Project
Monte Carlo Tolerance Stack-Up Project
Project guide for Monte Carlo tolerance stack-up using worst-case bounds, RSS propagation, yield, assumptions, redesign and validation evidence.
This project produces a tolerance-analysis package for an assembly clearance using worst-case bounds, root-sum-square propagation, Monte Carlo-style probability estimates, yield decision logic, tolerance allocation, and validation evidence.
The project is intentionally not just a formula exercise. A useful tolerance stack-up must define the functional requirement, dimensional chain, datum and sign convention, input evidence, distribution assumptions, capability basis, decision threshold, and inspection plan. The final deliverable should be reviewable by design, manufacturing, quality, suppliers, and operations.
Project Objective
Analyze the axial clearance of a sliding carriage assembly. The clearance must remain large enough to avoid binding and small enough to prevent excessive backlash.
The deliverable must include:
- dimensional chain and sign convention;
- worst-case clearance limits;
- statistical RSS propagation;
- Monte Carlo-equivalent yield estimate;
- sensitivity or variance contribution ranking;
- redesign recommendation;
- validation and production evidence package.
Analysis Boundary and Assumption Register
The project boundary is the complete path from drawing dimensions to functional clearance in an assembled product. It includes the design model, datums, part measurement method, manufacturing process capability, supplier process behavior, selective assembly rules, inspection uncertainty, and production release criteria.
Before running any simulation, the team should record an assumption register:
| Assumption | Why it matters | Required evidence |
|---|---|---|
| stack equation is complete | missing dimensions create false yield confidence | drawing review and assembly teardown |
| signs are correct | wrong sign can reverse sensitivity | datum-chain review with design owner |
| inputs are centered | yield estimate depends strongly on mean location | recent capability or pilot-build data |
| distributions are reasonable | tail risk changes with non-normal data | histogram, probability plot or supplier history |
| inputs are independent | correlations can increase or reduce spread | process-routing and fixture review |
| measurement error is small enough | inspection noise can hide true clearance risk | gage R&R or uncertainty budget |
| production process matches data source | prototype data may not represent production | process readiness and supplier confirmation |
This register is not administrative overhead. It is the difference between a simulation that supports release and a simulation that only decorates a design review. If an assumption has weak evidence, the project should either test it, model it conservatively, or document it as a release risk.
Functional Requirement
The assembled clearance is:
where:
- A is the housing window length;
- B is the carriage length;
- S is the shim thickness;
- C is the functional clearance.
The requirement is:
The required production yield is:
Assume the input tolerances are bilateral drawing tolerances corresponding approximately to \pm3\sigma process spread after process centering has been verified.
Input Data
| Dimension | Nominal | Drawing tolerance | Sign in stack |
|---|---|---|---|
| housing window A | 120.00\ \text{mm} | \pm0.20\ \text{mm} | +1 |
| carriage length B | 119.30\ \text{mm} | \pm0.12\ \text{mm} | -1 |
| shim thickness S | 0.20\ \text{mm} | \pm0.05\ \text{mm} | -1 |
Nominal clearance:
Engineering Comment
The sign convention matters. A larger housing window increases clearance, while a larger carriage or shim decreases clearance. Many stack-up errors come from losing this directionality or mixing drawing datums with assembly datums.
Functional Failure Modes
The lower and upper limits represent different failures. A clearance below (0.30\ \text{mm}) risks binding, high friction, assembly rejection, wear, thermal seizure or field complaints under contamination. A clearance above (0.80\ \text{mm}) risks backlash, rattle, positioning error, impact loading or perceived quality loss.
Because the failure modes differ, the tail risks should be reported separately. A combined yield number can hide which side of the requirement is driving the decision. In this project, the lower tail is the initial problem, so a redesign that only reduces total spread may be less effective than a redesign that moves the mean clearance.
Step 1: Worst-Case Stack-Up
Minimum clearance occurs when A is low, B is high, and S is high:
Maximum clearance occurs when A is high, B is low, and S is low:
The worst-case range is:
This fails the requirement because:
and:
Engineering Comment
Worst-case analysis is conservative because it assumes all dimensions hit their harmful limits at the same time. That may be appropriate for safety-critical, non-adjustable, or low-volume assemblies. For a high-volume adjustable assembly, a statistical analysis can be justified if the process distributions, centering, independence, and inspection plan are credible.
Step 2: Convert Tolerances to Standard Deviations
Using the stated \pm3\sigma assumption:
For independent inputs in a linear stack:
Therefore:
Engineering Comment
This is the same variance propagation behind many RSS tolerance screens. It is valid here because the model is linear and the inputs are assumed independent. If dimensions share the same machine setup, fixture, operator, thermal condition, or supplier lot, the independence assumption must be challenged.
Correlation Check
Independence is often the weakest assumption in tolerance analysis. If A and B are produced by related fixtures, paired machining offsets, supplier-lot drift, thermal expansion, or common inspection compensation, their errors may be correlated.
For a linear stack, correlated inputs require covariance terms. If (C=A-B-S), the variance is:
The signs depend on the stack equation. A positive correlation between A and B can reduce clearance variation because the dimensions move together in opposite signed terms. A negative correlation can increase variation. The project should not assume that correlation is always harmful or always helpful; it should model the actual process relationship.
If correlation data are unavailable, the release review should include a sensitivity case such as (\rho=\pm0.3) for plausible shared-process pairs. If the yield decision changes under a realistic correlation case, more production data are needed before release.
Step 3: Estimate Yield from Z-Scores
Lower-limit z-score:
Upper-limit z-score:
For a normal approximation:
and:
Estimated yield:
or:
This fails the required:
Engineering Comment
The nominal clearance is centered enough to avoid the upper limit most of the time, but the lower clearance tail is too close to the binding limit. A nominal design that “looks centered” can still miss a production-yield target when the process spread is large relative to the margin.
Tail-Specific Decision Screen
The project should report:
and:
separately from total yield. This prevents a misleading statement such as “99.4 percent yield” when almost all failures are binding failures. Manufacturing and design actions differ by tail: lower-tail risk may call for mean shift, reduced carriage length, thinner shim, selective assembly or tighter housing control; upper-tail risk may call for the opposite.
Step 4: Monte Carlo Simulation Plan
A Monte Carlo simulation would sample:
and compute:
for k=1,\ldots,N.
For each sample:
Estimated yield:
Use at least:
for the review estimate. The expected sampling standard error near p=0.994 is:
or about:
Engineering Comment
Monte Carlo is not needed mathematically for this linear normal example, but it is useful as a project workflow. The same simulation structure can later handle nonlinear geometry, clipped distributions, supplier mixtures, selective assembly, correlations, inspection uncertainty, and non-normal data.
Simulation Quality Controls
The simulation plan should define:
- random seed policy for reproducibility;
- sample count and convergence metric;
- input distribution source for each dimension;
- treatment of drawing limits: normal, truncated normal, empirical, capability-based or supplier-lot mixture;
- correlation model and covariance evidence;
- measurement error model, if inspection noise is not negligible;
- output fields: yield, lower-tail risk, upper-tail risk, mean, standard deviation and percentile clearance values;
- acceptance rule before the simulation is used for release.
For a yield target of (99.5%), the simulation is judging rare tail events. A small sample count can understate risk by chance. The team should repeat the run or increase (N) until the estimated yield and tail probabilities are stable enough for the decision.
Non-Normal and Clipped Inputs
Real manufacturing data may not be normal. Tool wear, process adjustment, supplier sorting, inspection screening, rework, cavity-to-cavity differences and lot changes can create skewed, clipped or multimodal distributions. A normal model with (\pm3\sigma) tolerances can be optimistic if the tails are heavier than expected or if the mean drifts after setup changes.
When empirical data exist, the simulation can sample from measured distributions or fit a distribution justified by process knowledge. When data are limited, a conservative comparison should show whether the decision survives plausible non-normal behavior. A robust project does not claim precision that the input evidence cannot support.
Step 5: Rank Variance Contributions
Variance contributions are:
Total:
Contribution fractions:
Engineering Comment
The housing window dominates variation. Tightening the shim tolerance alone will not be the most efficient way to reduce spread. However, changing shim nominal thickness can shift the mean clearance, which may be enough when the main issue is one tail rather than total spread.
Tolerance Allocation Interpretation
Variance contribution is not the same as improvement priority. The housing window dominates spread, but reducing housing variation may require tooling investment, supplier renegotiation or slower machining. The shim has small variance contribution, yet its nominal value is an efficient mean-shift lever.
The project should therefore separate:
- spread-reduction levers: tighter process control, better fixturing, improved supplier capability;
- mean-shift levers: nominal shim, nominal carriage length, design offset;
- selection levers: measured shims, part pairing, rework policy;
- validation levers: inspection plan, pilot build and field feedback.
The recommended change should be chosen by yield improvement per cost, manufacturability risk and failure-mode severity, not only by the largest variance fraction.
Step 6: Redesign the Shim Strategy
Use a thinner nominal shim:
and tighten its process tolerance to:
New mean clearance:
New shim standard deviation:
New clearance standard deviation:
Lower z-score:
Upper z-score:
Normal approximation:
Estimated yield:
or:
This meets:
Engineering Comment
The redesign works mainly by moving the mean away from the lower binding limit while staying far enough from the upper backlash limit. The small shim tolerance improvement helps, but the mean shift is the dominant decision.
Redesign Risks
The thinner shim strategy should be checked against assembly handling, supplier availability, minimum practical thickness, deformation during installation, material flatness, burrs, coating thickness, wear-in and thermal expansion. A tolerance analysis that changes a part nominal can create new manufacturing or durability risks outside the clearance equation.
If shim selection is manual, the project must also check whether operators can identify shim sizes reliably and whether the inspection method can separate adjacent shim classes. If the shim is supplied in lots, supplier lot-to-lot drift should be included in the validation plan.
Step 7: Define the Validation Package
The tolerance-analysis package should include:
- released drawing chain with datum references and sign convention;
- process capability evidence for A, B, and S;
- measurement-system analysis for the clearance and contributing dimensions;
- histogram or empirical distribution for each input dimension;
- correlation check for dimensions made on the same fixture or machine setup;
- Monte Carlo model version, random seed policy, sample count, and convergence check;
- estimated yield, lower-tail risk, upper-tail risk, and decision threshold;
- pilot-build clearance measurements;
- nonconformance plan for binding and excessive backlash;
- change-control record for shim nominal thickness and supplier capability.
Pilot-Build Release Gate
The pilot build should not only measure average clearance. It should demonstrate that the production process can reproduce the assumptions used in the analysis.
A practical release gate includes:
| Evidence item | Release question |
|---|---|
| measured A, B and S distributions | are input means and spreads inside the model envelope? |
| clearance distribution | does assembled clearance meet lower and upper functional limits? |
| gage R&R or measurement uncertainty | can inspection distinguish true failures from measurement noise? |
| correlation check | do shared processes change stack variation? |
| supplier lot review | does one lot dominate the simulated capability? |
| assembly trial | do thin shims create handling, seating or durability issues? |
| nonconformance disposition | what happens to binding or backlash units? |
The release decision should be held if pilot data show mean drift, distribution shape mismatch, unexpected correlation, measurement-system weakness or a new assembly failure mode. A simulation is not a substitute for evidence that the production process matches the simulated process.
Final Decision
The defensible engineering decision was:
Release the revised shim strategy after pilot-build measurements confirm process centering, input distributions, measurement repeatability, and clearance yield above the 99.5 percent requirement.
The main lesson is that tolerance analysis is an uncertainty and decision problem, not only a dimensional arithmetic problem. Worst-case, RSS, and Monte Carlo methods answer different questions. A strong project explains which method is appropriate, what assumptions support it, and what evidence will keep the production process within the functional clearance requirement.