AI Glossary / Statistics and probability / Advanced
Analysis of manufacturing variance
Definition
Analysis of manufacturing variance, or AMV, is a computer-experiment approach that decomposes projected manufacturing-process variation into contributions associated with input factors and their interactions. It evaluates the output distribution implied by a process model and specified input distributions, rather than simply comparing observed group means in a conventional ANOVA.
What it means
A manufacturing simulator may predict a response for any specified set of inputs. Actual production does not hold every input perfectly fixed: material properties, settings and operating conditions vary. AMV asks how that assumed input variation translates into output variation and which factors contribute most.
The source glossary emphasizes distributions for the factors, often Gaussian in the described setting. Those distributions are part of the analysis, not incidental details. Changing their spreads or dependencies can change the contributions even when the simulator remains exactly the same. The term is closely related in purpose to variance-based sensitivity analysis, but terminology varies across sources.
How it works
Define the manufacturing response and a validated process model or suitable emulator. Specify the input distributions and any dependencies using relevant process knowledge and data. These choices determine the production conditions to which the projected output distribution refers.
Propagate the input variation through the model analytically or computationally. Where an appropriate decomposition applies, separate contributions associated with individual factors and combinations of factors. For independent inputs, orthogonal variance decompositions can separate main effects and interactions without double counting; dependent inputs require additional care and a clearly defined attribution method.
Check uncertainty from the simulator, fitted emulator, input-distribution estimates and numerical calculation. Compare scenarios such as tighter input tolerances to see how projected output variation changes. A variance contribution is conditional on the model and input distribution; it is not automatically the amount a real intervention will remove. Changes to production can affect other inputs, dependencies or the response relationship.
Three practical examples
1. A process engineer uses a simulator to estimate whether variation in material thickness or temperature contributes more to the spread of final dimensions.
2. A manufacturing team compares projected output variability under current and tighter tolerances for one input, including its interactions with other factors.
3. An analyst evaluates an emulator's sensitivity results against simulator checks before recommending where additional measurement or process control could be valuable.
Think of it this way
Think of several fluctuating ingredients feeding a recipe. The variation in the finished product depends on how much each ingredient varies and how strongly the recipe responds to it. Two ingredients may also interact, so their combined variation cannot always be understood by considering each alone.
Worked example
Consider the original simulator Y = 100 + 2X + 3Z + XZ. Let X and Z be independent, centered normal input deviations with variances 1 and 4 respectively. The mean output is 100 because the inputs and their product have mean zero.
The three varying terms are uncorrelated under these assumptions. Their variances are Var(2X) = 4, Var(3Z) = 36 and Var(XZ) = E(X squared)E(Z squared) = 4. Total output variance is therefore 44. The individual X contribution is 4/44, about 9.1%; the individual Z contribution is 36/44, about 81.8%; and the interaction contributes another 9.1%.
If Z's standard deviation is halved from 2 to 1 while other assumptions remain unchanged, its variance becomes 1. The output variance becomes 4 + 9 + 1 = 14. This is a projected reduction of 30 variance units, about 68.2%, including changes to both Z's direct and interaction contributions. It is a model calculation, not observed proof that a production intervention will achieve that reduction.
Illustrative case study
Fictional case study: Varianceforge studies inconsistent output dimensions using a process simulator. Its first analysis ranks factors using broad engineering ranges, but those ranges do not resemble everyday production variation. The team revises the input distributions using relevant operating data and checks dependence between factors.
The revised results identify a different priority for investigation. Engineers examine whether tighter control is feasible and run physical checks before changing specifications. The report states the assumed distributions and distinguishes simulator uncertainty from projected manufacturing variation.
This fictional case illustrates why input assumptions matter. It does not establish that the largest variance contribution is always the cheapest or most practical improvement target.
Common misconceptions and mistakes
AMV is not simply another label for a conventional ANOVA significance table. The source meaning emphasizes projected process distributions from computer experiments. It also differs from estimating variance components in a random-effects model, although all these methods discuss variation.
A large model coefficient alone does not imply a large contribution: input spread and interactions also matter. Normal input distributions do not guarantee a normal output distribution when the simulator is nonlinear. Finally, contributions derived under independence should not be applied unchanged when inputs are dependent, because the attribution and covariance structure can change.
Business use
AMV can help prioritize tolerance studies, measurement improvements and process-control investigations. It makes the connection between assumed operating variability and projected output variability explicit, which can be useful when physical experiments are expensive.
Use distributions representative of the intended operating regime and document the model's validated range. Compare improvement scenarios with implementation cost and feasibility rather than ranking contributions alone. Examine tail behavior separately when specification failures matter, because equal variances can accompany different failure probabilities. For AI emulators, validate predictions and sensitivity conclusions in relevant regions; accurate average predictions do not automatically guarantee accurate variance attribution or reliable conclusions about rare outcomes.
People also ask
Does AMV require a computer simulator?
The source term is associated with computer experiments. Its broader variance-attribution purpose can be implemented with an appropriate mathematical or surrogate model.
Are factor contributions fixed properties of the process?
No. They depend on the response model, input distributions and attribution assumptions.
Can an interaction contribute to output variation even when inputs are independent?
Yes. Independence of inputs does not eliminate interaction terms in the response function, as the worked example shows.
Check your understanding
A linear simulator is Y = 5 + 2X, with Var(X) = 3. What output variance is projected if there are no other varying inputs or errors?
Answer: Var(Y) = 2 squared times 3 = 12. The constant five changes the mean, not the variance.
Related terms
- Computer experiment
- uncertainty propagation
- variance-based sensitivity analysis
- interaction
- input distribution
- process capability
- emulator and variance components
- its attribution and other statistical uses of variance
Learn next
vary an input's standard deviation in a simple simulator and recalculate output variance. Then add an interaction and compare how tightening one input changes both its direct contribution and the shared interaction term.
More in Statistics and probability
- Almost sure hypothesis testing - Almost sure hypothesis testing studies sequences of tests whose decisions eventually remain correct with probability one as observations accumulate, under...
- Alternative hypothesis - An alternative hypothesis is the statistical claim considered in contrast to a null hypothesis.
- Analysis of covariance - Analysis of covariance, or ANCOVA, is a regression-based method for comparing groups while including one or more quantitative covariates.
- Analysis of rhythmic variance - Analysis of rhythmic variance, abbreviated ANORVA, is a method proposed by Peter Celec for detecting periodic patterns in biological time series.
- Analysis of similarities - Analysis of similarities, or ANOSIM, is a permutation-based statistical method for comparing predefined groups using ranked pairwise dissimilarities.
- Analysis of variance - Analysis of variance, or ANOVA, is a family of statistical methods that partitions variation in an outcome into components associated with modeled factors...
Sources
Primary sources: NIST/SEMATECH, Glossary, analysis of manufacturing variance: https://www.itl.nist.gov/div898/handbook/glossary.htm
European Commission Joint Research Centre, Sensitivity analysis methods: https://joint-research-centre.ec.europa.eu/sensitivity-analysis-samo/methods_en
European Commission Joint Research Centre, variance-based sensitivity analysis with dependent inputs: https://publications.jrc.ec.europa.eu/repository/handle/JRC75937
Previous: Analysis of covariance Next: Analysis of rhythmic variance