Product Liability Defenses: Exposing Failure Analysis Without Peer Review

A finite element analysis that predicts a stress fracture is not physical evidence. It is a numerical simulation whose output is fully determined by choices the analyst made: material constants, mesh density, element type, boundary constraints, applied loads, and failure criterion. Change any one input and the peak stress the model reports can move by a wide margin. That is why a product liability expert witness challenge aimed at an FEA opinion should target the configuration of the model, not the plausibility of the conclusion.

Under Federal Rule of Evidence 702, as amended effective December 1, 2023, the proponent must show by a preponderance that the opinion reflects a reliable application of a reliable method to the facts of the case. The amendment was written to stop courts from admitting an expert's say-so and leaving reliability to the jury. For simulation testimony that means the analyst has to account for the actual numbers fed into the solver rather than the software's reputation.

Finite element analysis admissibility turns on whether the model is traceable and reproducible. If the material tolerances came from a peer-reviewed allowables source and the mesh was shown to converge, the opinion is defensible. If the tolerances were hand-selected and the mesh was tuned until a crack appeared, the model is an argument wearing the costume of physics.

Every structural FEA rests on a handful of material constants: elastic modulus, Poisson's ratio, yield and ultimate strength, and, for fatigue claims, an S-N or strain-life curve. The decisive question is where those numbers came from and whether they carry a statistical basis.

• Peer-reviewed allowables. Sources such as MMPDS (the Metallic Materials Properties Development and Standardization Handbook, successor to MIL-HDBK-5) and the ASM Handbook publish statistically derived values. MMPDS distinguishes A-basis and B-basis allowables, meaning a defined fraction of the material population is expected to exceed the value at a stated confidence level. Those numbers are reproducible and were vetted outside the litigation.
• Hand-picked constants. An analyst who wants a fracture can quietly substitute a low yield strength, use a single coupon result instead of a population, or pull a typical value and treat it as a guaranteed minimum. Each move shifts the margin between predicted stress and the strength the model compares it against.

Ask which properties are governed by a recognized test method. Tensile inputs should trace to ASTM E8/E8M; axial fatigue inputs to ASTM E466 or an equivalent strain-life protocol. An input with no cited test method and no handbook lineage is an assumption, and Rule 702 does not treat an assumption as data.

The manipulation to watch for is asymmetric conservatism. A defensible model applies the same statistical basis to both the load side and the strength side. A results-driven model uses a worst-case low strength for the part while feeding it an aggressive load, stacking two conservative assumptions to force the stress past the fracture threshold.

FEA discretizes a part into elements. Coarser meshes underreport stress; finer meshes capture gradients more accurately, but only up to a point. A credible analysis includes a mesh convergence study: the analyst refines the mesh in stages and shows that the stress of interest stabilizes toward a value that stops changing as elements shrink. Without that study, the reported stress is an artifact of an arbitrary element size.

The exploit lives at geometric singularities. At a perfectly sharp re-entrant corner, linear elastic theory predicts stress that grows without bound as the mesh is refined. The number never converges. An analyst who wants a fracture can place the peak-stress query at exactly such a node, then refine the local mesh until the reported stress clears the material's strength. The result looks precise and is physically meaningless, because a real part has a finite fillet radius that caps the concentration.

• Confirm a convergence study exists and that the reported stress is the converged value, not the finest-mesh value at a singular point.
• Check whether sharp corners in the model correspond to actual fillet radii on the physical part. A missing fillet is a common way to manufacture a stress riser that does not exist.
• Look at element type and order. Swapping linear for quadratic elements, or shell for solid, changes stiffness and stress recovery. The choice should match the geometry and be stated, not silent.

Boundary conditions define how the modeled part is held and loaded. They are the least visible and most manipulable part of the setup, because a plausible-looking constraint can inject or erase stress anywhere in the model.

• Over-constraint. Fixing more nodes or degrees of freedom than the real mounting provides makes the part artificially stiff and drives up local stress at the constraint. Applying a load through a single node instead of a realistic contact area produces a spike with no physical counterpart.
• Load definition. The magnitude, direction, and distribution of the applied load decide the whole answer. A static load used where the event was an impact ignores dynamic amplification; an impact factor invented without test support inflates it. Either direction can be the thumb on the scale.
• Contact and friction. In assemblies, whether surfaces are bonded, allowed to slide, or given a friction coefficient changes load paths substantially. A bonded contact where the real joint slips can route stress into the fracture location the theory needs.

The through-line is that each assumption should map to a documented physical fact: a mounting drawing, a measured load, a known duty cycle. When an assumption cannot be traced to the record and instead traces to the desired conclusion, that is the analytical gap General Electric Co. v. Joiner authorizes a court to exclude on.

Verification and validation are distinct, and analysts sometimes blur them. Verification asks whether the equations were solved correctly: mesh convergence, numerical error, correct solver settings. Validation asks whether the right equations were solved, by comparing the simulation against physical measurement. ASME V&V 10, the standard for verification and validation in computational solid mechanics, formalizes this separation.

A model that predicts fracture but was never correlated to a physical test is an untested hypothesis. Correlation can take the form of strain-gauge readings on an exemplar part, coupon testing of the actual material, or a component test that reproduces the loading. When the simulated strain field matches measured strain within a stated tolerance, the model earns weight. When there is no comparison at all, the analyst is asking the jury to accept software output as fact.

• Ask for the validation basis. If none exists, the opinion rests on verification alone, which confirms the math ran, not that it describes the failure.
• Ask whether the material in the model was tested, or only assumed from a handbook. A handbook value is a reasonable starting point, but a fracture opinion about a specific part is stronger when the specific material was characterized.
• Ask whether the model reproduced the known failure. A model that cannot recreate the observed fracture origin, surface, and mode under the claimed loads is contradicted by the physical evidence it purports to explain.

Simulation testimony is technical rather than scientific in the narrow sense, so its gatekeeping authority runs through Kumho Tire Co. v. Carmichael, which extended the Daubert reliability inquiry to engineering and other technical experts. The Daubert factors translate cleanly onto FEA.

• Testability and known error rate. Was the model validated against measurement, and is the numerical error bounded by a convergence study? An FEA with neither has no stated error rate.
• Peer review. Do the material allowables and methods trace to peer-reviewed sources such as MMPDS, the ASM Handbook, and recognized ASTM test methods, or were the constants generated for the case?
• Standards controlling the operation. Did the analysis follow a recognized verification and validation framework such as ASME V&V 10, or the analyst's own undocumented practice?
• General acceptance. Are the element choices, failure criterion, and boundary conditions the ones a disinterested analyst in the field would use for this geometry and loading?

Rule 702(d) supplies the sharpest edge: the opinion must reflect a reliable application of the method to the facts of the case. Joiner lets a court exclude when too great an analytical gap separates the data from the opinion. A model built on untraceable inputs and an unconverged mesh is that gap made visible.

An FEA report delivered as a PDF is a conclusion. The reliability lives in the native files, and those are what to compel. A summary that omits the input deck cannot be independently reproduced, and reproducibility is the whole point of Rule 702.

1. Native model and input files. The solver input deck or database, not a rendered report, so your rebuttal expert can rerun it. Include the mesh file and the material cards.
2. The convergence study. The staged mesh refinements and the stress-versus-element-size data that show the result stabilized.
3. Material property provenance. The source for every constant: handbook edition and page, or the test report and specimen data, with the statistical basis identified.
4. Boundary condition and load documentation. The physical drawings, measured loads, or standards each constraint and load is drawn from.
5. Validation records. Any strain-gauge, coupon, or component test used to correlate the model, or a clear statement that none exists.
6. Version history. Prior iterations of the model. A sequence of runs in which the mesh was refined at one corner until the stress crossed the strength threshold is the fingerprint of a result-driven analysis.

The deposition goal is to separate what the analyst measured from what the analyst assumed, and to lock the provenance of each input before trial. Concrete lines, each tied to a mechanism above:

• For every material constant in the model, what is the source, and does it carry an A-basis or B-basis statistical designation? If it came from a single test, how many specimens?
• Did you perform a mesh convergence study, and did the stress you rely on converge, or is it the value at the finest mesh you happened to run?
• Is the location of peak stress at a geometric feature that has a fillet radius on the actual part? What radius did you model, and how did you determine it?
• How are the boundary constraints justified by the physical mounting, and did you test whether relaxing them changes the result?
• Was the model validated against any physical measurement of this part or material? If so, what was the correlation tolerance? If not, why not?
• Can your model reproduce the actual fracture origin and mode observed on the failed part under the loads you assumed?

Answers that repeatedly reduce to analyst judgment without a traceable source are the record you need. They convert a polished simulation into a chain of unverified assumptions, which is the ground on which a court decides whether the opinion reaches the jury at all.

None of this guarantees exclusion of the opinion or a particular outcome at trial. It is deposition-preparation and procurement guidance, not legal advice on a specific matter.