collimation
**A collimator is a filter that throws away most of what it is given, and the reason a fab would accept that is that a sputter target emits into the wrong shape.** Atoms leaving a magnetron target leave with something close to a cosine angular distribution about the target normal — the same broad lobe that governs thermal evaporation from a surface. Spread over the wafer, that distribution is superb for coating a flat field and useless for reaching the bottom of a deep, narrow feature. A via with an aspect ratio of three accepts only the small cone of trajectories that can pass through its opening without striking a wall; every atom arriving outside that cone lands on the field, on the mouth shoulder where it builds the overhang, or on the upper sidewall. The deposition is not failing — it is delivering exactly the distribution it was given. Collimation is the decision to fix the distribution before it reaches the wafer rather than fix its consequences afterward, and to do so by simple geometry: put a honeycomb of parallel tubes between the target and the wafer, and let the tubes absorb everything that is not travelling nearly straight down.
The geometry of that filter is entirely determined by the aspect ratio of a single channel, and the arithmetic is unforgiving:
$$\theta_{acc} \;=\; \arctan\!\frac{1}{\mathrm{AR}_{c}}, \qquad T \;\simeq\; \sin^{2}\theta_{acc} \;=\; \frac{1}{1+\mathrm{AR}_{c}^{2}}$$
The acceptance half-angle is set by the channel length over its diameter, and the transmitted fraction of a cosine emitter falls as the reciprocal of one plus the square of that ratio. Put numbers to it. A channel as long as it is wide accepts forty-five degrees and passes about half the flux. Go to an aspect ratio of two and the acceptance narrows to twenty-six degrees while transmission falls to one fifth. At three, the acceptance is eighteen degrees and roughly a tenth of the sputtered material reaches the wafer; the other ninety percent is plated onto the inside of the collimator. That is the whole economics of the technique in one line — **directionality costs throughput as the square of the directionality you asked for**, and the material you paid to sputter is not merely slowed down but permanently removed from the process.
Long-throw sputtering reaches the same end by the same physics with the walls taken away. Instead of blocking off-angle atoms, increase the target-to-wafer distance so that the wafer subtends a small solid angle from any point on the target; an atom leaving at a steep angle simply misses the wafer and lands on the shield. The angular selectivity improves with throw distance, and so does the loss, in the same inverse-square manner. Long throw carries an extra constraint that collimation does not: it only works at low pressure. The whole point is that atoms travel in straight lines from target to wafer, and that requires the gas mean free path to exceed the throw distance — typically a few millitorr or below, which limits the plasma density and therefore the sputter rate again. Both approaches are passive. Both discard rather than redirect. Both were, for a period in the nineteen-nineties and early two-thousands, the only way to get titanium and titanium nitride into a contact that had become too deep for conventional sputtering.
**What makes a collimator genuinely difficult to run, rather than merely expensive, is that the filter changes while you use it.** The ninety percent of material that does not reach the wafer is not vented away; it condenses on the channel walls, which is to say it grows inward, which is to say the channel diameter shrinks while the channel length stays fixed. The effective aspect ratio therefore climbs monotonically from the moment a fresh collimator is installed, and transmission falls with the square of the shrinking diameter:
$$\mathrm{AR}_{eff}(t) \;=\; \frac{L}{d_{0}-2\,\dot{s}\,t}, \qquad \frac{T(t)}{T_{0}} \;\approx\; \Bigl(1 - \frac{2\,\dot{s}\,t}{d_{0}}\Bigr)^{\!2}$$
This is a component whose transfer function is consumed by the process it serves, and it makes the collimator unlike anything else in the chamber. A target erodes but its angular emission barely changes. A shield collects material but does not sit in the beam path. The collimator does both — and because its aspect ratio is drifting, both the deposition rate and the angular distribution of what reaches the wafer are drifting together. A rate that falls ten percent over a collimator's life is not a rate problem that can be corrected by extending the deposition time, because the film deposited at end of life was built from a narrower cone than the film at beginning of life and has different bottom coverage, different sidewall coverage, different stress and different resistivity. Time compensation restores the thickness and not the film.
The practical consequences follow directly. Collimator life becomes a scheduled consumable replacement measured in kilowatt-hours rather than wafers, and the interval is set not by when the channels close but by when the drift exceeds what the downstream process can absorb. Every replacement is a chamber open, a pump-down, a season and a requalification. The accumulated deposit is under stress and eventually flakes, so particle excursions cluster near the end of collimator life and provide the other, less negotiable, limit on the interval. Because the honeycomb is a physical structure with a pattern, it prints: the web between channels shadows the wafer, and unless the geometry, the target-collimator-wafer spacing and the plasma are arranged to blur it out, the film carries a faint hexagonal thickness signature that is invisible on a nine-site measurement and obvious on a dense map. And the captured metal is expensive, particularly for tantalum and for the platinum-group and rare-earth targets used in magnetic stacks, so the reclaim value of a used collimator is a real line in the cost model.
| Approach | How the arrival cone is narrowed | Rate penalty | What drifts in use |
|---|---|---|---|
| Conventional magnetron | not narrowed — broad cosine lobe reaches the wafer | none | target erosion profile only |
| Collimator | off-angle atoms physically intercepted by a honeycomb | severe, scales as one over one plus AR squared | channel diameter closes, so cone and rate drift together |
| Long-throw | wafer subtends a small solid angle from the target | severe, and requires low pressure to preserve line of sight | shield loading and any pressure creep that shortens the mean free path |
| Ionized PVD | sputtered metal is ionized, then steered by the wafer sheath | modest — flux is redirected rather than discarded | coil or cathode erosion, ionization fraction with power and pressure |
**Ionized PVD replaced collimation for high-volume interconnect work, and the reason is worth stating precisely because it is a general lesson rather than a detail of one tool.** A collimator is a passive filter: it improves directionality by deleting the trajectories it does not want, so improvement and loss are the same operation. Ionized PVD is an active one: it ionizes a substantial fraction of the sputtered metal — with a secondary inductively coupled coil, a hollow cathode magnetron, or high-power impulse operation — and then accelerates those ions across the wafer sheath, which is by construction perpendicular to the wafer. An off-angle atom is not discarded but turned. Directionality improves without a corresponding loss of flux, and the acceptance cone becomes an electrical parameter set by bias and pressure rather than a mechanical one set by a part that must be replaced. The same bias also drives resputtering, so a single knob controls both delivery to the via bottom and redistribution to the sidewall. Nothing in the chamber has to be scheduled for replacement to keep the angular distribution constant. Once that existed, a technique whose central characteristic was a consumable transfer function had no path forward in a cost-per-wafer argument.
Collimation nonetheless persists where its weaknesses do not bind, and the surviving applications share a common shape: small wafer counts, a strong requirement on angular distribution, and no meaningful throughput pressure. Magnetic tunnel junction stacks for MRAM are the clearest case — the layers are angstroms thick, interface abruptness dominates device behaviour, and a narrow arrival cone reduces intermixing and shadowing across a patterned surface in a way that matters more than the deposition rate. Compound semiconductor and photonic processes use it for the same reason. Advanced packaging uses it for seed layers in deep, high-aspect-ratio through-silicon vias where nothing else geometric will reach the bottom and the wafer counts are low enough that rate is not the constraint. Some legacy contact modules never converted because the process was qualified, the tools were paid for, and the devices are still selling. And in research and pilot work, a collimator is simply the cheapest and most predictable way to obtain a narrow angular distribution without building an ionization source.
Reading a collimated process therefore means reading two things that a conventional PVD process lets you read separately. Deposition rate alone is ambiguous: a falling rate could be target erosion, could be a power or pressure shift, or could be the collimator closing — and only the last of those also changes the angular distribution. The measurement that separates them is bottom coverage in a test feature, tracked alongside rate across collimator life. If rate falls and bottom coverage holds or improves, the collimator is closing and the film is changing character; if rate falls and bottom coverage falls with it, the source is losing output and the cone is unchanged. A collimated recipe that will transfer therefore has to state the collimator aspect ratio, the accumulated kilowatt-hours at which the data was taken, the acceptance criterion on both rate and coverage rather than rate alone, and the replacement trigger — because a specification that names only a deposition time is describing a film that existed on one particular day in the life of a part that was already changing underneath it.