An advisor for the column that will not converge
Every simulator user knows the moment: the column failed, the log is long, and there is no obvious reason why. The Advisor tab on the distillation column asks Anthropic’s Claude to read the solver’s evidence and say what it would change.
Is it the solver, or the specification?
The Luyben depropanizer sample, asked for 95 mol% propane overhead, with its design specification allowed to raise the reflux ratio only to 2. It stops at 73.1 mol% propane and shows WARN.
- Diagnose deeper re-solved the column at reflux 2, 3 and 4 on a private copy, in about 35 s
- Its verdict: reflux 2 is below this column’s minimum, so no answer meets the targets there, and the solver itself is fine
- With the design specification’s reflux limit raised from 2 to 5, the column solved in 0.4 s: 95.0 mol% propane overhead at a reflux ratio of 3.90, and 3.0 mol% in the bottoms
Captioned, no sound needed. 1 min 25 s. The advisor’s run is shortened, and the captions say where.
A quick read, or an investigation
Confer with advisor
Seconds, and cents of API credit. The advisor receives the failed column’s configuration, feeds, specifications, the solver’s verdict, the residual history and the profile it was holding when it gave up, and returns a diagnosis with ranked recommendations. Good at familiar failure shapes: two duties fighting each other, a design specification pinned at a bound, a wide-boiling feed on the wrong engine.
Diagnose deeper
Minutes. Instead of inferring from one snapshot, the advisor investigates: it reads the rest of the flowsheet, then re-solves the column with changed specifications to test its hypotheses, such as a different engine, a higher reflux ratio or fewer stages, and reports what actually happened. That is how it tells “the solver could not find the answer” from “there is no answer at these specifications”.
The solver stays the only authority on every number. The advisor does not do thermodynamics and does not compute results. It reads what the solver produced and, in Diagnose deeper, asks the solver to run experiments. Nothing it reports is its own calculation.
It cannot damage your model
The advisor is read-only. Its test solves run on a private copy of the column’s inputs and never touch the column, its results, the connected streams or your saved file. Every change it suggests is yours to make or ignore.
It remembers what worked
Every solve attempt on a column goes into a journal saved with the case. The advisor always receives the last settings that converged and exactly what has changed since, so the working recipe cannot scroll away.
Your key, your account
The advisor uses your own Anthropic API key, entered on the Advisor tab and stored encrypted on your machine. Requests go straight to Anthropic and are billed to your account. Build with AI uses the same key. A Claude.ai subscription does not include API use.
A chemical engineer, with AI as the development partner
AI Process Simulator was built by one chemical engineer with more than twenty years in industry, working with Anthropic’s Claude as the development partner. The engineer decides what the software must do and checks every answer; the AI writes the code under that direction.
That way of working only holds up with a strict test discipline, which is why the software carries 1,138 documented checks and publishes what each one is measured against.
Try the advisor on a stubborn column
The advisor is in every copy of version 1.6.2. You bring the API key.