How Accurate Is CFD? What Acceptable Error Really Means

The short answer

The presentation is over. The contours looked good, the conclusions were clear, and then someone at the back asks the question every simulation engineer knows is coming: how accurate is this, plus or minus what?

There are two bad answers. One is "about 5%", said with confidence and no basis. The other is a ten-minute lecture on turbulence closure that ends without a number.

The honest answer is that CFD has no single accuracy. A simulation is accurate for a particular quantity, in a particular flow, to a degree you can estimate. Whether that degree is good enough depends on the decision you are about to make with it.

Why is there no single number?

Because one simulation produces hundreds of quantities, and they are not equally hard to get right. Pressure drop through a duct with attached flow is forgiving. A local heat transfer coefficient near a separation point is much harder. The place where a flame anchors, or the NOx at a stack, is harder again. One model can be within a few percent on the first and far off on the last, in the same run.

So when someone quotes an accuracy for CFD in general, ask which quantity and which flow. If they cannot say, the number means nothing.

Where does the error come from?

Three places. Keeping them apart is most of the skill.

• Numerical error. The equations are solved on a finite mesh and stopped after a finite number of iterations. You can estimate this part without any experiment, by refining the mesh and checking that the solution has stopped changing.
• Input error. The inlet flow rate came from a flow meter with its own tolerance. The properties came from a table. The geometry is the CAD model, and the real part has weld beads, fouling and a gasket that sits two millimetres into the stream.
• Model error. Turbulence, combustion and multiphase flow are modelled with closures that were calibrated on simple flows. This is the error people usually mean, and the only one that needs measurements to estimate.

A fourth number gets forgotten: the uncertainty of the measurement you compare against. An experiment is another estimate, with its own error bars.

What does a published comparison look like?

Supersonic ejectors are a hard test, because the whole device depends on turbulent mixing. A 2021 study in Applied Thermal Engineering screened several turbulence models against measurements. The k-omega SST model agreed best, and its entrainment ratio still differed from experiment by 14% on average and by up to 20%.

That is a careful study with a suitable model, and the answer is tens of percent on a quantity that depends on mixing. For pressure drop in attached internal flow you can expect far better. The spread between those two cases is the reason a single figure for CFD accuracy cannot exist.

A worked comparison

Here is how I would answer the question from the back of the room. The figures are an invented example to show the arithmetic.

You measure a pressure drop of 12.4 kPa. The transducer and its installation give an uncertainty of 0.5 kPa. Your simulation gives 13.1 kPa, so the comparison error is 0.7 kPa, about 5.6% of the measured value.

That does not yet mean the model is wrong by 5.6%. A mesh refinement study puts the numerical uncertainty at 2% of the computed value, 0.26 kPa. The inlet flow rate is known to 3%, and pressure drop in this regime scales with flow rate to a power of about 1.8, so that becomes roughly 5.4% in pressure drop, 0.71 kPa.

The comparison error of 0.7 kPa is smaller than 0.91 kPa. The difference between simulation and test is within what the measurement, the mesh and the inlet uncertainty can explain. The model error is somewhere below about 0.9 kPa, and with this data you cannot resolve it further. This is the reasoning of the ASME V&V 20 standard, reduced to one line.

It also tells you what to do next. The largest term is the inlet flow rate, so a better flow measurement would sharpen the comparison far more than a finer mesh.

So what error is acceptable?

The one that does not change your decision.

Comparisons are where CFD earns its keep. Two similar designs run with the same mesh strategy and the same models share most of their errors, so the difference between them is more trustworthy than either value alone.

Mistakes I see repeatedly

• Quoting a general accuracy. "Our CFD is accurate to 5%" is a sentence without a subject.
• Comparing with data of unknown uncertainty. If you do not know how good the measurement is, a 3% agreement and a 10% disagreement tell you equally little.
• Tuning until one point matches. Adjust a turbulence constant until the outlet temperature agrees and you have fitted one number. Every other field is now less trustworthy.
• Skipping the numerical part. Without an estimate of mesh sensitivity you cannot separate numerical error from model error.

What I would do

Before the study, write down the decision and the accuracy it needs. During the study, estimate the numerical uncertainty on the quantity that matters. Do it on the number itself; a contour plot will not show it.

When you have test data, compare with all the uncertainties on the table. When you do not, say so in the report, state which results are comparative and which are absolute, and name the measurement that would settle it.

Then the question from the back of the room has a two-sentence answer. For the pressure drop, the numerical uncertainty is 2% and we agree with your test within its tolerance. For the wall temperatures we have no data, so treat them as a ranking between the options.