Photolithography

Dry-film photoresist on copper-clad board runs the identical five operations with equipment you can buy for the price of a meal. Do it properly and measure the result rather than admiring it. MAKE THE MASK. Print your pattern on inkjet transparency film at the highest density your printer manages, or better, print two identical copies and stack them in register to double the optical density — thin ink is the commonest cause of a washed-out image, because resist responds to total dose and grey is not black. Design the pattern as a resolution test: a series of line-and-space pairs at 500, 400, 300, 250, 200, 150, 100, 80 and 50 micrometres, plus some large areas and some isolated dots. You are not making a circuit, you are measuring your process. PREPARE THE SUBSTRATE. Scrub the copper with a fine abrasive block, rinse, degrease with isopropanol, and dry. From here handle by the edges only. Fingerprints are grease, grease stops resist adhering, and every void in your pattern will sit exactly where a finger was. COAT. Peel the matte protective layer from the dry film and laminate it onto the warm copper — a domestic laminator at about 100 degrees, two passes. You are aiming for no trapped air. Bubbles become holes. EXPOSE. Sandwich mask against resist under a sheet of clean glass, emulsion side down and in firm contact, and expose to UV. A 365 nm LED array or an unfiltered fluorescent UV tube works. Contact is critical and it is worth pressing hard: any air gap lets light diffract around the mask edge, and the penumbra that results is the difference between a 100 micrometre line and a smear. Bracket the dose. This is the single most useful thing you will do. Expose a strip of identical test patterns for 30, 60, 90, 120, 180 and 240 seconds by sliding an opaque card along in steps. Underexposed negative resist washes away entirely; overexposed resist spreads, because scattered light reaches beyond the mask edge and lines close up. There is a window, it is narrower than people expect, and you cannot find it by reasoning. DEVELOP. Sodium carbonate solution, about 1 percent, at 30 degrees, with gentle agitation. Watch it: unexposed film lifts and clears in 30 to 90 seconds. Stop as soon as the copper is clean, because the developer slowly attacks the exposed resist too and over-development thins your lines. ETCH. Ferric chloride, or the hydrogen-peroxide-and-hydrochloric-acid mixture. Warm and agitated etches faster and, importantly, more evenly. Wear goggles and gloves and work over a sink; ferric chloride stains everything it touches permanently. STRIP. Stronger sodium hydroxide, 3 to 5 percent, lifts the cured resist in a minute or two. MEASURE. Under the microscope with a calibrated scale, find the finest line and space pair that came out cleanly separated and correctly sized. That number is your critical dimension, and it is the honest output of this step. 100 to 150 micrometres is a good result on a first attempt with a domestic laminator and an inkjet mask. Then measure the ERROR as well as the size. Compare the width of your developed 500 micrometre line against the mask. It will not be 500. Wet etching removes copper sideways as fast as downwards, so a line etched through 35 micrometres of copper loses about 35 micrometres from each side. That is undercut, it is limit 4 from the previous step, and you have just measured it.

One patterned layer is a printing exercise. Two is semiconductor manufacturing, and the difference is registration. Every layer after the first must land in a defined position relative to the ones already there. A transistor gate that sits 200 nanometres off its channel is not a slightly worse transistor, it is a short circuit. The tolerance on this — overlay — is typically a fifth to a third of the critical dimension, so a 5 nanometre process needs layer-to-layer alignment to about 1 nanometre, across a 300 millimetre wafer, at temperature. DEMONSTRATE IT YOURSELF, because it is far harder than it sounds. Take a fresh board and run the process twice. First layer: a pattern of large crosses and a grid of 200 micrometre squares. Strip, re-laminate fresh resist, and expose a second mask carrying complementary crosses that should nest inside the first, and squares that should land exactly on the first set. Then measure how far off you are. Under the microscope, measure the offset of the second-layer crosses from the first in x and y. You will not be at zero. A hand-aligned transparency on a domestic setup typically lands 100 to 500 micrometres out, and the sources are worth separating because each has an industrial counterpart: TRANSLATION — the whole second layer is shifted. Simple positioning error. Industrially solved with an interferometric stage that knows its position to a fraction of a nanometre. ROTATION — the offset grows along the board, in opposite directions at the two ends. You put the mask down at a slight angle. Industrially, the stage rotates to null it. SCALE — the offset grows steadily from one edge to the other in the same direction. Your transparency has stretched, usually from the heat of the printer or the laminator, or the substrate itself has expanded. This is the interesting one: it is why fabs control temperature to a tenth of a degree, and why an inkjet transparency is fundamentally unsuitable for anything demanding. Measure the distance between your two outermost crosses on the mask and on the board and you can calculate the stretch directly, usually a few parts per thousand. RUN-OUT AND DISTORTION — the offsets vary in a way that is neither uniform, rotational nor linear. Non-flat substrate, non-flat mask, or local stretch. Industrially this is corrected with high-order distortion terms, and residual distortion is one of the practical limits on overlay today. What you should take away is that the alignment error is not one number, it is a FIELD, and separating it into these components is how you fix it rather than merely notice it. That analysis is identical in kind to what a modern lithography tool does between every wafer, from measurements on dedicated alignment marks placed around the die for exactly this purpose. And the count matters. A modern chip needs 60 to 100 patterned layers, every one registered to those before it. If each layer yields 99.9 percent, a hundred layers yield 90 percent. That arithmetic — small errors compounding over many steps — is the real reason semiconductor fabrication is difficult, and it is a far better explanation of the industry's structure than the size of the features.

HOW IS THIS DIFFERENT FROM THE PCB ETCHING BLUEPRINT? Honestly: the bench procedure in step 2 IS PCB etching, and that blueprint is linked from this one deliberately. The difference is what the procedure is being used to teach and where it is pointed. There, the aim is a working circuit board and the process is the means. Here, the aim is the process itself — you build a resolution test rather than a circuit, you bracket the exposure to find the window, you measure undercut and overlay as numbers, and every one of those numbers is then placed against the industrial equivalent. If you want a board, follow that blueprint. If you want to understand why a lithography tool costs as much as an aircraft, follow this one. Both use the same chemicals, and that is the point of having both. WHY POSITIVE RESIST WON. Negative resists cross-link, and cross-linked polymer SWELLS when the developer soaks it, which distorts small features and limits resolution. Positive resists dissolve rather than swell. Negative resist was standard through the 1960s because it was more sensitive and more durable; positive took over as features shrank below a few micrometres and swelling became the binding limit. Modern chemically-amplified resists are a further generation again, where one photon triggers an acid that catalyses many reactions, buying the sensitivity back. WHY IS 193 NANOMETRES STILL EVERYWHERE? Because the step past it was a cliff. At 157 nanometres, fused silica stops transmitting usefully and the programme was abandoned. The next viable wavelength is 13.5 nanometres, where nothing transmits at all — so EUV uses reflective optics of forty-plus alternating layers, in vacuum, with light made by hitting tin droplets with a CO2 laser fifty thousand times a second. Everything not requiring the finest features is still done at 193 nanometres because it works and it is paid for. WHAT IS MULTIPLE PATTERNING? A way of beating the Rayleigh limit with two exposures instead of one. Print alternate lines in two separate passes, each individually within the resolution limit, interleaved to give a pitch neither could reach alone. It works, it doubles or quadruples the cost and time of that layer, and it makes overlay error a direct contributor to line placement. It is why EUV was worth its price. WHY ARE MASKS 4x LARGER THAN THE PRINT? Projection tools demagnify, typically by four. Making the mask four times oversize eases the mask-making tolerance by the same factor, and shrinks the apparent size of any defect on it. The cost is field size: the exposed area is a fraction of the wafer, so the tool steps across the wafer exposing one field at a time. That is what a stepper steps. WHAT IS THE PELLICLE? A thin transparent membrane held a few millimetres off the mask surface. Dust that lands on it is far outside the focal plane and images as a harmless blur rather than a printed defect. A simple, cheap idea that saved the industry an enormous amount of scrap, and a good example of solving a defect problem with geometry instead of cleanliness. WHAT DID ANDRUS ACTUALLY CHANGE? He made patterning parallel. Before photolithography, features were placed one at a time, by hand or by mechanical mask. After it, every feature on the wafer is defined in a single exposure and the cost per feature falls as features multiply. Every subsequent economic fact about semiconductors — that transistors became effectively free, that complexity got cheaper rather than more expensive — descends from that one property.