Verification & Validation (V&V) in CAE

Great question — and honestly, most engineers don't ask it enough. V&V stands for Verification and Validation, and it's the formal process of building trust in a simulation. Think of it as two very different questions: Verification asks "Am I solving the equations correctly?" — meaning, is my software and mesh giving me the right numerical answer to the math problem I set up? Validation asks "Am I solving the right equations?" — meaning, does my math model actually reflect what happens in the real world?

Exactly. A converged solution is not a validated solution. Classic ASME V&V 10 (for FEA) and V&V 20 (for CFD) formalize this distinction. The AIAA Guide (G-077) defines it the same way. Your 250 MPa might be a perfectly computed answer to a model that uses the wrong material law or boundary conditions — that's a validation failure, not a verification failure.

There are three main tools for code verification. First, the patch test — this checks that a FEM element reproduces a constant strain state exactly on an arbitrary mesh. Any element that passes the patch test converges to the correct solution as the mesh is refined. If your element fails the patch test, throw it away. Second, the Method of Manufactured Solutions (MMS) — you pick an exact answer first (say, a sinusoidal stress field), derive what body forces and boundary conditions would produce it, run the solver, and compare. Third, you do a grid convergence study — keep halving the mesh size and watch the error drop at the expected rate (e.g., quadratic elements should give 4× error reduction for 2× mesh refinement, which is 2nd-order convergence).

Richardson extrapolation is a mathematical trick to estimate the "exact" answer from two or more mesh solutions. If you have a coarse mesh result \(f_{2h}\) and a fine mesh result \(f_h\), and your method is \(p\)-th order accurate, the extrapolated "exact" value is approximately \(f^* \approx f_h + (f_h - f_{2h})/(2^p - 1)\). The Grid Convergence Index (GCI), standardized by Roache, goes further — it gives a confidence band around that extrapolated value. In practice, aerospace and nuclear codes are required to report GCI values.

Just plotting simulation vs. experiment curves and saying "looks good" isn't enough for serious engineering. You need formal validation metrics. The most common are: E-value (the simple relative error at a point), the u-value (normalized error accounting for combined experimental and numerical uncertainties), and for field comparisons, the Theil inequality coefficient or L2 norm of the residual. In fatigue and stochastic problems, you also check that confidence intervals overlap — called a "feature comparison" or "Bayesian model validation."

Great distinction. Uncertainty quantification (UQ) splits uncertainty into two fundamental types. Aleatory uncertainty is irreducible randomness — like the natural scatter in material properties between different batches of steel. You can characterize it statistically but you can't eliminate it. Epistemic uncertainty is reducible ignorance — like not knowing exactly which friction coefficient to use. Getting more test data or a better model reduces epistemic uncertainty. In aerospace certification, you must quantify both and show your simulation predictions remain conservative even at the worst-case uncertainty bounds.

V&V is woven throughout the project, not bolted on at the end. Here's a typical sequence: First, during software selection and model setup, you do code verification (patch tests, order-of-accuracy checks on benchmark problems). Then, before your main production runs, you do a mesh convergence study — this is solution verification. You find the mesh size where results stop changing significantly. Then, once results are in, you compare against available experimental data for model validation. Finally, you quantify residual uncertainty and document everything. In practice, companies often maintain a "V&V matrix" — a table mapping each simulation type to the verification tests and validation experiments that support it.

Yes, there are several, and for aerospace work you'll encounter all of them. ASME V&V 10-2006 covers the guide for verification and validation in computational solid mechanics (FEA). ASME V&V 20-2009 does the same for computational fluid dynamics and heat transfer. AIAA G-077-1998 (updated 2010) is the guide for V&V in computational fluid dynamics specifically. And for NASA-funded work or missions, NASA-STD-7009A is the standard for models and simulations — it defines maturity levels (1–6) for simulation credibility and requires documented evidence at each level.

In aerospace, the simulation evidence package for certification typically includes: (1) a Software Qualification Report showing the solver itself has been verified against benchmark problems, (2) a Model Description Document detailing every modeling assumption and its justification, (3) a Mesh Convergence Report with GCI values for each QoI, (4) a Validation Report comparing simulation to coupon-level and component-level test data, (5) an Uncertainty Quantification Report showing the simulation uncertainty band covers the test scatter, and (6) a Traceability Matrix linking each simulation claim back to supporting evidence. For nuclear, NUREG-2121 requires a validation matrix covering all relevant phenomena, often involving thousands of test cases.