Diagnose a diverging run

The residuals are rising, or the solver stopped, and you want to know which of the four usual causes it is.

This page is the order to check things in. The messages themselves are indexed in Troubleshooting.

Stop it first

A diverging run does not recover on its own. Select Simulation runs and press Stop gracefully — it writes the current step out before exiting, which leaves you something to look at. Kill does not, and asks before it does.

Pause, Stop gracefully and Kill are drawn on that pane only while the run is running or paused, and Pause is disabled for the whole of a solve: a solver cannot be suspended without corrupting its decomposition, and the pane says so beside the button. Stop gracefully is also always on the ribbon at Model → RUN, greyed with a reason when nothing is running — worth knowing, because a diverging run is exactly when somebody is hunting for the button.

Then work down this list

1. Read the Problems tab. It reports what is wrong before a run, and it is the easiest part of the interface to skip. Red means the run will fail or produce nonsense.

2. Look at where it fell over. The Log tab, not the summary — OpenFOAM names the entry or the cell it objected to on the last line.

3. Check the mesh. Quality reports non-orthogonality, skewness and aspect ratio. A poor mesh where the flow is interesting usually announces itself as divergence rather than as a meshing failure. Apply to numerics adapts the schemes to the mesh you actually have, and Apply to mesh settings backs the layer stack off when coverage is what did it. The Acceptance criteria section on the same pane is what the mesher rejects its own work against — worth reading when the mesh looks fine and the run diverges anyway.

4. Check the time step, for a transient case. Simulation control → Max Courant, with Adjust time step to Courant number on. The hint states the ranges the two algorithms tolerate.

5. Check the boundary conditions are physical. A pressure outlet on the inlet side, or velocity specified into a wall, will not settle. Each patch's This writes block shows exactly what it produces per field.

6. Back off the numerics. Numerics → the Robust preset. Relaxation that is too aggressive for the case is the fourth of the four causes, and the cheapest to test.

7. Improve the start. A transient run from rest, or a steady run from a uniform guess on a long internal-flow domain, spends its early iterations in a state the solver was never meant to handle. See restart from an existing solution.

When it stops with nothing useful

Simulation runs → Trap floating-point exceptions decides what happens when a NaN appears. On, the run stops at once, which is what you want while a case is being set up. Off, it runs on — some turbulence models form a quotient and clip it immediately, and the trap fires on the quotient rather than on the result, so a healthy run dies with no message. If a run dies silently and early, try it off.

Flat is not the same as diverging

Residuals that go flat well above convergence have stopped improving, not diverged. That is usually too coarse a mesh, or a genuinely unsteady problem being solved as steady — a steady solver applied to vortex shedding sits at a plateau forever, because there is no steady answer to find.

Check it worked

• Residuals fall steadily by several orders of magnitude.
• The Result block on Simulation runs gives a mass balance near unity.
• The Report's "Can these numbers be trusted?" table reads OK rather than CHECK or PROBLEM. Every row states the number it was judged on beside the verdict, so you can disagree with it.

Related: the residuals rise instead of falling — indexed by the symptom rather than the intent.