SPC, Cpk and Gauge R&R for 3D Printed Parts — What Actually Works | Precise3D

Why Machining SPC Rules Do Not Transfer to a Print Cell

Run a conventional SPC programme on a print cell and it will look healthy right up to the moment a customer rejects a shipment. The reason is that the standard rules assume a process whose variation is short-term noise around a stable mean. A print cell violates that assumption in four specific ways, and each one corrupts a different control chart.

Drift within a build. Chamber temperature, and therefore part shrinkage, changes as the build progresses. The first layer and the four-hundredth layer are not produced under the same conditions, so they do not belong in the same subgroup.

Material state. Filament moisture content rises over a week of open storage and falls after a drying cycle. Extrusion quality tracks it, which means an assignable cause enters the process on a schedule nobody logs.

Consumable wear. Nozzle wear is monotonic. It moves a dimension steadily in one direction across a period measured in kilograms, not hours, so a chart reacting to it as special cause will fire continuously.

Low volume per characteristic. A print cell making eight different parts on rotation may produce twenty data points per part per week. Text-book x-bar charts with five-piece subgroups cannot be filled honestly at that rate.

The practical conclusion is that SPC is not wrong for additive, but a programme copied from a machining cell is. The three things that must be redefined are which characteristic you control, how you form a subgroup, and how you attribute your measurement variation. Quality planning that precedes all of this — the FMEA and control plan — is covered in our quality planning guide, and this page picks up where the control plan ends.

Choosing the Control Characteristic Before Choosing the Chart

The highest-leverage decision in a print cell's SPC programme is not the chart type. It is which characteristic you put on the chart at all. Control five characteristics badly and you will generate data nobody acts on; control one that matters and the programme pays for itself.

Rank the candidates in this order. First, what the customer's assembly actually touches: a hole that receives a bearing, a face that seals, a boss that takes a screw. If a dimensional feature mates with anything in the customer's product, it is first in line. Second, what is visible on the assembled product, because a cosmetic defect on a visible face is a rejection even though it is dimensionally irrelevant. Third, a free-state dimension that tracks the process — often overall length or wall thickness, chosen because it moves early when the process drifts, which makes it a good indicator even though the customer never measures it. Fourth, part weight, which is the cheapest and most sensitive proxy for extrusion consistency and is badly underused in this industry.

The trap to avoid is controlling a dimension because it appears on the drawing in bold. Challenge dimensions are the ones the designer was unsure about; they are frequently the ones the customer will never inspect, and putting them under control consumes the budget that should have gone to the mating features.

Rational Subgrouping on a Print That Runs Fourteen Hours

A rational subgroup is a group of measurements taken under conditions where only common cause varies. In machining, five consecutive parts off a machine qualify, because the machine is in the same state for all five. In a print cell that logic inverts: five consecutive parts come off the same build, so they share the same chamber temperature, the same material spool, and the same nozzle state. They are not five independent observations of the process. They are one observation of the process, sampled five times.

Put those five into an x-bar chart and the range bar will be small — so small that the control limits collapse and the chart fires on every normal between-build shift. The operator then spends the week chasing common-cause variation and eventually ignores the chart, which is the worst outcome available.

The correction is to subgroup by plate, not by part. Take one measurement per plate at a fixed part position, and let the subgroup be a sequence of plates. Within-plate variation then becomes visible as its own statistic — the range across parts within a plate — which is exactly the number that tells you whether the process is uniform across the build volume. Between-plate range tells you whether the process is stable from run to run. Both are useful; neither is available from the naive chart.

What belongs to within-plate variation

Within-plate variation carries the build-volume signature: thermal gradient from the heated bed upward, flow-front differences by position, and the fan or chamber airflow pattern. A part in the corner of the plate is not made under the same conditions as a part in the centre, and pretending otherwise hides the single most actionable signal a print cell generates.

What belongs to between-plate variation

Between-plate variation carries the consumable signature: nozzle wear, filament moisture, ambient conditions, and the state of the machine after a maintenance event. This is the variation that responds to scheduled interventions, which is why it needs its own chart rather than being blended into the same one.

Gauge R&R: Separating Printer Repeatability From Gauge Error

The most common way a print cell misreads its own capability is by blaming the gauge for variation the printer created. A conventional Gauge R&R study measures how much of total observed variation comes from the measurement system, and if the answer is above thirty percent the measurement system is declared unfit. On a print cell this conclusion is frequently wrong, because the study has silently bundled the printer's within-part repeatability into the gauge's error term.

The physical fact that creates the problem is that an additively manufactured surface is not smooth. A calliper jaw sits on layer texture and on support scar, so repeated measurement of the same feature by the same operator at the same point varies by an amount that has nothing to do with the gauge. If that variation is attributed to the gauge, the gauge looks unusable and the real process problem is never addressed.

The fix is a two-stage study. First, run a nested Gauge R&R in which each operator measures three different parts, which is the standard structure, but record a separate repeatability trial in which one operator measures the identical point on one part ten times without removing the part. The spread of that ten-measurement trial is the true gauge-plus-surface repeatability. Subtract it from the study's reported repeatability and the difference is the printer's contribution, made visible rather than hidden.

Two further rules apply and are routinely broken. The parts chosen for the study must span at least twice the tolerance band, otherwise the study cannot distinguish a measuring problem from an adequately uniform process — and on a print cell producing good parts, every part is close to every other, so the study needs deliberately varied samples rather than whatever came off the plate. And the gauge resolution must be at least a tenth of the tolerance, which on a printed feature with a wide tolerance is usually easy and on a tight metal-insert feature is usually not.

Realistic Capability Targets for As-Printed Features

Quoting a Cpk of 1.33 for an as-printed dimension is the most reliable way to fail a customer audit on paper before anything is built. The 1.33 figure is the machining and injection-moulding convention, and it assumes a stable process with a short-term distribution that is well behaved. A print cell does not clear the same bar on an as-printed feature, and promising that it will is what turns a healthy process into a paperwork failure.

What is achievable depends almost entirely on how much of the feature is cleaned up after printing. A raw as-printed dimension carries layer thickness, thermal shrinkage, and support-surface roughness, and its capability is limited by the total spread of those three. A feature that is machined after printing, or printed oversize and finished to size, sheds most of that spread because the finishing operation overwrites it. A resin-cast feature made in a printed mould inherits the mould's smoothness rather than the print's layer texture. And a feature on a support side is the worst case in every direction.

The practical guidance is to negotiate capability per feature class rather than per part. A part with an as-printed tolerance and a machined datum is not one capability; it is two, and quoting a single number for it invites an argument at delivery.

The translation matters when you talk to a customer. A Cpk of 1.33 corresponds to roughly sixty-six parts per million outside the limits in a stable normal process; a Cpk of 1.00 corresponds to about two thousand seven hundred. For a lifetime of two thousand parts, that difference is the difference between expecting no defect and expecting five. Making that arithmetic explicit in the quotation is far more defensible than promising a number the process cannot deliver, and it is the same discipline that sits behind the receiving inspection regime described in our incoming QC guide.

Control Chart Selection: Which Chart Survives an AM Signature

With the characteristic chosen and the subgrouping corrected, the chart type is almost a formality — but only if you pick from the right three.

X-bar and R charts remain the right choice for a stable high-volume cell where each plate carries enough parts to form a genuine subgroup and the process is not drifting on the timescale of a week. The caution is that the R chart must be read as a build-volume uniformity indicator, not as process noise.

Individuals and moving-range charts (I-MR) are the practical default for a low-volume cell making several different parts, because they need only one measurement per build. They are far less sensitive to small mean shifts than x-bar charts, which on a print cell is usually an acceptable trade: a mean shift large enough to matter on an as-printed feature tends to be large enough for I-MR to catch it.

EWMA charts earn their place exactly where the print cell's real problem lives: slow, monotonic drift from nozzle wear or material moisture. An EWMA chart with a weighting between 0.1 and 0.3 detects a drift of about one sigma several builds earlier than a Shewhart chart does, which on a consumable-wear signature is the difference between catching tooling degradation and discovering it in a reject bin.

Defect-based cells should use a p-chart or u-chart rather than a variables chart, because when the rejection criterion is visual and pass/fail, coding it into a numeric dimension throws away the information. Many print cells find that a p-chart on first-pass yield, combined with one I-MR chart on the assembly mating feature, outperforms a full battery of variables charts nobody reads.

The First Thirty Plates: A Commissioning Sequence for a Controlled Cell

A control programme that is switched on all at once produces noise. The following sequence brings a print cell under statistical control in the order that keeps each step interpretable, at a cost of about thirty plates.

Plates one to five: establish the baseline. Measure the chosen assembly mating feature on one part per plate and compute a plain mean and standard deviation. No charts yet, because there are not enough points to set honest limits.

Plates six to twelve: quantify measurement error. Run the nested Gauge R&R and the ten-measurement repeatability trial. Do this before setting any control limits, because limits computed from an unmeasured system will be wrong by exactly the amount of gauge error.

Plates thirteen to twenty: set limits and verify no false alarms. Compute limits from the baseline and run without intervention. If the chart fires more than twice in eight plates, the limits are too tight for the real process and the subgrouping is probably still wrong.

Plates twenty-one to twenty-six: introduce capability measurement. Compute Cpk per feature class, not per part, and compare against the table above rather than against a machining convention.

Plates twenty-seven to thirty: audit the control plan. Whatever the plan says to do on an out-of-control point must now be tested once deliberately. A control plan whose reaction column has never been executed is a document, not a system, and it is the specific gap that the cost of poor quality work and the FMEA control plan are both designed to close.

The measurement work itself is only as good as the methods behind it. The instruments, datums and inspection strategies used to generate these numbers are covered in our part metrology guide, automated pass-fail limits in the vision inspection guide, and the finishing operations that shift a feature class from as-printed to machined in the tolerance compensation guide.

← Back to Blog