Semiconductor CNC Precision Machining: How Tight Tolerances Get Held

• 1
  Why semiconductor parts are not ordinary machined parts
• 2
  What 5-axis actually buys you in semiconductor CNC precision machining
• 3
  Material choice decides the finish you can actually hold
• 4
  Surface finish, burrs and the particle problem
• 5
  How the tolerance is verified before the part ships
• 6
  Where semiconductor CNC precision machining stops being the right answer
• 7
  Matching the machine setup to the semiconductor part

A bracket on a pump can be off by 0.1 mm and nobody notices. A wafer-handling arm cannot. In semiconductor CNC precision machining the part usually sits inside a closed environment where a burr, a stray particle or 20 μm of runout turns into a scrapped wafer lot. The tolerance on the drawing is often not the real requirement. The real requirement is that the part does not disturb the process around it.

That shifts the problem. You are not only cutting metal to a number. You are controlling edge quality, surface finish, cleanliness and how the part behaves when it heats up. A vacuum chamber wall machined to ±0.005 mm still fails if a coolant residue outgasses at 10⁻⁶ mbar. Engineers who have built these tools learn to read a drawing for the requirements that are not written down.

Volume matters too. Semiconductor equipment runs from a handful of prototypes to a few hundred units a year, rarely millions. That favors machining over stamping or high-volume casting. Setup cost is amortized over small batches, so the process has to be flexible rather than fast.

• 1
  Cleanliness is a specBead blast media, polishing compound and tapping fluid all have to be removed.
• 2
  Small batches are normalOne prototype to a few hundred units, so tooling-heavy processes rarely pay off.
• 3
  Thermal drift countsA 100 mm aluminum part grows about 0.23 mm over a 100 °C rise.

A 3-axis machine holds the part still and moves the tool in X, Y and Z. Every new face needs a new setup, and every setup adds a datum error. On a part with six machined faces and a true position callout of 0.02 mm, those stacked errors eat most of the tolerance before the cutter touches metal.

A simultaneous 5-axis center adds two rotary axes, usually A and B, so the tool can reach the part from almost any angle in one setup. The practical gain is not speed. It is that the positional relationship between features is held by the machine, not by how carefully someone reloaded the vise.

For semiconductor work the second gain is tool access. Deep pockets, undercut sealing grooves and port geometry that a straight tool cannot reach become single-setup features. On a 4,000 mm maximum processing size machine, large chamber frames can be finished without breaking the setup.

The limit is stiffness. A 5-axis trunnion at full tilt is less rigid than a 3-axis table, so heavy roughing in hard steel still belongs on a 3-axis or mill-turn platform. We rough on the rigid machine and finish on the 5-axis one when the part allows it.

• 1
  Use 5-axis whenMultiple faces, angled ports, or true position under 0.03 mm across features.
• 2
  Use 3-axis whenOne or two faces, generous tolerance, heavy material removal in steel.
• 3
  Use mill-turn whenThe part is mostly round with milled flats, slots or cross-holes.

Aluminum 6061-T6 is the default for structural semiconductor parts. It machines cleanly, anodizes predictably and holds ±0.005 mm on a stable setup. Its weakness is stiffness. A thin 6061 plate will chatter, and the answer is often a light finishing pass rather than more clamping force.

Stainless 304 and 316L turn up wherever corrosion resistance or vacuum compatibility matters. Both work-harden, so a rubbing cut ruins the surface. Sharp tools, constant feed and no dwelling in the cut are the rules. 17-4PH (SUS630) is the choice when you need strength plus corrosion resistance, and it can be aged after machining to reduce distortion.

Copper and beryllium copper appear in heat spreaders and RF components because of thermal and electrical conductivity. They are gummy, they burr, and they pick up on the tool. Beryllium copper also needs dust control. Titanium TC4 (Ti-6Al-4V) and Inconel are used for high-temperature fixtures, and both cut slowly with high tool wear.

Plastics behave differently again. PEEK and POM hold tight tolerances if you control heat; PMMA and PC crack at tapped holes if the wrong tap is used. For all of these, the finish callout drives the material decision as much as the strength callout does.

• 1
  Aluminum 6061-T6General structure, ±0.005 mm, best finishing behavior.
• 2
  Stainless 316LCorrosion and vacuum service; control work hardening.
• 3
  Copper C110Thermal paths; expect deburring effort.
• 4
  PEEKInsulating parts; watch heat build-up in deep pockets.

Ra 1.6–3.2 μm is a normal as-machined finish and is fine for most frames and brackets. Sealing surfaces and sliding interfaces usually need Ra 0.8–1.6 μm. Optical and vacuum-critical faces can go to Ra 0.2–0.8 μm, but that finish is fragile and expensive to reach on a large part.

Burrs are the failure mode people underestimate. A 50 μm burr on a gas inlet can shed particles into a process chamber. Deburring by hand is inconsistent on complex geometry, so we plan the toolpath to leave minimal burrs and then use controlled methods: tumbling, bead blasting at low pressure, or a deburring tool on edges only.

Anodizing adds roughly half the coating thickness per surface to the part. A hardcoat at 25 μm per side grows a 10 mm slot to about 9.95 mm. If a tight-tolerance feature will be anodized after machining, the drawing should say so, or the feature will be undersized.

Plating and laser marking follow the same logic. Electroless nickel adds a uniform layer and can be specified as a corrosion barrier. Laser marking needs a minimum character height of 1.5 mm to stay legible after finishing.

• 1
  Name the process on the drawingAnodize, plate and blast all change dimensions.
• 2
  Specify deburring method"Deburr all edges" is not a controllable instruction.
• 3
  Keep marking out of seal areasLaser marks can create local stress risers.

A tolerance you cannot measure is a tolerance you do not have. On a ±0.005 mm callout, the measuring system needs to be at least four times better than the tolerance, so a CMM in a temperature-controlled room is the baseline. Hand calipers will not resolve it, and using them creates arguments rather than data.

GreatLight inspects 100% of parts before shipment, covering raw material check, in-process monitoring and final inspection, with reports available on request. In-process checks catch drift before a batch is finished. On a 200-piece run, catching a thermal shift at part 30 saves the remaining 170.

The qualification rate we work to is 99.99%, and historical late-delivery probability is below 2%. Those numbers are only meaningful because the inspection step is not skipped for small orders. A single prototype still gets measured.

For semiconductor customers, inspection reports often need to travel with the parts into a build record. If that is your workflow, say so at quoting, because it changes how the data is formatted and stored.

• 1
  In-process monitoringCatches drift before the batch is complete.
• 2
  Final inspectionEvery part measured, not sampled.
• 3
  Report formatsTell us the format your build record needs.

Machining is the wrong process when the geometry is genuinely a sheet metal part. A flat cover with a few holes, no sealing face and ±0.2 mm tolerance is cheaper and faster as formed sheet metal. Forcing it onto a mill wastes money and adds a setup that buys nothing.

It is also the wrong answer for a part that will be made in tens of thousands with no design changes. At that volume, die casting or injection molding amortizes tooling and wins on unit cost. Machining stays competitive because there is no tooling and no minimum order quantity, from one prototype to 10,000+ part runs.

There is a middle case worth naming. A part that is machined from billet for the prototype, then cast for production, will not behave identically. Cast material has different porosity and damping. If the prototype has to predict production behavior, keep both versions machined until the design is frozen.

Finally, some tolerances are better solved by design than by process. If a feature needs ±0.002 mm and the assembly can tolerate ±0.01 mm with a shim, change the design. Spending machining time to hold a tolerance the function does not need is the most common waste we see in semiconductor work.

• 1
  Sheet metal whenFlat geometry, no sealing surface, loose tolerance.
• 2
  Casting whenHigh volume and a frozen design.
• 3
  Redesign whenThe tolerance is tighter than the function needs.