Skip to content
SASemiAgoraSemiconductor knowledge & engineeringOpen lab

METROLOGY / VARIABILITY / UNCERTAINTY

Keep the contract attached to the number.

An uncertainty budget becomes useful only after the quantity, unit, measurement grain, denominator, evaluation method, correlation, and evidence boundary are visible.

observed spreadALEATORICmissing knowledgeEPISTEMICmeasurement systemMEASUREMENTmodel discrepancyMODEL

BUDGET ANATOMY

Four questions come before combination.

The V136 fixtures are original teaching values. They expose bookkeeping and refusal behavior; they do not imitate a fab distribution or transfer a source result.

01

Name the quantity

Edge position, linewidth, local CD, read margin, and a normalized profile response are different quantities. A shared word such as sigma does not make them comparable.

02

Lock grain and denominator

A trace segment, wafer site, cell-condition observation, condition row, AFM-like map, and Gauge R&R study represent different counted units.

03

Separate the class

Observed stochastic spread, missing knowledge, measurement-system effects, and model discrepancy answer different questions and stay visible as separate components.

04

Declare evaluation

Type A and Type B describe how a component is evaluated. They do not rename aleatoric, epistemic, measurement, or model uncertainty.

COMBINE ONLY AFTER COMPATIBILITY

Variance terms plus named covariance

Every component must describe the same output quantity, unit, grain, denominator, and evidence class. An explicit sensitivity coefficient maps each standard uncertainty to that output.

u_c^2 = sum((c_i u_i)^2) + 2 sum(rho_ij c_i u_i c_j u_j)

k = 2 is shown only as a teaching interval. It is not a universal confidence level, coverage statement, acceptance limit, or disposition rule.

ONE DISCIPLINE / DIFFERENT GRAINS

The same reading order travels across domains.

Patterning

LER/LWR and PSD

Sampling pitch and trace length define the visible spatial band. Edge correlation changes linewidth variance.

Open V127
Memory

Hierarchical variation

Condition, lot, wafer, site, replicate, and residual components do not gain independent samples at the same rate.

Open V128
Process

Plasma-etch uncertainty

Condition-dependent spread remains a normalized teaching signal, not a recipe or transferable process window.

Inspect signal
Packaging

Hybrid-bonding metrology

Surface and mechanical characterization requires map identity, aperture, uncertainty, and replacement evidence before readiness claims.

Inspect signal

READING ORDER

Move from definition to a visible stop.

  1. 01Define

    Quantity, unit, sign convention, extraction method, operating condition, grain, and denominator.

  2. 02Decompose

    Aleatoric, epistemic, measurement, and model components, including unknown or zero states explicitly.

  3. 03Correlate

    Named covariance terms only; no hidden independence assumption and no rho outside the declared component pair.

  4. 04Filter

    Sampling pitch, measurement length, PSD band, aperture response, detrend, and finite degrees of freedom.

  5. 05Stop

    Refuse mixed dimensions, cross-case ranking, missing classes, insufficient samples, and production requests.

Evidence stop

No physical calibration, universal uncertainty budget, process recipe, acceptance threshold, qualification, material disposition, silicon correlation, production readiness, or signoff follows from this lesson or its deterministic lab.