it, and a plate that swells under a hot lamp moves its pattern. Cheaper soda-lime glass, the family of window glass, expands more, blocks more ultraviolet, and is typically only flat to 5–15 µm, which is fine for coarse work.

The dark film is called the absorber, and on most plates it is chrome. What mask shops call chrome is a stack: about 80–100 nm of chromium metal with 10–30 nm of chromium oxide on top. The oxide is an anti-reflective coat, so light bouncing off the metal does not expose the coating a second time. The stack is about 100 nm thick and has an optical density of 3, which means it lets through about one part in a thousand of the light. That is dark enough that the coating underneath never notices.

The pattern stays out of a 3–10 mm exclusion zone at the border, where coating and chrome are less uniform, and on a stepper plate it has to sit centered, often within 0.5 mm. For features of 10 µm or larger there is a cheap cousin: a film mask, a polyester sheet printed on a photoplotter at 4,000 to 40,000 dots per inch.

What the light does to the wafer

The factory is not taking a photograph of your screen. It is exposing photoresist, a coating that changes chemically where light lands. A developer then washes away either the lit areas or the dark ones, depending on the resist. What stays on the wafer is a stencil. The next step works through that stencil: it etches away a film, implants ions into the silicon, or lets metal stick only in the open areas. Then the resist is stripped, a new film goes on, and the next plate is loaded.

One printed step needs one plate. Diffusion, gates, contacts, every metal level and every via level: each has its own glass. A study of mask costs from 2006 counted up to 30 plates in a set. When a layer is too dense to print in one exposure, it is split across two or more plates and printed twice, so newer processes need even more. The plates are aligned to each other with alignment marks, small crosses and boxes that every plate prints in the same place. If a mark lives on a layer that never becomes a plate, the next print has nothing to find.

Chrome where you drew, or chrome everywhere else

A plate has two tones: opaque and clear. Which one your shapes become is called the tone, or polarity, of that plate. A clear-field plate is mostly clear glass, with chrome only where you drew. A dark-field plate is mostly chrome, with clear glass only where you drew. A gate plate is often clear-field. Contact and via plates are usually dark-field, because the factory wants small holes of light there, not small islands of shadow. The resist then decides again: a positive resist dissolves where light hit it, a negative resist stays where light hit it.

The picture is a LayoutEditor view of a small teaching cell, a single inverter drawn on the layer numbers of the open SKY130 process. On the left is the drawing as a designer sees it, on a cream background: the tan n-well, two green diffusion areas, the plum poly gate crossing both, the small square contacts (licon), and the teal local-interconnect straps (li1) for power, ground, input and output. Every layer is visible at once, in colors someone picked. In the middle is the poly plate in clear field: pale glass with chrome only where poly was drawn, so the gate and its stub to the input contact appear as one dark T. On the right is the contact plate in dark field: chrome everywhere in the window, and seven clear holes where the seven contacts were drawn. That plate was not painted. LayoutEditor's rule engine subtracted the contacts from the plate window, the same kind of Boolean step a mask shop runs, and the resulting polygon with seven holes was read back into the view.

A layout file in GDSII, the long-standing layout format, has no field that says "this layer is chrome," and the colors in your editor have nothing to do with the plate. Tone travels on a separate document, the mask order form. Do not invert the data yourself to help: if you invert it and also ask for dark field, the shop inverts it again, and the holes become the only chrome on the plate.

The plate also has a front and a back. Plates are written chrome up and usually used chrome down, toward the wafer, so the pattern looks mirrored from one side. Mask people call the original orientation right-reading and the mirror image wrong-reading, and the order form states which one is wanted from which side. Text in the layout is the quickest check on the finished plate.

Same size, or four times bigger

The photograph above is a contact aligner, the simplest machine that uses a plate. The square plate sits in clamps a hair above the round wafer, and the microscope column looks down through both. An operator turns the micrometer screws until the marks on the plate line up with the marks already on the wafer, then fires the lamp. The whole wafer is exposed in one flash, and one millimeter on the glass is one millimeter on the wafer. The plate either touches the resist or floats tens of micrometers above it, and contact printing pulls bits of resist onto the plate, so those plates wear out. Many university labs and makers of micro-machines (MEMS) still work this way, because the tool is a bench and a 1:1 plate is cheaper.

Historically, "photomask" meant exactly this kind of plate: a 1:1 copy that covers the whole wafer. A reticle was a plate carrying one enlarged chip, first ten times the size, used to make those photomasks. Projection tools later used the reticle directly, and today the two words mean the same plate.

The second photograph is a reduction stepper. The reticle sits on its stage at the top, the long black lens column is in the middle, and the wafer lies on the chuck at the bottom; the person in a cleanroom suit in the background gives the scale. The lens shrinks the image four or five times: five on older steppers, four on newer scanners, including the extreme ultraviolet (EUV) machines. The lens can only see a field, not a whole wafer. A common scanner field is 26 mm × 33 mm on the wafer, so the chrome for that field is 104 mm × 132 mm on the glass, which fits a 6-inch plate with room for marks and clamps. The machine prints one field, steps the wafer, and prints again. If several small chips fit in one field, the plate carries several copies, and each flash prints them all.

The shrink also forgives the mask shop: an error of 4 nm on a 4× plate ideally becomes 1 nm on the wafer. At the smallest sizes that forgiveness fades. A 25 nm line on the wafer is a 100 nm line on the plate; a 5 % tolerance is 1.25 nm on the wafer and 5 nm on the glass. And a mask error may print larger than a quarter of its size. That ratio is the mask error enhancement factor, and it can rise above 1.

The pellicle: a dust cover that stays out of focus

A stepper reticle usually wears a pellicle: a transparent polymer film about 1 µm thick, stretched over a metal frame that is glued to the chrome side of the plate. Frame heights run from 5 to 10 mm, and 6.35 mm is common, the same as the thickness of the plate. The lens focuses on the chrome with a depth of focus of less than 1 µm, so dust that lands on the film is millimeters out of focus and prints as nothing. With a 6.3 mm stand-off, particles smaller than about 90 µm do not shrink the usable process window. A 6-inch plate needs a pellicle only on the chrome side, because the quarter inch of glass already keeps dust on the back out of focus; thinner 5-inch plates often get one on both sides. Dust that reached the chrome before the pellicle went on prints on every chip of every field.

Binary, phase-shift, and EUV masks

A plate that only blocks or passes light is a binary mask. As features shrank toward the wavelength of the light, mask makers began to use the phase of the light as well; features of about 150 nm and below generally need some form of phase shifting.

Mask type	Pattern film	Typical use
Binary	~100 nm chrome on quartz, blocks light	most layers, aligners, older nodes
Attenuated phase-shift	MoSi, passes ~6 % of the light, shifted 180°	in production since the 90 nm node
Alternating phase-shift	quartz etched for a 180° shift	gates at 65 nm, plus a trim exposure
EUV, reflective	40–50 Mo and Si layers, Ru cap, Ta-based absorber	the finest layers, 7 nm and below

A binary plate is the chrome plate described above. An attenuated plate replaces the chrome with a film of molybdenum silicide (MoSi) that lets a few percent of the light through, half a wave behind the light from the clear areas, which cancels some of the blur at each edge. An alternating plate etches the quartz itself, so light through neighboring openings is 180° apart and the dark line between them gets darker; the edges of the etched areas also print, so a second exposure trims them away. A plate for EUV light is a different object. Nothing is transparent at 13.5 nm, so the plate is a mirror: a low-expansion glass substrate under dozens of thin layers of molybdenum and silicon that reflect by interference, a thin ruthenium cap to protect them, and a tantalum-based absorber patterned on top. The light arrives at an angle of about 6°. Pellicles for EUV exist but are hard to make, because the film has to survive that light twice. The newest high-numerical-aperture (high-NA) EUV scanners shrink by four in one direction and eight in the other, which halves the field to 26 mm × 16.5 mm.

The plate is not a copy of your drawing

At these sizes the plate rarely carries your shapes as drawn. Light this small does not print a sharp rectangle. In the words of an early design-automation paper, line-end shortening, corner rounding and neighbor-dependent width changes are "fundamentally consequences of subwavelength lithography." A separate step after layout, optical proximity correction (OPC), reshapes the plate to compensate: small square serifs on outer corners, hammerheads on line ends, and thin assist bars beside lonely lines that are not meant to print. One lithography teacher puts the line at the 250 nm node: below it, chips need OPC. The plate maker is part of the error budget too: a mask writer does not reproduce every tiny serif and jog perfectly.

How a plate is written

None of the cameras above draws the chrome. A mask writer does, and the photograph above is one: a tall stainless vacuum column, a large chamber beside it, and a square glass blank on the load arm in front, not yet sitting above any wafer. The blank arrives already coated with chrome and a resist. The writer exposes the resist with a beam, the resist is developed, the uncovered chrome is etched away, and the remaining resist is stripped.

Before that, your file goes through mask data preparation. The shop reads the GDSII or OASIS file, picks the layers named on the order form, applies Boolean operations and sizing for its own etch, and converts the result into the writer's format, together with a jobdeck that says where each pattern goes on the plate. That conversion includes fracture: every polygon is cut into the rectangles and trapezoids the writer can expose. A circle drawn as a 16-sided polygon can become seven trapezoids; more vertices mean more pieces and a longer write. Coordinates also land on the writer's address grid, somewhere between 1 and 250 nm depending on the machine, so a design grid that is a whole multiple of it avoids surprise snapping.

Writers come in three broad kinds. Laser writers scan a modulated ultraviolet beam and suit plates with relaxed sizes. Shaped-beam e-beam writers at 50 keV expose one rectangle or triangle per shot, so their write time grows with the number of shapes; the first electron-beam writers of the late 1970s gave the industry the MEBES data format that is still in use. Multi-beam writers expose many pixels at once: one design uses 512 × 512 = 262,144 beams of 20 nm, at 50 keV, over an 82 µm × 82 µm field, and writes a full plate in under 10 hours no matter how complex the pattern is. Shops also run a closure check: a small mark is written at the very start of the exposure and its partner at the very end, and if the two meet, nothing drifted in between.

Checking and repairing the plate

A finished plate is inspected before it ships. Die-to-die inspection compares two copies of the same chip on one plate and flags where they differ. Die-to-database inspection compares the plate with an image rendered from the design data, and it is the only option for a plate that carries a single chip. A scanning electron microscope measures the critical dimension, the one feature size the order form names, and the deviation is written on the plate's label. A defect is anything that would print and hurt the chip: a chrome bridge between two lines is a hard defect, a fallen particle is a soft one. For advanced plates, a microscope that mimics the scanner's optics decides whether a suspect spot prints at all.

Many defects can be repaired. Laser pulses ablate unwanted chrome or phase-shift film. A focused ion beam or an electron beam deposits material where chrome is missing, and electron-beam etching and deposition give the finest repairs; ion beams are avoided on EUV plates because they damage the mirror layers. Only after inspection and repair is the pellicle mounted, and a light-scattering inspection then looks for particles that arrived during mounting.

What one plate costs

A 2006 study of mask prices found single plates ranging from $250 at the low end to $100,000 for a high-end phase-shift plate. A set multiplies that by its plate count, and every revision of a layer means a new plate. That fixed cost is why designs are checked before they go out, and why small designs often share one reticle on multi-project runs.

A common mix-up

Beginners often think the colored shapes on the screen are the mask. The screen shows a drawing; the plate and the wafer it prints are separate objects that share its numbers, as the drawing versus the thing you build shows with a plate, a wafer, and a board on one table. The shop sees one layer per plate, in one of two tones it reads from your order form. The second mix-up is size: your drawing is usually at wafer size, and the shop enlarges it by the scanner's reduction. Enlarge it four times yourself and the shop enlarges it again; the chip comes out four times too big on the wafer. The third is the file: one layout file is not one mask. It holds the whole set, plus notes and helper layers that must never become chrome.

What the file has to tell the mask shop

The glass will not guess. For each plate, the order needs:

• the mask name and the GDS layer and datatype that become this plate, and which layers stay paper only;
• the tone: are the insides of your polygons clear or chrome;
• the parity: right-reading or wrong-reading, chrome up or chrome down;
• the scale: 1× for an aligner, 4× or 5× for a stepper or scanner;
• the top cell, the one cell whose contents become the plate; a shop left to guess may pick an empty frame cell and write a clear plate, the failure that opens from idea to mask;
• a critical dimension: one feature, its size, location and tone, so the shop can measure it and confirm tone and scale at the same time.

Larger orders follow the SEMI P10 specification for this data. Before writing the order, preview each plate in your layout editor: in LayoutEditor, hide every layer except the one that becomes the plate and look at it alone. If it looks wrong as a single layer, it will look wrong in chrome.

A layout program does not draw chrome. It draws layers. The plate at the top of this page is what one of those layers becomes after data preparation, a writer, an etch and an inspection: one sheet of glass in a set, waiting for its pellicle and the lamp.