A design office proposes to justify a load path by a finite-element calculation rather than a static test. The authority answers neither yes nor no: it asks which test, already performed, has seen this load case. The answer to that single question decides the schedule, the budget and the acceptance of the file — and it is prepared years before it is asked.
Analysis and test are not two levels of rigour, but two means
The vocabulary of certification establishes no hierarchy between « calculated » and « tested ». It knows means of compliance: for each requirement of the applicable specifications, the applicant proposes the way in which it intends to demonstrate that it meets it, and the authority accepts or rejects this proposal. The families are few and stable: statement of compliance, design review, calculation and analysis, safety assessment, laboratory test, ground test on the aircraft, flight test, design inspection, simulation with crew, equipment qualification. Each carries a code that is found, line by line, in the certification programme.
The decisive point concerns timing. The certification programme, which shows requirement by requirement the means selected, is agreed before the work begins. An accepted means commits both parties; a means proposed after the fact, to make up for a test that went badly or a milestone that has slipped, arrives under the worst possible review conditions. The question « can this test be replaced by a calculation? » is therefore raised upstream, when the programme is negotiated, and not in front of the debris of a specimen.
The binary opposition also needs to be dispelled. The great majority of requirements are demonstrated by analysis substantiated by test: the calculation covers the range of cases, the test anchors the calculation at a few points. The pure test — you load, you look — and pure analysis — you calculate, you conclude — are the two rare extremes of a continuum. What is negotiated in practice is the position of the cursor: how many test points, at what level of assembly, for what domain credited.
Certification doesn’t prefer tests over calculations or vice versa. It’s about proving compliance, and both methods have their place. The key is planning: decide early whether a test or a calculation will be used, and stick to that plan.
For an Airbus A350 wing, finite-element analysis (FEA) is used to model load cases across a wide range of flight conditions. However, a full-scale static test is still required to validate the model at critical load points, such as 150% of the limit load, to ensure the wing meets certification standards.
What is credited is not the calculation: it is the validated model
A calculation demonstrates nothing by itself. What can receive a certification credit is a model whose credibility has been established through two distinct operations that everyday language constantly confuses: verification and validation.
To verify is to make sure that the equations chosen are being solved correctly: convergence with mesh refinement, insensitivity to time step, test cases with an analytical solution, non-regression between solver versions. Verification says nothing about reality; it says that the code does what one believes it does. To validate is to compare the model’s output against a measurement taken on a physical object, under known conditions, with its own uncertainty. A model that is verified but not validated is only a very fast opinion-generating mechanism.
From this follows the notion that governs the whole subject: the validation domain. A model is not valid in the absolute; it is valid over the envelope of parameters where it has met measurement — load levels, temperatures, strain rates, damage states, assembly configurations. Inside it, it interpolates, and the credit is defensible. Outside it, it extrapolates, and extrapolation is not a means of compliance.
The price of this domain depends directly on the physics at play. As long as the response remains elastic and linear, a few points suffice to calibrate a model that will cover a wide envelope. Every nonlinearity sharply narrows its extent: plasticity, post-buckling behaviour, contact and friction, delamination, rupture, combustion. These are exactly the phenomena on which requirements are decided, and where testing takes back the lead.
The corollary is visible on the flight line. What has not been demonstrated does not disappear: it comes back down as a limitation. An envelope not covered becomes a restriction in the flight manual, a placard in the cockpit or cabin, an operating condition in the operational data. Reading a type’s limitations is reading, in negative, the outline of the domain the demonstration actually reached.
The pyramid: how a model buys its credit
The justification of an aircraft structure has long been organised into stacked building blocks: material coupons, elements, structural details, sub-assemblies, components, then the full airframe. The base commonly counts thousands of elementary specimens, the summit a single example.
This pyramid is readily presented as a test plan. It is first a calibration schedule. At each tier, the model predicts, the test measures, the discrepancy is documented, and the model moves up a level with reduced uncertainty and a widened domain. The full-scale test, at the summit, then discovers almost nothing: it confirms. A programme whose summit discovers something is a programme whose base was rushed — and the discovery is paid for with an entire specimen, plus the time to rebuild it.
The material dictates the shape of the pyramid. A wrought metal alloy benefits from well-known behaviour, controlled scatter and a considerable service history: the cursor can go far toward analysis. A composite laminate shifts the cursor the other way, because its performance depends on the process as much as on the design — lay-up, cure cycle, porosity rate, absorbed moisture, post-impact strength. Allowable values there are established statistically on real batches; the effect of hot, humid environments is measured; the effect of damage not detectable to the eye is measured. None of these three quantities comes out of a calculation.
What must be produced for a validation to be credited
An acceptable validation file rests on a handful of pieces, always the same ones. It is their absence, not the quality of the calculation, that sends files back for further review.
- The intended use. A model is never validated « in general »: it is validated for a named output quantity, at a named location, over a named domain. The same mesh may be credited for a global stiffness and refused for a local stress at a notch root.
- Numerical verification. Convergence study, sensitivity to the mesh and to boundary conditions, solver version, the calculation chain and its settings frozen.
- Measured input data. Properties from tests on the batches actually used, manufacturing clearances and tolerances, assembly pre-stresses. A model fed with catalogue values is not validated, it is illustrated.
- The discrepancy metric, announced before the test. Which quantity, at which sensor, with what acceptable discrepancy, and for what decision. A prediction published before the shot has probative value; a model adjusted after the fact to fall back onto the measurement has none, however seductive the overlaid curves may look.
- Uncertainties. Material scatter, measurement uncertainty, residual model bias. The credit is taken at the conservative bound of this set, never at the mean value.
- Configuration control. The model is a definition item: versioned, traced, under identified responsibility, with a written rule stating when it must be revalidated.
The right reflex. Before every test intended to validate a model, record the numerical prediction, the metric and the acceptance threshold in a dated, distributed document. This few-line sheet weighs, under review, more than all the post-processing produced afterwards: it is the only proof that the model predicted instead of commenting.
The two paths, term by term
Honestly comparing the two paths requires taking them over their entire lifetime, not at the quotation stage. The table below lists the criteria that actually come back in progress meetings.

| Criterion | Physical test | Analysis validated by test |
|---|---|---|
| What is proven | the behaviour of the article actually manufactured, processes and tolerances included | the behaviour of the definition, over the domain confronted with measurement |
| Coverage | only the points tested, in the configuration tested | the entire envelope interpolated between calibration points |
| Cost of the first case | high and concentrated: specimen, test rig, instrumentation, test-bench slot | high and diffuse: material data, verification, calibration campaign |
| Cost of the next case | nearly identical: a new specimen is needed | low, as long as one stays within the validated domain |
| Lead time | set by specimen manufacture and rig availability | set by the production of calibration data, then very short |
| What brings the demonstration down | a configuration discrepancy between the specimen and the production series | a change that falls outside the domain, including one deemed to have no effect |
| Typical human error | faulty setup or instrumentation, generally visible at data reduction | a wrong boundary condition: invisible, and the result remains plausible |
| Expected traceability | test report, calibrated measurement chain, exact state of the specimen | model version, input data, metric, discrepancies and justification |
The economic reading is more counter-intuitive than it appears. On an isolated justification, the validated-analysis path rarely costs less: it requires buying the calibration data, hence tests, plus the construction, verification and documentation of the model. It pays for itself on subsequent configurations, and the gap becomes considerable as soon as the family counts a dozen — fit-out variants, weight and centre-of-gravity ranges, installation options, successive derivatives. This is why a decision taken « for that case alone » is almost always wrong: the choice of means is reasoned at the scale of a product family, over a decade. The logic actually reverses cleanly: a product known to have neither derivative nor installation variant has no economic reason to pay in advance for an extended validation domain, and direct testing remains the shortest path there.
What test proves that calculation never will
There is a fundamental reason, rarely stated, for the irreducibility of certain tests: a full-scale test does not only validate a design, it validates an object. The specimen has been drilled, bored, riveted, laid up, cured, bonded, assembled by operators, with real clearances and real torque values. The model, for its part, contains only the definition. Between the two lies everything that manufacturing adds and that no one knows how to put into equations: process scatter. Removing the test means removing the only organised confrontation between what was designed and what comes out of the shop floor.
The second reason is physical. Where the criterion is an event — a rupture, a fire, an impact, human behaviour under stress — calculation describes a scenario, not its outcome. Hence a fairly stable map of the domains in which the cursor stays on the test side.
| Domain | Dominant means | What analysis does not replace |
|---|---|---|
| Static strength up to ultimate load | full-scale airframe test | the proof that the manufactured article withstands the regulatory factor applied to limit load |
| Fatigue and damage tolerance | full-scale fatigue test, then analysis | real crack initiation, on assembly details the mesh does not represent |
| Containment of a failed blade, ingestion | full engine test | the actual trajectory of fragments and the casing’s response to a single event |
| Flammability, heat release, propagation | standardised laboratory tests | the decomposition chemistry of real materials, outside the scope of credited models |
| Seats and impact loads | dynamic tests, analysis for variants | the strength of anchorages and the response of the instrumented dummy |
| Icing | icing wind tunnel and natural-icing flight | the actual shape of accreted ice, on which the rest of the calculation depends |
| Flight-manual performance and handling qualities | flight tests, then a calibrated performance model | the behaviour of the complete aircraft, engines, systems and handling included |
| Time-limited evacuation | demonstration with occupants | the behaviour of a real population in a constrained flow |
This table reads both ways. It states where testing remains mandatory, but above all it states that, everywhere else, validated analysis is already the rule and not the exception: flight-load cases, installation strength, thermal margins, electromagnetic compatibility, most modification justifications have long run on credited models.
Derivatives: where analysis pays off, and where it traps you
The return on investment of a validated model is collected on evolutions. A modification, a derivative, an installation variant is then justified by comparison: it is shown that the definition discrepancy stays within the domain where the model has met measurement, and analysis covers the rest. It is this mechanism that makes bearable the pace of cabin fit-out variants, avionics options and local structural repairs, none of which would on its own justify a test campaign.
The trap lies exactly there, and it does not look like a risk taken knowingly. It looks like a minor change: a different fastener supplier, a different resin batch, a modified machining process, a mastic thickness, a revised torque value, a solver version updated by the IT department. None of these events presents itself as leaving the validation domain, and no process flags it spontaneously: they travel through the chain of procurement, methods or software tooling, not through certification’s.
The trap. A validated model does not lose its credit with a bang, it loses it silently, the day an input parameter changes outside the certification loop. The question to ask before any modification, even one deemed to have no effect, is mechanical: which test in the pyramid would need to be redone if this change had been present on the day that test was conducted? If even one test is concerned, the validation domain has moved and the analysis credit must be re-examined beforehand, not afterward.
What to check in your own file
Three entries suffice to gauge a justification by analysis, and they can be read in a few minutes: the statement of the model’s intended use, the plotted validation domain with the tests bounding it, the discrepancy metric and its issue date compared with that of the test. A file that does not answer these three entries is not a weak file, it is an incomplete one, and it will be sent back on that ground before the results are even opened.
A grey area remains, held by practice more than by the texts: the boundary between a legitimate recalibration — correcting a boundary condition shown to have been wrong — and the accommodating adjustment that makes the curve pass through the measured point. No text draws this boundary; only traceability and the metric’s priority in time hold it. Machine-learning-derived models sharpen the difficulty, because they interpolate remarkably well without exposing their domain of validity — precisely the piece that certification demands.
For anyone preparing to take up a role in a design organisation, or a qualification touching on structural justification, the most rewarding reading order is not that of the regulation: start with the certification programme of a real file, to see how the means are distributed requirement by requirement; continue with a model validation report; finish with the certification specifications, which then become readable. Discussions on the FL330 technical forum and the FL330 training pathways approach the subject from the same angle.
The certification specifications applicable to the type, the certification programme accepted by the authority and the approved documentation take precedence over this article, whose purpose is to shed light on a mechanism, not to substitute for it.
Questions fréquentes
Can a very well validated model exempt from the full-scale static test?
Not for the demonstration of strength up to ultimate load of a new type, however good the model’s quality. A full-scale test does not validate the definition, it validates the object actually manufactured, with its processes, its clearances and its tolerances, which the model does not contain. On a derivative or a local repair, however, analysis calibrated on the original type’s test covers the definition discrepancy, provided that discrepancy stays inside the validated domain.
What is the practical difference between verifying and validating a model?
To verify is to demonstrate that the equations chosen are being solved correctly: convergence with the mesh, insensitivity to time step, test cases with a known solution, non-regression between solver versions. To validate is to compare the model’s output against a measurement made on a physical object, with its uncertainty. A model can be perfectly verified and completely wrong; only validation ties it to reality, and only validation opens a certification credit.
Does a model recalibrated after the test retain probative value?
Much less, and sometimes none at all. A recalibration is only defensible if one demonstrates which hypothesis was wrong and why, independently of the result obtained. Adjusting parameters until the measurement is recovered produces a flattering overlay of curves but proves nothing about the next case. This is why the prediction, the metric and the acceptance threshold must be written down and distributed before the test.
How can one tell that a modification falls outside the validation domain?
By asking the question in reverse: which test in the pyramid would need to be redone if this modification had been present on the day that test was conducted? If at least one calibration test is concerned, the domain has moved. The danger comes from changes that do not travel through the certification loop — fastener supplier, material batch, process, solver version — and that no alert flags spontaneously.
Why does a composite structure require more testing than a metallic structure?
Because its performance depends on the process as much as on the design. Lay-up, cure cycle, porosity, absorbed moisture and post-impact strength change properties from one part to the next, with wider scatter than that of a wrought alloy. Allowable values are therefore established statistically on real batches, and environmental effects, like those of damage not detectable to the eye, are measured rather than calculated.
Can the choice between analysis and test still be discussed once the programme is under way?
It can be discussed, but under very poor conditions. The means of compliance selected for each requirement appears in the certification programme, accepted by the authority before the work begins, and that agreement commits both parties. Proposing a calculation to replace a test already planned, after a technical difficulty or a slipped milestone, amounts to requesting a reopening of the review at the worst point in the schedule.

Leave a Reply
You must be logged in to post a comment.