The Correct Way to Interpret FEM/CFD Analysis Results — From the Pitfalls of Contour Plots to Report Preparation
Why "How You Read Results" Decides Everything
The Contour-Plot Trap — Auto-Scaling
Isn't looking at the contour plot enough? When the colours come out nicely I catch myself thinking "done!"
That is the single most common trap in practice. When the colour bar's maximum is auto-scaled, it gets dragged up by a stress singularity, the whole model turns blue, and the information you actually wanted is buried. A sharp re-entrant corner at a bolt-hole edge, for example, produces stress that keeps climbing without bound every time you refine the mesh — a stress singularity. Auto-scaling then hands the top of the colour bar to that singular value, so the "stress distribution over the whole flange face" you came to look at collapses into one uniform colour. Checking where the colour-bar maximum occurs and why is step one of reading any contour plot.
The Right Order — Deformed Shape, Then Balance, Then Probes
Rather than jumping straight to the contour plot, work through this sequence.
- Deformed shape (at a sensible magnification) — is the model moving in a way that is physically possible? Are restrained faces staying put, are parts interpenetrating, is the load direction right? Most result-interpretation errors are caught at this stage
- Global balance — sum of reactions equals applied load (structural); mass and energy balance (thermal/fluid). The same discipline as the reaction check in boundary condition verification
- Point probes — read numbers at representative points that can be compared against theory, hand calculations, or past results. Does the nominal stress agree with beam theory to within an order of magnitude?
- Contour plot — only now look at the distribution, with the scale set manually to suit the question
Reading Results Means Producing Material for a Decision
Interpreting an analysis is not appreciation of a picture; it is conversion into a decision — "does this design hold up?", "what should change?". So the definition of reading results correctly is: write down the acceptance criteria (allowable stress, target pressure drop, temperature limit) first, then shape the results into a form (numbers, margins, distributions) that can be compared against them. Staring at a contour plot with no criterion is kicking a ball without knowing where the goal is.
Behind the Numbers — What Is That Colour, Exactly?
How a Stress Value Is Made — Integration Points, Extrapolation, Nodal Averaging
FEM does not compute stress at nodes directly. The actual pipeline is: (1) compute stress at the integration points inside each element, (2) extrapolate to the element's nodal positions, (3) average at nodes shared by several elements. A post-processor's averaged display goes through step (3); the unaveraged display stops at step (2). Two practical rules follow.
- The gap between averaged and unaveraged is a mesh-quality metric — with a sufficiently fine mesh the two converge. Regions where they differ substantially (roughly more than 5–10%) have large inter-element stress jumps, a sign the mesh is too coarse
- Averaging across material or thickness boundaries is invalid — stress is genuinely discontinuous at a dissimilar-material interface. Nodal averaging across such a boundary manufactures an intermediate value that does not physically exist, so restrict the averaging scope to within each part and material
Equivalent stress carries one more trap. Von Mises stress is a nonlinear function of the components, so "average the components, then compute von Mises" does not equal "compute von Mises at each point, then average". Argue about peak values without first checking which one your tool displays and the few-to-fifteen-percent difference shows up as a mysterious discrepancy.
Behind CFD Display Values — Cell Values, Interpolation, Wall Quantities
In finite-volume CFD the variables live at cell centres, and the contour display has been interpolated to nodes. The ones to watch are wall quantities (skin friction coefficient, heat transfer coefficient, y+), which are reconstructed from the first-cell solution and the wall treatment (wall function or low-Re). Because tools and settings differ in how wall heat flux is evaluated, a tool-to-tool comparison has to be aligned down to "which definition of heat transfer coefficient" — that is, how the reference temperature is taken. Cross-sectional averages change too depending on whether they are area-weighted or mass-flow-weighted; pressure-drop evaluation is normally done on mass-flow-weighted total pressure.
Reading Deformed Shapes, Vectors, and Principal Stresses
The deformation magnification factor can silently sit at several hundred, and misreading "this is a large deformation" without checking the factor is a classic accident. Get into the habit of looking at a scale factor of 1, and judge whether geometric nonlinearity is needed from the ratio of deflection to thickness or span. For stress, go beyond the scalar equivalent stress: principal stress vectors (with direction) let you read the load path. Buckling risk on the compression side, tensile principal stress running perpendicular to a weld line — many design questions can only be settled with directional information, which makes "contour plus principal stress vectors" a set worth standardising.
The Craft of Reporting
Rules for Figures That Communicate
| Item | Rule | Reason |
|---|---|---|
| Colour bar | Fix the range manually; use one identical scale across all comparison figures | Two auto-scaled figures cannot be compared |
| Upper limit | Set it at the allowable or yield stress and saturate the exceeding region in a single colour | "Where it fails" reads at a glance |
| Singularities | If excluded from evaluation, annotate them and keep them out of the scale | Prevents a singular value from crushing the scale |
| Path plots | Include a 1-D graph along the evaluation line | Supports a quantitative argument that colour cannot |
| Deformed shape | State the magnification factor on the figure | Prevents the exaggerated view from being misread |
The Verification Information a Report Must Carry
Alongside the figures that carry the conclusion, include a verification section that lets a third party judge credibility. The minimum set is four items: (1) mesh information and the convergence check (GCI or a level-to-level comparison), (2) the numbers from the reaction and balance checks, (3) a summary of the convergence history (all increments converged, final residuals), (4) the principal modelling assumptions (idealised boundary conditions, material model, geometry omitted) and an estimate of their influence. A report carrying all four becomes an asset you can reuse in later design changes and failure investigations.
Misreadings You Will Actually Meet
- Computing fatigue life from a singular peak value — a mesh-dependent, meaningless number. Follow the conventions of the assessment method itself: hot-spot stress, nominal stress, and so on
- Declaring "converged" while averaging hides the stress jumps — build the unaveraged check into the verification procedure
- Calling it "safe" on a single case — check sensitivity to scatter in the inputs (material, load, boundary stiffness), even with a small number of cases (screening sensitivity analysis)
- Judging better or worse by the impression the colours give — comparing two figures on different scales. Always use an identical scale, or a difference plot
Post-Processing Features by Tool
Key Points per Tool
| Tool | Averaging control | Practical note |
|---|---|---|
| Abaqus/Viewer | Averaging threshold (default 75%): no averaging where adjacent element values differ by more than the threshold | Set it to 100% to always average, 0% for unaveraged. Left at the default, the averaging scope changes from figure to figure |
| Ansys Mechanical | Averaged / Unaveraged / Nodal Difference displays | Nodal Difference (the gap between averaged and unaveraged) doubles as a handy mesh-quality map |
| Nastran-family post | Choice of element stress or nodal stress output (corner output) | Whether a CQUAD4 reports centre or corner values changes how sharp the distribution looks |
| ParaView (OpenFOAM etc.) | Explicit Cell Data / Point Data conversion filters | Cutting an isosurface from cell data straight away gives a staircase result — stay conscious of whether the conversion was applied |
Automating Post-Processing — Making Figures Reproducible
Build report figures from hand-taken screenshots and the viewpoint and scale drift on every re-analysis, making comparison impossible. ParaView (Python trace), Abaqus (Python scripting), and Ansys (Mechanical scripting/ACT) all support fully automated post-processing, and scripting a "standard figure set with fixed viewpoint, scale, and annotations" cuts the cost of case comparison, revision handling, and audit response by an order of magnitude. The ideal state is that every design change regenerates the same figures under the same conventions.
Frontiers — Making Interpretation More Powerful
From Single-Case Views to Difference and Statistical Fields
For design comparison, plotting the difference field \( \Delta\sigma(\mathbf{x}) \) directly carries far more information than putting the contours of A and B side by side, and in the UQ context visualising the mean field and standard-deviation field over many cases (where does it scatter?) is becoming standard. Regions of large scatter are the regions sensitive to input uncertainty, which feeds straight into the discussion of design robustness.
Sharing Results via Web3D and Lightweight Formats
Exporting results to a lightweight 3-D format such as glTF and sharing them so they can be rotated and probed in a browser is spreading. Unlike a still contour image, the recipient controls viewpoint and scale, so you hand over the data itself rather than "the picture you wanted them to see" — a genuine advance. Design reviews reach agreement faster.
AI-Assisted Anomaly Detection and Interpretation
Research is progressing on anomaly detection that flags "results unlike the usual" against a database of past analyses, and on learned models that propose critical regions from a result field, putting AI in charge of the initial interpretive screening. Its realistic role today, though, is an aid against oversight: it does not replace the verification discipline described here — the balance checks, the averaged/unaveraged comparison.
Troubleshooting
Symptoms, Causes, and Fixes
| Symptom | Likely cause | Fix |
|---|---|---|
| Peak stress keeps rising as the mesh is refined | You are evaluating a stress singularity | Identify the singularity with a convergence test; move to a hot-spot or nominal-stress assessment method |
| Averaged and unaveraged displays differ greatly | Mesh too coarse; averaging across a dissimilar-material interface | Refine locally; restrict the averaging scope to a single material |
| The same model gives different wall heat transfer coefficients in different tools | Differing definitions of reference temperature, wall treatment, and reconstruction | Confirm the definitions in the documentation and compare on heat flux itself |
| Contour plot looks patchy or noisy | Cell data displayed directly; poor-quality elements | Convert to point data; check element quality |
| After reporting, you are told "this can't be compared with last time" | Scale, viewpoint, and averaging settings differ figure by figure | Script the post-processing to fix the figure conventions |
| Result values do not match the theoretical order of magnitude | Inconsistent unit system; different definition of the output quantity (total vs. static pressure, etc.) | Trace the unit system end to end from the input; confirm the definition of the quantity in the manual |
The Last Line of Defence Is Physical Intuition
That's a lot of checks, and I'm not confident I can do them all. If I had to pick just one, which matters most?
If it is one, make it "compare against an order-of-magnitude hand calculation". Bending stress in a beam, pressure drop in a pipe, temperature rise from a concentrated heat source — if the simplified hand calculation and the analysis agree within a factor of two, there is almost certainly no large input error. Conversely, if the order is off, something is broken no matter how beautiful the figure is. The more sophisticated the analysis tool becomes, the more this humble cross-check is worth. A contour plot is not "the answer"; it is the detailed information you earn after passing the cross-check — keeping that order is the professional's way of reading results.
Related: verification methods for boundary conditions, checking and diagnosing solver convergence, mesh convergence verification with GCI.
Feel the theory hands-on with the interactive simulators in this field
Simulator LibraryRelated Fields
detail
error