🧮 1. Assemblies Do Not Care About Your Nominal Dimension
Every dimension on a drawing is a wish, and every part is a compromise. The machined housing sits 0.02 millimetre past its nominal size, the shaft comes in 0.01 millimetre under, the bearing bore drifts with temperature, and the assembled mechanism still has to work. Tolerance stack-up is the discipline of adding up all these individual compromises along a line of fit and predicting, before the parts are ever made, whether the assembly will fit, move, and function.
Designers who skip the stack analysis discover the assembly problem at first fit-up, when the wrong ends of the tolerances arrive together and the mechanism binds, leaks, or rattles. The fix at that point is scrap, rework, or a field change that was never quoted. The fix upstream is a stack-up study that costs an afternoon at the drawing board. This article builds the two dominant stack methods, worst case and statistical, shows how to read a loop diagram, and connects the analysis to the GD&T system that supplies its dimensions.
🔄 2. The Loop Diagram: Walking Around the Fit
Any stack-up begins by drawing the loop: a closed path that starts at one feature, travels through the chain of parts that form the fit, and returns to the start. The loop diagram is the geometry of the stack, and getting it right is worth more than the arithmetic that follows. Start from the functional requirement, for example the clearance between a pin and the far wall of an assembly, and walk from feature to feature through each contributing part, assigning every dimension a positive or negative sign according to its direction along the loop.
The rule of the loop is that the sum of all signed dimensions must close exactly on the nominal, and the sum of all the signed tolerances gives the total stack. A missing dimension in the loop is the classic error: the stack comes out suspiciously tight, the first assembly binds, and the re-check reveals the neglected shim or bearing race that was never entered into the sum. The loop diagram, drawn as a sketch or a spreadsheet row by row, is the audit trail that keeps every contributor visible.
The dimensions in the loop must be independent. Two dimensions that both derive from the same machining setup share their error, and double-counting that shared error exaggerates the stack. The analyst must also know the sign convention of each datum, because a stack that reads the same side of every part is an exercise in self-deception.
⚠️ 3. Worst-Case Stack: The Honest, Brutal Assumption
The worst-case stack adds every tolerance at its extreme, assuming all parts at their worst simultaneously. For a chain of dimensions, the worst-case stack tolerance is simply the sum of the absolute values of the individual tolerances. The method guarantees that if every part meets its drawing tolerance, the assembly will always fit, with zero statistical sophistry. That guarantee is the strength and the cost of the method.
The cost shows up in numbers. A five-part chain, each holding plus or minus 0.05 millimetre, produces a worst-case stack of plus or minus 0.25 millimetre, a clearance that may be absurdly generous for the function. Worst case is therefore mandatory where a failure to fit is catastrophic, such as a hard interference in a safety latch or a seal that must never leak, but it is economically brutal when many contributors feed an unconstrained clearance.
Worst case also sets the tolerance arithmetic for assemblies whose parts are assembled by select fit or whose failure mode has no statistical forgiveness. The engineer who quotes the worst-case number first, then asks whether the assembly can tolerate that worst case, has a defensible baseline for every loosening decision that follows.
📊 4. RSS and Statistical Stack: The Real Distribution of Parts
Real manufacturing processes do not produce every part at its tolerance limit. A well-controlled process centers on the nominal and spreads parts in a bell curve, so the chance of every dimension in a chain arriving at its extreme simultaneously is vanishingly small. The root-sum-square method captures this reality: the RSS stack tolerance is the square root of the sum of the squares of the individual tolerances, under the assumption that the contributors are independent and normally distributed.
The arithmetic difference is large and meaningful. The five-part chain from the previous section stacks to plus or minus 0.25 millimetre worst case, but only to about plus or minus 0.11 millimetre by RSS, less than half. That difference is the design budget the method releases when the process is predictable, and it is precisely what makes RSS the workhorse of stack-up analysis for production assemblies.
RSS carries assumptions that must be checked before it is trusted. The contributors must be independent, so shared machining datums must be handled with care. The distributions must be approximately normal and centred, which is only true when processes are capable and centered. And the target, often a three-sigma or six-sigma band, must match the assembly quality target. Where those assumptions fail, the honest answer is a Monte Carlo simulation with measured process data, not a faith-based extension of RSS.
| Method | Stack tolerance | Assumption | When to use |
|---|---|---|---|
| Worst case | Sum of |tolerances| | All parts at extreme | Fit is critical or cheap chains |
| RSS | Sqrt of sum of squares | Independent, normal | Production, capable processes |
| Monte Carlo | Simulated distribution | Real measured data | Complex or uncontrolled chains |
Monte Carlo simulation deserves its seat at the table whenever the geometry is complex, the parts share setups, or the distributions are skewed. It replaces the closed-form formulas with thousands of random assemblies and reports the actual probability of a fit failure, which is the number the business case actually needs.
🎯 5. GD&T in the Stack: Position, Datums, and the Shift That Hides
Geometric dimensioning and tolerancing changes the stack in a way that surprises engineers trained on linear tolerances. A positional tolerance on a hole, expressed as a diameter in the feature control frame, caps the axis of the feature within a tolerance zone anywhere along its height, and it converts a two-sided linear box into a round zone that the stack can use more honestly. The datum reference frame defines which surfaces, holes, or edges anchor the measurement, and the datum precedence transfers the stack from the drawing to the physical fixture.
The hidden player in GD&T stacks is datum shift, which appears when the primary datum feature has additional tolerance. A shaft located by a sloppy datum than the datum feature itself, allowing the part to shift within the datum feature tolerance. Datum shift can be a friend who absorbs misalignment or an enemy who widens a stack that the analyst forgot to include, and the discipline is to declare it explicitly in the loop rather than let it appear at first assembly.
A positional tolerance also interacts with the tooling that makes the holes. A hole drilled to a positional tolerance relative to a datum transfers that datum to the drill bush, which is precisely the chain the fixture designer built in the previous article. The stack analyst therefore reads the drawing, follows the datum reference frames into the fixtures, and audits the reality of the part against the ideal of the print.
🔧 6. Fixing a Stack That Does Not Close
When a stack analysis reveals an assembly that cannot fit, the engineer has a ladder of remedies, each with its own cost. The cheapest is to relax the least critical contributor: the dimension whose tolerance can open without hurting function. The next is to tighten the few contributors that dominate the stack, because the RSS method makes the largest tolerances the loudest voices in the sum of squares. Re-datuming the drawing to a common datum can cut the chain length, and redesigning to a self-locating feature, such as a dowel or a register step, removes a dimension entirely from the loop.
Selective assembly, sorting parts into matched bins, closes tight stacks without tightening individual processes, at the price of inventory and administration. And the final, often best remedy is to change the design so the stack no longer matters, for example a floating fastener with clearance, or a shim adjusted in assembly. The engineer who lists the remedies before choosing one keeps the decision an engineering choice rather than a firefight.
📝 7. Worked Example: A Bearing Housing Clearance
Consider a bearing housing whose flange is bolted to a plate, with a dowelled bore centre that must align within 0.15 millimetre of the mating plate bore. The chain runs from the housing flange face, through the dowel pin, through the plate, and back to the mating bore. The loop carries the flange face to bore position tolerance, the pin clearance, and the plate bore position tolerance, four independent contributors.
Worst case, the stack sums to about 0.30 millimetre, which fails the 0.15 millimetre requirement outright. Switching the same contributors to RSS, with three-sigma distributions, brings the stack to about 0.15 millimetre, marginally acceptable. Loosening the least-critical contributor and tightening the dowel fit closes the loop at a comfortable margin. The final assembly uses a light press-fit dowel, whose clearance contributes almost nothing, and the RSS stack lands near 0.10 millimetre, inside the requirement with a statistical margin that the production records will eventually confirm.
✅ 8. Stack-Up Checklist
Draw the loop first and list every contributor with its sign before doing any arithmetic. Confirm that the dimensions are independent, or re-datum to remove shared error. Run worst case to confirm the floor of the design, then RSS or Monte Carlo to find the realistic operating margin. Read the GD&T datum reference frames into the loop, and declare datum shift explicitly. Identify the dominant tolerances by their contribution to the sum of squares, and loosen or tighten precisely those. Check the process capability of the suppliers against the distributions the analysis assumed. Finally, verify the first ten assemblies with real measurements and compare them to the predicted distribution, because the stack analysis is only as good as the process data that feed it.
🔚 9. Conclusion
Tolerance stack-up analysis is the accountant of the design process: it counts every micrometre that every part contributes to the fit, and it forces the designer to confront the arithmetic before the parts exist. Worst case reveals the brutal floor, RSS reveals the realistic margin, and GD&T reveals the true geometry that feeds both. Perform the loop early, revisit it when the drawing changes, and verify it on the first production batch, and the assembly that used to fight at first fit-up becomes the assembly that simply closes.
💻 10. Software Tools and the Discipline of Documentation
Tolerance stack-up software has evolved from spreadsheet templates into dedicated packages that read the datum reference frames, apply the GD&T rules, and compute worst case, RSS, and Monte Carlo results in one run. The software removes the arithmetic drudgery, but it does not remove the analysis. A tool that computes automatically will happily compute a wrong loop that the user sketched incorrectly, so the loop diagram and the list of contributors remain the human contribution that no package replaces.
Documentation is the unsung deliverable. A stack-up study without a dated record of the loop, the tolerances, the method, and the assumptions is a private opinion; the same study written into the quality file is a design record that supports the tolerance decision when the supplier pushes back or the field reports a fit problem. Record the version of the drawing, the process distributions assumed, and the result, and the stack analysis becomes the shared engineering language that keeps the whole supply chain honest.
