Checking Mesh Convergence in PDE Simulations

Why Convergence Is a Modeling Requirement

A spatial PDE model represents a continuous tissue with values computed at a finite set of locations and times. The method of lines turns the spatially discretized model into a system of ordinary differential equations, which is then advanced in time. The resulting prediction depends in part on the spatial mesh and time step. A simulation may look smooth and plausible while still changing materially when either resolution is increased.

A convergence check asks whether a specified numerical output approaches a stable value as the spatial mesh is refined and the time step is reduced. The check is not simply whether two plots look alike. It should be planned around outputs that matter biologically or for the intended analysis: for example, peak concentration, time to threshold, integrated exposure, or concentration at a specified tissue location and time.

A convergence study holds the PDE, parameters, initial and boundary conditions, geometry, and output definition fixed. Only the numerical resolution changes. This distinction matters: changing a diffusion coefficient while refining a mesh is not a convergence test, because the model itself has changed.

Refining Space and Time Separately

Let \(Q_{h,\Delta t}\) denote a computed quantity of interest using representative spatial mesh spacing \(h\) and time step \(\Delta t\). A useful initial study refines one resolution while making the other sufficiently fine that its contribution to the change is small. For example, to assess spatial convergence, compare several values of \(h\) using a time step whose further reduction has negligible effect on \(Q\). Then assess temporal convergence on a sufficiently fine fixed mesh.

For a regular mesh and a method operating in its asymptotic regime, the leading spatial error for a scalar output is often approximated by

where \(Q^*\) is the limiting numerical output as both resolutions are refined, \(p\) is the spatial order, \(q\) is the temporal order, and \(C_h\) and \(C_t\) are coefficients that depend on the problem and method. This expression is an approximation, not a guarantee. It is most useful when the leading error terms dominate, the refinement sequence is systematic, and the output changes smoothly with resolution.

If time error is negligible and \(h\) is halved at each level, the observed spatial order can be estimated from three outputs:

The corresponding estimate of the limiting output, under the same asymptotic assumptions, is

These calculations help characterize a refinement pattern; they do not replace checking whether the pattern is stable or whether the residual difference is acceptable for the intended use. For nonuniform meshes, halving a single representative spacing may not describe the refinement consistently, so use a well-defined measure of mesh resolution and document how it changes.

Testing Temporal Convergence

Time-step convergence should be assessed with the spatial mesh held fixed and fine enough that spatial error does not obscure temporal changes. Halve \(\Delta t\) systematically, preserve the same output definition, and compare outputs at the same physical times. If outputs are sampled only at step times, ensure that the sampling procedure itself does not shift the comparison time.

A practical study should also examine at least one combined refinement. Spatial and temporal errors can interact, and the leading-error expression may not describe a coarse regime. If refining both at once, however, a stable result alone cannot reveal whether spatial or temporal error was responsible for an earlier change. Separate checks make the source of numerical variation easier to diagnose.

Outputs, Tolerances, and Evidence

Choose outputs before running the study. A model can be converged for an integrated exposure but not for a sharp local peak; a field can be visually similar while a threshold-crossing time differs enough to affect interpretation. Use outputs tied to the intended scientific question, and include both global and local summaries when both matter.

Report absolute differences when the output has a natural scale, and relative differences when the output is strictly positive and a percentage change is interpretable. Relative error is not useful when the reference value is near zero. State units for dimensional outputs and the resolution levels used. Set tolerances based on how much numerical variation could affect the intended conclusion, rather than choosing a tolerance solely because it is easy to achieve.

For a field-valued solution, comparisons can include selected point values, spatial averages, or a norm of the difference between fields on a common grid. Interpolation onto a common grid adds its own numerical operation, so document it and avoid treating a field-level comparison as a substitute for checking key scalar outputs. A convergence claim should identify the output, tested resolutions, observed changes, and practical tolerance.

Separating Numerical Error from Biological Sensitivity

A PDE prediction can change because the numerical approximation changes or because a biological parameter changes. These are different questions. A mesh study holds parameters fixed and varies resolution. A sensitivity analysis holds the numerical setup adequately resolved and varies a parameter, such as diffusivity or a reaction rate. If both are varied together, their effects cannot be cleanly distinguished.

Diagnostics, Alternatives, and Limitations

Convergence evidence can be weak or misleading. Watch for these patterns:

• Changes do not shrink. The sequence may not be in an asymptotic regime, or the output may be affected by a numerical instability or an unresolved spatial or temporal feature.
• Changes alternate in sign or size. This can occur with oscillatory error, competing error terms, or nonsmooth outputs. Do not infer a reliable order from a single refinement ratio.
• Different outputs converge differently. A stable average does not establish that a local peak or threshold time is stable. Assess each output relevant to the scientific claim.
• Only a smooth-looking plot is offered. Plot appearance is not a quantitative convergence check; report changes in defined outputs.
• Parameters or output definitions change between runs. Such comparisons do not isolate numerical resolution.

Refinement studies are not the only way to assess numerical accuracy. When an analytic solution is available, direct comparison provides a benchmark. A manufactured solution can test whether an implementation reproduces a chosen solution with the expected refinement behavior. These approaches test a method or implementation against known behavior, whereas a convergence study of the biological model assesses resolution dependence when the exact tissue solution is unknown. They complement rather than replace one another.

Convergence is also distinct from validation. A converged numerical solution can accurately solve a PDE whose assumptions or parameter values do not describe the biological system. Conversely, agreement with measurements at one resolution does not establish numerical convergence. Refinement can be computationally expensive, and extrapolation can be unreliable when the tested levels are too coarse, the solution lacks the smoothness required for the expected order, or the output is sensitive to small shifts in a steep front or threshold crossing.

A concise report should state the mesh and time-step sequence, the fixed model settings, the outputs and units, the differences across levels, and the tolerance used to judge practical adequacy. If a clear refinement pattern is absent, report that the test is inconclusive rather than labeling the result converged.