📚 Course ContentsChapter 3 Process capability indices (Cp, Cpk, Pp, Ppk) for CAE outputs

3. Process capability indices (Cp, Cpk, Pp, Ppk) for CAE outputs

In a CAE process, "capability" answers a specific question: given the natural scatter of a simulation output under stable, in-control conditions, how comfortably does that scatter fit inside the tolerance band the design requires? The four indices Cp, Cpk, Pp, Ppk are the standard vocabulary for that question, but they are routinely misapplied to simulation data because CAE outputs are not interchangeable with machined-part measurements. This section defines each index, states the assumptions they rest on, and shows how to compute and interpret them on a realistic CAE output.

Definitions and the formulas behind them

All four indices share the same skeleton: a tolerance width divided by a spread measure, with Cpk and Ppk additionally penalizing off-centering. Let USL and LSL be the upper and lower specification limits, μ̂ the estimated process mean, and σ̂ the estimated standard deviation. Then:

Cp  = (USL − LSL) / (6σ̂_within)
Cpk = min( (USL − μ̂) / (3σ̂_within), (μ̂ − LSL) / (3σ̂_within) )
Pp  = (USL − LSL) / (6σ̂_overall)
Ppk = min( (USL − μ̂) / (3σ̂_overall), (μ̂ − LSL) / (3σ̂_overall) )

The only structural difference between the C-family and the P-family is the standard deviation in the denominator. Cp/Cpk use a within-subgroup (short-term) σ, typically estimated from the average moving range or pooled subgroup ranges, which reflects the variation a process exhibits when only common causes are present. Pp/Ppk use the overall (long-term) sample standard deviation computed across all observations, which absorbs drift, shift-to-shift differences, and any other between-subgroup variation. In a perfectly stable process the two σ estimates converge and Cp ≈ Pp; the gap between them is itself diagnostic information.

The relationship between the pairs is also worth internalizing: Cp and Pp measure potential capability (what the spread alone would allow if the mean were perfectly centered), while Cpk and Ppk measure actual capability (what the spread and the current centering jointly deliver). Cpk can never exceed Cp, and Ppk can never exceed Pp. A large Cp with a small Cpk is the classic signature of a capable spread that is badly centered.

Assumptions that CAE data frequently violate

The indices are not assumption-free. Before quoting a number, verify the following, because each violation silently distorts the result:

Stable, in-control process. Cp/Cpk presume the within-subgroup σ is a meaningful estimate of common-cause variation. If the CAE output is trending (e.g., mesh convergence not yet reached, solver tolerance loosening across runs), the within-subgroup estimate is contaminated and Cpk is optimistic.
Normally distributed output (or a justified transformation). The 6σ denominator and the 3σ tail logic assume normality. CAE outputs such as maximum stress, fatigue life, or peak deflection are often right-skewed; applying the raw formula to a lognormal life distribution will misstate the tail probability. Transform (e.g., log) or use a non-normal capability method.
Independent observations. Repeated runs that share the same mesh, solver settings, and random seed are not independent replicates; they understate σ and inflate every index.
Specification limits that are real and fixed. USL/LSL must come from design requirements, not from the simulation output itself. Deriving limits from the data (e.g., mean ± 3σ of the same sample) makes the index circular and meaningless.
Sufficient sample size. σ̂ from fewer than roughly 25–30 observations is too noisy to support a capability claim; report a confidence interval, not a point estimate alone.

A common CAE pitfall: treating a single deterministic simulation as one "measurement" and then computing Cp from a handful of design-of-experiments (DOE) runs. If the DOE varies input factors deliberately, the resulting scatter is between-design variation, not process variation, and the index describes the DOE space, not a manufacturing or operational process. State explicitly what population the σ represents.

Worked example on a CAE output

Consider a bracket stiffness study where the output of interest is peak von Mises stress at a fixed load case, with a design limit of USL = 240 MPa and no meaningful lower limit (LSL effectively unbounded, so only the upper-side term governs Cpk/Ppk). Suppose 30 independent runs vary manufacturing tolerance inputs (wall thickness, fillet radius) within their specified ranges, and the observed stresses have a mean of 205 MPa. The within-subgroup σ, estimated from 6 subgroups of 5 runs each, is 9.0 MPa; the overall sample σ is 12.5 MPa. Then:

Cp  = (240 − 0) / (6 × 9.0)  ≈ 4.44   (not meaningful with one-sided spec; use Cpk)
Cpk = (240 − 205) / (3 × 9.0) ≈ 1.30
Pp  = (240 − 0) / (6 × 12.5) ≈ 3.20   (again one-sided; use Ppk)
Ppk = (240 − 205) / (3 × 12.5) ≈ 0.93

The interpretation is instructive. The within-subgroup Cpk of 1.30 suggests the process is marginally capable under stable conditions, but the overall Ppk of 0.93 — below the common 1.33 threshold — reveals that between-subgroup drift (here, batch-to-batch variation in the tolerance inputs) is eroding the margin. The gap between Cpk and Ppk (1.30 vs 0.93) is the actionable signal: reduce the between-batch variation before claiming capability. Note that with a one-sided specification, only Cpk and Ppk are meaningful; quoting Cp or Pp here would be a category error.

Report the index together with its inputs: the σ used (within vs overall), the sample size, the specification source, and whether the output was transformed. A bare "Cpk = 1.30" is not reproducible and cannot be audited.

Reading the numbers: thresholds and their limits

Conventional automotive and general-manufacturing thresholds treat Cpk ≥ 1.33 as "capable," ≥ 1.67 as "highly capable," and < 1.00 as "not capable." These thresholds were calibrated on high-volume discrete-part production and should be applied to CAE outputs with caution. A CAE capability study often has a small sample, a skewed output, and a σ that reflects input-tolerance propagation rather than a physical process. The threshold is a starting reference, not a verdict. More defensible practice is to pair the index with the estimated defect rate (from the normal or transformed tail) and with a confidence interval on Cpk, then judge against the consequence of exceeding the limit.

The figure below summarizes how the four indices relate: the shared tolerance-width numerator, the within-versus-overall σ split, and the centering penalty that separates Cp from Cpk and Pp from Ppk.

Practical guidance for CAE capability reporting

When you present a capability index for a simulation output, make the following explicit so the number is defensible. First, define the population: are the runs independent replicates of a manufacturing/operational process, or a DOE over design factors? The σ must match the population you claim to describe. Second, state which σ was used and how it was estimated (moving range, pooled subgroups, or overall sample). Third, confirm the distributional assumption or apply a transformation, and report the transformed index. Fourth, give the sample size and a confidence interval — a Cpk of 1.30 from 30 runs carries a wide interval and should not be treated as a precise pass/fail. Finally, pair the index with the estimated defect rate and the engineering consequence of exceeding the limit, because a capability number without a risk statement is only half the decision.

Used this way, Cp/Cpk and Pp/Ppk become a compact diagnostic pair rather than a compliance checkbox: the C-family tells you what the process could achieve under stable conditions, the P-family tells you what it actually delivers once real-world drift is included, and the difference between them points directly at where to invest improvement effort.