Correlating FEA to a durability failure — how close is close enough?

Rohan TanPlastics Engineering · 20 yrs

Verified expertise in Plastics Engineering

Rather than treating the 1.6 as an error to be eliminated, calibrate it and keep the physics honest. Because your two omissions both bias toward over-prediction and your prediction is over by 1.6, applying a single calibration factor derived from this test is defensible — provided you apply it only to components in the same alloy, casting process and surface condition, and you document it as an empirical correction rather than folding it silently into the knockdown factor. The failure mode is worrying only if you later use the calibrated model on a machined or differently cast part and the correction no longer applies.

If you want to reduce the uncertainty properly rather than calibrate it, porosity is where the return is. Radiograph or CT the failed part in the failure region and check whether the initiation site is at a pore or at the surface. If it initiated at a subsurface pore, your surface knockdown factor is not the governing mechanism at all and the agreement you are getting is partly coincidental — which matters a great deal when you use the model to assess a design change that alters the surface but not the porosity distribution. That is a single scan and it tells you whether your model is right for the right reason.

Mateo KowalskiAdditive Manufacturing Engineering · 16 yrs

One question worth answering before assessing the model further: what is the scatter on the test itself? A single durability result at 62 percent of required cycles is one sample, and fatigue scatter on die castings is wide — a factor of two to three between the shortest and longest life in a set of nominally identical parts is common, driven by exactly the porosity distribution you have not modelled. If your model is over-predicting by 1.6 against a single test that could itself be a low outlier, the real correlation may be considerably better than you think, or considerably worse.

If more samples are not available, at least check where this part sat in the production population — batch, casting cell, position in the die if that is traceable. It will not give you a statistical distribution but it will tell you whether you are correlating against a representative part or against the worst one from a bad batch, and that changes how much weight the 1.6 deserves. Correlating a model to a single test result and then using the correction on future designs is the trap here, and it is worth naming in the report.