# Tolerance Stack-Up Analysis in CAD Assemblies: A Practical Method
Every assembly tolerances compile. A bearing bore, a shaft step, a spacer, a housing face — each holds its own tolerance, and the accumulated error lands on the critical gap at the end of the chain. Plenty of assemblies are designed with each part perfectly reasonable in isolation and unusable in combination. Tolerance stack-up analysis is the discipline of computing that accumulation on purpose, before the first part is machined. This article gives a working method, the worst-case versus statistical distinction, and the mistakes that make stacks lie.
Reading the Chain
A stack-up begins by drawing the chain of dimensions from one fixed datum to the critical feature. The critical feature might be a clearance, an engagement depth, a preload, or a concentricity between two bores on opposite sides of a housing. Work in one direction: start at the datum, list every dimension, and sign each one as adding to or subtracting from the stack. Every dimension in the chain contributes its full tolerance in worst case. If the chain passes through a welded section with its own allowance, that allowance is part of the chain whether you like it or not.
The discipline of writing the chain down by hand, even when a tool can draw it, is what catches the mistakes. Each dimension must be traceable to a drawing or a feature of the CAD model, each sign justified, and each tolerance justified by a process. A chain that relies on a dimension the drawing does not dimension is a chain built on hope. The signature of a solid stack is that every term can be pointed to on a real drawing.
Worst Case Stack
Worst case adds the tolerances in the chain arithmetically. If four parts each hold 0.1 mm, the worst-case stack is up to 0.4 mm regardless of how unlikely that combination is. Worst case is the right tool when failure is catastrophic: a clearance that becomes an interference, a seal that loses its squeeze, a hard stop that cannot be reached. The result is conservative and the design is guaranteed to work if the stack fits. The cost is over-tolerance, meaning parts that carry tighter tolerances than needed to satisfy a combination that will never happen.
Worst case also has a place in design verification. For a one-off mechanism assembled by capable hands, the worst-case stack that fits is the answer; there is no production volume to justify statistical gambling. Run worst case first on every chain, because it costs nothing and anchors the discussion. When worst case fits, the analysis is done and the design is robust. When it does not, the decision moves to statistics or to a design change.
Statistical Stack
When a clearance that narrows only drives performance, not safety, statistics can buy back tolerance. RSS (root sum of squares) and its refinements assume the part variations are independent and roughly normal around their nominal. RSS gives a stack of about two-thirds of the worst case for four equal tolerances, and the 3-sigma result means roughly 0.3% of assemblies fall outside. Statistical stacking is justified when the parts come from stable processes and the consequence of an occasional tight fit is rework, not failure. Never apply statistics to a chain where every part could simultaneously be at its limit — and never use it where a single extreme part is acceptable in isolation but lethal in combination.
The assumptions deserve a harsh look before they are trusted. Are the processes actually independent? Two parts machined on the same fixture in the same batch can carry correlated errors, and correlation destroys the RSS assumption. Is the distribution actually centred and roughly normal? A process with drift or bi-modal behaviour fools the statistic. The safe habit is to treat statistical results as a bonus that justifies a looser nominal, not as a licence to ignore the worst case. Every statistical stack should be reported alongside its worst case so the reader sees both bounds.
The Mean Shift Problem
Real manufacturing is not centred. Machining tends to hold the nominal, but moulding, casting, and bending bias toward one side. A statistical stack computed with centred assumptions hides a systematic shift: every casting comes out 0.05 mm on the thick side, and the gap closes by 4 × 0.05 = 0.2 mm regardless of the RSS maths. The remedy is to bias the nominal values in the stack by the expected mean shift of each process, then run the statistical computation. If you do not know the shift, measure it; a dozen samples from the real process beat a textbook assumption.
Mean shift is the difference between a stack that predicts an occasional tight part and a stack that predicts a systematically tight assembly. When customers report that “every third unit is scrap” or “the first unit of every batch needs a shim,” mean shift is usually the culprit. Feed actual measurement data into the stack model and the prediction catches the shift that the nominal arithmetic misses.
CAD Tools and Their Limits
Modern CAD has tolerance stack modules (and manual spreadsheet methods work too) that build the chain graphically and report worst case and RSS with a few clicks. The tool is convenient, and it carries the same risk as any calculator: garbage in, gospel out. The model is only as good as the dimension scheme and the process assumptions. The most common CAD-specific mistake is letting the model pull dimensions that do not match the physical datum scheme on the drawings. If the drawing dimensions originate from one datum and the model stack reads another, the numbers are wrong even though the software reported them.
A CAD stack module also flatters the user with precision. The software prints four decimal places, so a result that depends on an assumed mean shift and a guessed process distribution looks exact. Keep the report honest by writing the process assumptions next to the numbers, and by treating the tool’s output as a fast calculation to be reviewed, not a verdict to be quoted.
The Datum Scheme Matters More Than the Arithmetic
The stack is only as good as the datums it references. A part whose dimensions reference the wrong face, or a chain that ignores the datum shift between two faces of the same part, computes a fiction. The first step of any stack analysis is to confirm the datum scheme against the functional requirements: which face locates the part, which faces register against mating parts, and which feature is measured by the customer. A stack built on the drawing’s datums is a stack that matches the drawing; a stack built on a guessed scheme is a stack that matches nothing.
A Worked Order of Steps
- Identify the critical feature and its required window
- Draw the dimension chain from the datum
- List every contributing dimension with tolerance and sign
- Run the worst-case stack first; if it fits, stop
- If worst case overflows, check whether statistics and mean shift are justified
- Document process assumptions and centring behaviour
- Confirm the result against the physical part at first article
Stack-Up in Design Review
Tolerance stacks belong in design reviews as a standing agenda item, not a last-minute rescue. A design that is stack-clean on paper earns its approval; a design that has a marginal stack and no mitigation is a known risk that a review should catch before release. When a stack comes back tight, the options are ranked: reorder the dimension scheme, relax a non-critical tolerance, tighten the critical one, or change the design to remove the feature from the chain. The stack analysis is the tool that names which option is cheapest.
The review should also record the residual risk. Every stack analysis ends with an answer and an assumption list; the assumptions are where the next reviewer looks for trouble. Writing “RSS assumed normal, mean shift unknown” on the review record is honest engineering. The part that later misbehaves in the way the unrecorded assumption predicted is a lesson only if the assumption was written down.
Stack-Up in Assembly and the Forgotten Parts
The chain often forgets the parts nobody draws as a proper part: shims, spacers, and bonded joints. A stack that includes a shim that appears only as a note, or a bonded joint whose thickness is absorbed by the adhesive, is a stack with a hidden term. Shims exist precisely to absorb stack error, so locate the shim position in the chain and size its range against the worst case and the mean shift. A bonded joint should either be excluded deliberately (because the adhesive is applied to close the gap) or measured, not silently assumed at zero.
The washer under the bolt head and the gasket between two flanges are the same story. Each carries a thickness tolerance that lands in the chain. When the stack analysis targets the preload of a bolted joint or the compression of a seal, the gasket tolerances are the working variables. The discipline is naming every physical layer in the chain, even the ones that seem too small to matter. Small layers in large stacks add up, and the analysis is the place to discover the total.
Statistical Tools and the Reality of the Shop
Statistical stack-up software and spreadsheet methods produce confident numbers, and the shop floor produces parts with drift, batch effects, and tool wear. The mismatch is where the analysis earns its reputation. A tool philosophy worth following: treat the statistical stack as a range to negotiate, not a precise promise. The moment a customer’s QC argues the stack number to the third decimal, the analysis has been given a certainty it does not have. Keep the report in round, honest bounds and attach the process data that produced them.
The shop-floor reality also includes measurement uncertainty. The part that measures out of the computed stack range may be within the measuring instrument’s own error. When a stack conflict reaches the inspection bench, verify with a higher-accuracy method before declaring the design wrong. A stack dispute that blames the drawing while the instrument drifts is a dispute that wastes a redesign on an illusion.
The First Article as the Verdict
Every stack-up analysis has its moment of truth at first article. The analysis predicted the gap; the physical unit shows it. The verdict classifies the analysis: confirmed, close, or wrong. Feed the verdict back into the process assumptions — the mean shift guessed, the tolerance distribution assumed, the k-factor measured — and update the model. The stack that was right for the first article is trusted for the next; the stack that drifted teaches the correction.
The first-article feedback is also the bridge to the customer. A stack that was analysed, measured, and confirmed on the first unit is a stack the customer can trust in the field. The quality record that shows the predicted range and the measured value side by side is the strongest evidence that the engineering was done, not assumed.
The Supplier and the Generation of Tolerances
The tolerance stack does not begin at the drawing; it begins at the supplier. The stamped bracket’s thickness, the cast housing’s datum shift, and the machined bore’s positional scatter all originate in someone else’s process, and their real distributions arrive with the goods, not with the datasheet. A stack analysis that assumes a symmetric normal distribution for a process that actually drifts with tool wear will publish a range the parts will not respect. When the stack is critical, spend the effort to learn the supplier’s actual capability — from their PPAP data, their historical measurements, or a batch study — and feed that into the analysis instead of the catalogue numbers.
The supplier discipline also names the process capability index that the stack needs. A stack whose slack depends on a ±0.05 dimension will only hold if the supplier actually holds it, and the purchase order should say so. Tolerancing that is analysed but not contracted is tolerancing that arrives on-site as a negotiation. The stack document and the supplier’s capability record together are the real requirement; the drawing alone is a hope.
Conclusion
Tolerance stack-up analysis turns assembly luck into assembly design. The worst-case calculation is the safety net; the statistical method with honest mean shifts is the efficiency tool; the CAD stack module is the convenience layer over both. Run the worst case before the statistics, and put the part measurement at first article next to the prediction. A stack that is calculated on purpose, not discovered on the bench, is what separates assemblies that assemble every time from assemblies that assemble in the hands of a patient fitter.
