Identical ceramic test bars break at very different loads: how do I use the Weibull modulus?

Ceramics fail from the worst flaw in the stressed region: a pore, an inclusion, a large grain or a machining scratch. There is almost no plastic flow to blunt it, so strength follows σ_f = K_Ic / (Y × √(π × a)), where a is the flaw size. With a toughness of about 4 MPa√m for alumina and Y near 1, a strength of 400 MPa corresponds to a flaw of about (4 / 400)² / π = 32 µm, and 280 MPa to about 65 µm.

So your spread of strengths is a spread of largest-flaw sizes between bars, all of them well under a tenth of a millimetre. It is a real property of the batch, not test error.

The Weibull distribution describes this weakest-link behaviour. For bars of one size: P_f = 1 - exp(-(σ / σ_0)^m). σ_0 is the characteristic strength, at which 63.2 percent of the bars have failed, and m is the Weibull modulus, which measures scatter: high m means consistent strength. Technical ceramics typically fall between about 5 and 20.

To find m, sort the strengths, assign each a failure probability P = (i - 0.5) / N for rank i, and plot ln(ln(1 / (1 - P))) against ln(σ). The slope of the fitted line is m and the stress where the ordinate is zero is σ_0. Thirty specimens is the usual minimum for a usable estimate.

Once you have the two parameters, the stress for any target failure probability is σ = σ_0 × (-ln(1 - P_f))^(1/m). Suppose σ_0 = 400 MPa and m = 10. For 1 percent failures: 400 × (0.01005)^0.1 = 400 × 0.631 = 252 MPa. For one failure in a million: 400 × (10^-6)^0.1 = 400 × 0.251 = 100 MPa.

That is the price of scatter: a material with a mean strength near 380 MPa can only be used at about 100 MPa if failures have to be rare. With m = 20 the one-in-a-million stress would be 400 × 0.501 = 200 MPa, so improving consistency is worth as much as improving average strength.

Size matters, and it works against you. A larger stressed volume is more likely to contain a large flaw. For two volumes under the same kind of stress, σ_1 / σ_2 = (V_2 / V_1)^(1/m). With m = 10, a component with ten times the stressed volume of the test bar has 10^(-0.1) = 0.79 of the strength, about 21 percent less, and a hundred times the volume gives 0.63.

The same effect makes test results depend on the method: three-point bending stresses only a small region at the peak level and reports higher strengths than four-point bending, and direct tension gives the lowest. Use an effective volume or area calculation when transferring bar data to a component, and check whether failures start at the surface or inside, since the two flaw populations scale differently.

For critical parts, designers do not rely on statistics alone. Proof testing loads every component to a stress above the service stress; the ones with large flaws break, and the survivors are known to have no flaw above a certain size, which truncates the weak tail of the distribution. It only works if the proof load stresses the part the way service does and if unloading is quick, because flaws in oxide ceramics can grow slowly under sustained stress in a humid atmosphere.

Design also helps: keep ceramics in compression where possible, avoid sharp corners and point contacts, and specify the surface finish, since grinding damage is often the strength-controlling flaw.