The run that worked
and was wrong.

The line you scrolled past.

Run a flowsheet with a property method that needs binary parameters and does not have them, and this appears in the Control Panel. Not an error. A warning, between the block that converged and the one that is about to.

Search your own Control Panel for ARE ZERO. If the property method names change, the wording tracks it: the same warning appears with UNIQUAC, WILSON, UNIF-DMD and the rest. The one that catches people out is that it can fire for some pairs rather than all of them, and the wording then names the pairs instead of saying every one.

It is a warning because Aspen cannot know whether you meant it. A mixture of heptane and octane really is close to ideal, and zero is the right answer. A mixture of water and anything organic is not, and zero is a fiction that will still converge.

One number for every pair.

An activity coefficient model does not ask what a molecule is. It asks how differently A sticks to B than A sticks to A. That difference, divided by RT, is a binary interaction parameter, and there is one for every ordered pair in the mixture.

For three components that is six numbers, not three, because the pair A to B is not the pair B to A. The diagonal needs nothing: a molecule against itself has no difference to describe.

What zero actually claims.

Set the parameters to zero in NRTL and the model does not go quiet. It collapses to a specific, confident answer.

Three consequences follow immediately, and all three are things the lab can contradict.

• No azeotrope. An azeotrope is the equilibrium curve touching the 45 degree line. An ideal curve of constant relative volatility can never touch it, so a zero-parameter model will cheerfully design a column through a composition the real mixture cannot cross.
• No liquid split. Two liquid phases exist because the activity coefficients get large enough to make one mixture less stable than two. At gamma equals one there is no such thing, so the model will never tell you a decanter belongs there.
• No heat of mixing. Excess enthalpy is a derivative of the same parameters. At zero it is zero, so a mixing tee that really does get hot stays at feed temperature on the flowsheet.

Note what it is not. It is not a cautious estimate, and it does not widen your error bars. It is a definite claim, in the only direction that makes a separation look easier than it is.

What it cost a real column.

Benzene and cyclohexane boil 0.6 degrees apart and form an azeotrope, so they are separated by extractive distillation with a solvent. In a 2017 study in Scientific Reports, the authors ran the same separation twice: once on Aspen's built-in NRTL parameters, once on parameters regressed against their own measured equilibrium data.

The relative volatility the two parameter sets predicted differed by a factor of roughly two. The column designed on each came out like this.

The second failure in the same study is worse than the first, because it is invisible in the results. On a toluene and DMF pair, the built-in parameters predicted an azeotrope. The measured data has none. A model that invents an azeotrope will route your design around a wall that is not there, and every downstream number will be internally consistent.

Do the arithmetic yourself.

You do not have to take the stage counts on faith. Fenske gives the minimum number of theoretical stages at total reflux from the relative volatility alone, and it is exact rather than correlated.

Two things to take from moving that slider. A real design is several times N(min), because nobody runs at total reflux, so treat the number as a ratio rather than a stage count: the published designs came out at 100 and 52, a factor of 1.9, against a Fenske ratio of 2.3. And the curve near alpha equals 1.02, which is what benzene and cyclohexane do with no solvent at all, is the reason extractive distillation exists.

The fix, in order.

Most of this is one checkbox, and the rest is a morning.

• Open Methods | Parameters | Binary Interaction and look at the form. If the cells are empty or every one reads 0, that is the warning made visible. The source column tells you which databank each filled row came from.
• On the same sheet, turn on Estimate missing parameters by UNIFAC. UNIFAC predicts interaction from functional groups, so it will give you something physical rather than zero for pairs nobody has measured. It is an estimate and it says so, but it is an estimate in the right direction, and it will find an azeotrope that zero cannot.
• Check which databank the filled rows came from. VLE-IG and VLE-RK are regressed on real data. APV140 VLE-LIT is literature. R-PCES is estimated, not measured, and should be treated the way you treat UNIFAC.
• For the pair your separation actually turns on, regress your own data. Aspen's Data Regression System takes measured T, P, x, y and fits the parameters directly. That is what turned 100 stages into 52 in the study above, and it is the only step here that gives you a number you can defend.
• Run a Analysis | Binary T-x-y before you size anything. If the curve comes back as a smooth bow with no azeotrope on a pair you know is awkward, that is the ideal assumption showing through, not good news.

None of this is hard. It is all invisible, which is the actual problem: a converged run with a warning in it looks exactly like a converged run without one.