Mesh Independence in Heat Transfer Simulations: A Practical Guide
A heat transfer result is mesh independent when refining the mesh further no longer changes the quantity you care about by more than you can tolerate. To show it, you solve the same case on a sequence of systematically refined meshes, track a few outputs tied to your design decision, and pick the cheapest mesh whose results sit inside an acceptable error band. The details matter, because temperature settles much earlier than heat flux, and a study that watches the wrong output can pass a mesh that is still wrong.
This guide walks through the whole process: setting up the sequence, choosing outputs, measuring change, deciding when to stop, and writing it up so a reviewer accepts it.
TL;DR
Change only the mesh. Boundary conditions, material properties, solver settings and convergence criteria stay fixed across every run.
Use at least three, preferably four or more, systematically refined meshes, and track the output that drives your decision, not just a convenient temperature.
Heat flux and wall gradients are usually far more mesh sensitive than temperature, so check them whenever heat transfer rate is the goal.
Stop refining when the change is small relative to your input uncertainty, then report the mesh counts, the metric, the change and the reasoning.
Extra refinement beyond what the accuracy need requires costs runtime and effort without fixing uncertain inputs.
Table of Contents
What mesh independence means for heat transfer
Every discretized model carries a discretization error: the gap between the numerical answer and the exact solution of the same equations. A mesh resolution study quantifies how sensitive the solution is to mesh size, using topologically similar meshes across a range of characteristic length scales, and the finest mesh is treated as the closest thing to the true solution when estimating that error (Sandia verification and validation report).
One heat-transfer paper defines grid independence as the point where results change so little between denser and coarser grids that truncation error can be ignored (Grid-independent Issue in Numerical Heat Transfer). Note what that does not say. It does not say the answer is correct. A mesh independent result can still be wrong because of a bad boundary condition, a poor turbulence model choice or incorrect material data. Mesh independence only removes one error source, the discretization, so that you can look at the others honestly.
It also does not mean “fine everywhere.” Spatial discretization should be fine enough that key model outputs no longer depend on it within an acceptable range, and regions with large temperature gradients need more nodes (ECSS Thermal Analysis Handbook). A thin boundary layer, a contact interface, a weld, a fin root or a thin coating deserves resolution. A large, nearly isothermal block does not.
Set up the refinement sequence
The goal is a clean experiment where the mesh is the only variable. Treat it like one.
1. Lock everything except the mesh
Freeze the geometry, boundary conditions, material properties, turbulence and radiation models, solver schemes and the convergence criteria. NASA’s spatial convergence tutorial is explicit that grid-generation parameters, boundary-condition settings and input data should stay unchanged between grids (NASA Glenn, Examining Spatial (Grid) Convergence). If you change a boundary condition between runs, you can no longer attribute differences to the mesh. If you are unsure your boundary setup is solid before you begin, see our guide to CFD boundary conditions.
2. Build a nested sequence
The cleanest recipe: generate a fine grid, then obtain coarser grids by removing every other grid line in each coordinate direction. That means the fine-grid point counts should be chosen so the coarser levels nest (NASA Glenn tutorial). On unstructured meshes you cannot drop grid lines, so scale the global and local size controls by a constant factor instead.
Sandia’s example shows how far a sequence can go. They generated eight meshes from one nominal mesh by scaling curve and surface sizes. The nominal mesh held 4,058,066 elements, and the sequence ran from 195,103 elements at the coarse end to 58,802,650 at the fine end (Sandia report). You rarely need that range. For an engineering decision, four meshes spanning a factor of roughly 8 to 10 in element count usually reveal the trend.
3. Pick a refinement method on purpose
NASA’s benchmark documentation separates three approaches: uniform refinement, quasi-uniform refinement that preserves relative grid density, and adaptive refinement driven by an error estimate (NASA case documentation). Uniform refinement is easiest to interpret and is the right choice for a formal study. Adaptive refinement is efficient for production, but you still need a sequence to show the trend.
Two practical notes for heat transfer models:
Refine the boundary layer mesh with intent. If you use wall functions, keep y+ in the range your model requires and do not refine into the buffer layer by accident. Our walkthrough on setting and validating y+ in a boundary layer mesh covers this.
In conjugate problems, refine both the solid and the fluid side near the interface, since the interface flux depends on both.
Choose the outputs that drive your decision
Pick the quantities before you run anything. Otherwise it is tempting to pick whichever curve flattens first.
The question “what am I going to decide with this model?” sets the output list:
Design question | Output to track | Likely sensitivity |
Will a component overheat? | Maximum temperature, temperature at sensor locations | Low to moderate |
How much heat does a surface shed? | Integrated heat rate, wall heat flux | Moderate to high |
What is the film coefficient? | Wall temperature gradient, Nusselt number | High |
How much pumping power is needed? | Pressure drop | Moderate |
Where are hot spots? | Local temperature field, peak location | Moderate |
The sensitivity column is a rule of thumb, not a guarantee, but it has a solid basis. A peer-reviewed study of heat conduction calculations for a tube in a fluidized bed found that heat flux can be much more grid sensitive than temperature. It recommends assessing discretization error in heat flux or temperature derivatives when heat-transfer quantities are the objective (Numerical Uncertainty in Heat Conduction Calculations, 1995). The lesson is that the quantity of interest controls the refinement strategy. Gradients and flux often need far more careful refinement than temperatures.
The same study found that, for the wall temperature gradient, a parabolic profile assumption converged faster than a linear one. So how your solver or post-processor computes the gradient at the wall also affects what you see across meshes.
Measure the change between meshes
Looking at a single probe value is a start. For anything you will defend in a design review, quantify the difference.
Point values and global norms
Sandia compared temperatures at monitored thermocouple locations and measured the differences between meshes, assessing convergence with normalized L2 and L∞ norms. L2 represents the overall, average solution difference, while L∞ represents the worst-case absolute difference (Sandia report). Use both when you have a field you can compare on common points. L2 tells you whether the model is broadly settled. L∞ tells you whether a single hot spot is still moving.
If you only track a few scalars, the same idea works with simple differences: the change in each output between successive meshes, expressed as an absolute value and as a percentage of the finest-mesh value.
An illustrative example
The numbers below are hypothetical, made up to show the bookkeeping for a heated plate in forced convection. They are not measured results.
Mesh | Elements | Max wall temperature (°C) | Change from previous | Wall heat flux (kW/m²) | Change from previous |
M1 | 0.5 M | 91.8 | n/a | 41.0 | n/a |
M2 | 1.0 M | 89.9 | 1.9 | 44.5 | 3.5 |
M3 | 2.0 M | 89.2 | 0.7 | 46.1 | 1.6 |
M4 | 4.0 M | 89.0 | 0.2 | 46.6 | 0.5 |
Read it two ways. Temperature moves by 0.2 °C between M3 and M4, which is negligible next to typical uncertainty in a convection coefficient. Heat flux, however, is still moving by 0.5 kW/m², about 1% of its value. If the decision rests on temperature, M3 may be plenty. If it rests on heat flux, you would either refine once more or at least report the remaining drift as part of your error budget.
Convergence rate on a log-log plot
To check that the sequence behaves like a sequence, plot your error norm against mesh scale on log-log axes and fit a slope. Sandia did exactly this. After excluding the three coarsest meshes, their L2 slope came out at −2.37 against an expected −2.0 (R² of 0.991), and their L∞ slope at −0.947 against an expected −1.0 (R² of 0.849) (Sandia report).
Two things stand out. First, the coarsest meshes were thrown out because they sat outside the asymptotic range, a common finding. Second, the observed slopes landed near the theoretical ones, but not exactly, and the worst-case norm was noisier than the average norm. If your fitted slope is wildly different from the scheme’s formal order, your meshes are probably not yet in the asymptotic range, or something else (iteration error, a changing mesh quality, a switch in a limiter) is contaminating the comparison.
Decide when the mesh is good enough
There is no universal percentage. The ECSS handbook says the acceptable variation from discretization is problem dependent and ideally much smaller than the uncertainty in the calculated values, stated there as typically 5 to 10 times smaller (ECSS Thermal Analysis Handbook).
That gives you a practical test:
Estimate how uncertain your answer already is from other sources: material conductivity, contact resistance, ambient temperature, a correlation-based film coefficient. A convection coefficient that is only good to ±20% makes a 0.5% mesh effect irrelevant.
Set the acceptable discretization error at roughly one fifth to one tenth of that uncertainty.
Pick the coarsest mesh in your sequence whose change to the next finer mesh, and whose estimated remaining error, falls under that threshold for every output on your list.
This is where an honest error budget pays off. For electronics, for instance, the discretization error often needs to be weighed against power dissipation, interface material and airflow assumptions. A practical PCB thermal analysis guide shows the kind of criteria that help you judge whether a given error is acceptable for the model’s intended use.
Justify the mesh you pick
If refining changes results only slightly, resist the urge to take the finest mesh “to be safe.” Sandia chose its nominal mesh because it sat within the linear convergence regime and in the coarse part of the sequence, which let it compute faster (Sandia report). The ECSS handbook adds that refinement beyond the accuracy need increases processing, runtime, storage, debugging, maintenance and verification burdens, and that it cannot recover inaccuracies caused by uncertain input parameters (ECSS Thermal Analysis Handbook).
That second point deserves emphasis. A 40 million cell model with a guessed contact resistance is not more accurate than a 4 million cell model with the same guess. It is just slower, and parametric sweeps and optimization loops that need dozens of runs will suffer first.
When you write the justification, include:
the mesh sizes or element counts and the refinement ratio
the outputs tracked and why they were chosen
the change between successive meshes, as absolute and relative values
the norm or error estimate used, and the fitted convergence slope if you computed one
the selected mesh and the remaining estimated error
what was held constant
One paragraph and one table are usually enough.
Common mistakes that make a study unreliable
Watching only temperature. As shown above, temperature can look settled while heat flux and Nusselt number still drift. Track the output tied to the decision.
Under-iterating the fine meshes. The heat conduction study found that iteration counts must increase as the grid is refined (1995 study). A fine mesh that stopped at the same iteration count, or the same loose residual, as the coarse one carries iteration error that looks like a mesh effect. Monitor the output values themselves for flatness, not only residuals. In conjugate problems, relaxation choices matter too, as covered in our note on conjugate heat transfer CFD relaxation.
Using only two meshes. Two points give a difference but no trend. You cannot tell whether you are converging, oscillating or sitting on a plateau that breaks at the next level.
Non-systematic refinement. Remeshing from scratch with different settings each time changes element shape, growth rate and quality along with size. Scale the controls instead so the meshes stay topologically similar.
Skipping the near-wall region. In turbulent flows, the wall treatment ties your mesh to your model. If you change the mesh and the wall treatment effectively changes with it, you are testing two things at once. See our guide on choosing and checking turbulence models for heat transfer.
Treating independence as validation. Mesh independence verifies the numerics. It says nothing about whether the physics matches reality. Compare against test data or a trusted analytical or correlation result separately.
Ignoring the starting point. For simple conduction problems you can build and verify a solver yourself, which makes discretization behavior transparent. Our article on a verified finite difference heat solver is a good way to see grid convergence on a problem you can check by hand.
Our perspective
We build thermal tools for engineers and students, and the pattern we see most is not too little refinement but refinement without a question. People run one coarse mesh and one very fine mesh, see a difference, and conclude the work is done or hopeless. Neither follows.
Our view is that the mesh study should be planned the way a test is planned. Write down the decision, write down the output that informs it, and write down the error you can live with before the first solve. Then the study is short, the plot is honest, and the answer to “is this mesh fine enough?” is a number rather than a feeling.
We also think access matters. Students and small teams often cannot afford a license seat just to learn convergence behavior, which is part of why we offer free engineering software alongside our paid toolkit. A hand-built, verified calculation of conduction, convection and radiation is often the best sanity check on a CFD result, and it costs seconds, not hours. If your CFD answer and a quick physics-based estimate disagree by a factor of two, no amount of meshing will settle which one to trust. Find the reason first.
— Joel
Cross-check your CFD with a fast thermal model
A mesh study tells you the numerics are stable. A quick independent estimate tells you the answer is plausible. The Jewlz Engineering Toolkit is an Excel-based tool that combines conduction, convection and radiation analysis with variable material properties, a wide temperature range and a built-in property database, so you can bracket a CFD result before or after you refine. For flow problems, see our CFD simulation software page, and for lumped and network-style checks, our thermal analysis tools.
FAQ
How many meshes do I need for a mesh independence study?
Use at least three, and four or more when the output is sensitive, such as heat flux. Three meshes give you two successive differences and a basic trend. Four or more let you drop the coarsest mesh if it falls outside the asymptotic range, as Sandia did when fitting convergence slopes.
What percentage change counts as mesh independent?
There is no fixed number. The ECSS handbook ties acceptable discretization variation to the problem and says it should ideally be much smaller than the uncertainty in the calculated values, typically 5 to 10 times smaller. Estimate your other uncertainties first, then set the mesh tolerance relative to them.
Why does heat flux need a finer mesh than temperature?
Heat flux depends on the temperature gradient at the wall, and gradients are derived from the solution, so they carry more discretization error than the temperature itself. A peer-reviewed conduction study found heat flux much more grid sensitive than temperature. That is why the output list should match the question you are answering.
Is a mesh independent result a correct result?
No. It shows that discretization error is small for the model as defined. Wrong boundary conditions, material data or physics models will still give a wrong answer, which is why validation against test data or a trusted reference is a separate step.
Can I skip the study if I use adaptive meshing?
Adaptive refinement guided by an error estimate is a legitimate way to build a mesh, and NASA lists it as a distinct refinement method. You still need to show the result is stable under further refinement for the outputs you care about, so run at least one more level and report the change.
Sources
Recommended
Comments