Do not judge convergence from residuals alone. Compare residual behavior with engineering monitors, global conservation and field stability. A residual plateau can coexist with a useful converged engineering quantity—or hide a genuinely unstable model.
Problem: the solver keeps iterating but the residual curve does not settle
Common patterns include residuals falling initially and then reaching a plateau, repeating cyclic oscillations, one equation remaining much higher than the others, or engineering outputs such as pressure drop, force or outlet temperature continuing to drift even though residuals look acceptable.
The important distinction is whether the flow itself contains an unsteady mechanism or whether the numerical method is failing to establish a stable solution.
Classify the residual behavior before changing settings
Stable plateau
Residuals stop decreasing but pressure drop, force, mass flow and heat balance are steady. This may be acceptable if discretization and conservation are verified.
Regular oscillation
Monitors and residuals cycle with a repeatable period. The flow may be physically unsteady and unsuitable for a steady-state model.
Growing oscillation
Amplitude increases with iteration. Suspect instability, bad cells, boundary feedback, excessive timestep/pseudo-timestep or inconsistent physics.
Slow drift
Residuals look calm while an integral quantity moves steadily. The solution is not yet statistically or numerically settled.
Diagnostic order
- Plot the quantities you actually need.
Track pressure drop, force, mass flow, heat transfer, outlet temperature or another engineering KPI together with residuals.
- Check mass and energy conservation.
A stable residual curve does not compensate for persistent conservation error.
- Inspect the flow field for inherently unsteady structures.
Separated wakes, vortex shedding, rotating interactions, buoyant plumes and free surfaces often create real time dependence.
- Check outlet and interface feedback.
Recirculation reaching an outlet, poorly placed boundaries or unstable coupling across interfaces can create iterative cycling.
- Localize problematic cells.
Inspect regions with high gradients, poor mesh quality, collapsed prisms, tiny gaps or abrupt size transitions.
- Compare solver aggressiveness with the physics.
Large pseudo-timestep, strong under-relaxation changes or aggressive high-order startup can alter the character of the oscillation.
Fix in an order that preserves the diagnosis
- Correct geometry, BC and mesh defects first.
Do not use numerical damping to hide a physically inconsistent setup.
- Improve the monitoring strategy.
Use integral quantities, extrema and conservation in addition to normalized residuals.
- Reduce pseudo-timestep or timestep if instability is numerical.
Compare the response systematically instead of changing many controls simultaneously.
- Use a robust startup scheme where appropriate.
Establish a stable field first, then restore accuracy-oriented discretization and verify the result.
- Switch to transient CFD when the physics demands it.
If a stable steady solution does not exist because the real flow is periodic or chaotic, forcing residuals flat is the wrong objective.
Need a structured convergence diagnosis?
Use the Abecator CFD Troubleshooter to separate setup, mesh, physics and numerical causes.
When should a steady case become transient?
Consider transient simulation when monitor oscillations are repeatable, coherent flow structures move through the domain, or the solution repeatedly flips between physically plausible states. A transient run should then be assessed using timestep sensitivity, Courant number where relevant, and statistically converged outputs rather than expecting instantaneous residuals to collapse monotonically.
Check timestep scale
Estimate Courant number from local velocity, cell size and timestep before the transient run.
Common mistakes
- Declaring convergence because residuals cross an arbitrary threshold while engineering outputs still drift.
- Assuming every oscillation is numerical and suppressing a real physical instability.
- Increasing damping before checking outlet placement, mesh defects or interface behavior.
- Comparing residual magnitude across equations without considering their normalization and meaning.
- Running more iterations indefinitely instead of diagnosing why the solution has plateaued.
Related Abecator resources
Still unsure whether the solution is credible?
Send a non-confidential project brief for an independent CFD review.