Scenario

Before a separation train is designed, the first question is can this mixture even be split by distillation — and where do azeotropes block the path? Engineers need a feasibility screen grounded in phase-equilibrium data before investing in rigorous column simulation.

Worked example. A solvent-recovery stream contains a suspected azeotrope that would cap achievable purity.

Agent workflow

1. Read the mixture: components and target purities.
2. Thermophysical property lookup — pull boiling points and VLE data.
3. Distillation feasibility screen — assess separability, detect azeotropes, and estimate the minimum number of stages / reflux.
4. Propose a feasible separation sequence or flag the azeotrope barrier.

Demo output

A feasibility verdict per split with the governing VLE behaviour, any azeotrope flagged, and a candidate column sequence. The magic moment: the screen reveals an azeotrope early and suggests an entrainer or pressure-swing route instead of a column that could never reach spec.

Deterministic vs LLM

• Deterministic — the VLE analysis, azeotrope detection, and stage estimates.
• LLM — the separation-strategy narrative.

Every equilibrium value is a validated calculation. The agent reads and routes; the engine decides.

Limits

Feasibility-level screening, not a rigorous column design. It shapes the flowsheet; detailed simulation and pilot data confirm it.