Fluid tuning with the AI Engineer

The AI Engineer can tune a compositional (equation-of-state) fluid model to your lab data. It runs the same regression as the Lab data & regression and Live tuning tabs, saves the result as a new tuned model linked to the original, and reports the error on every measurement before and after. You review the tuned model and decide whether to use it.

#Before you start

You need a compositional model with lab data. Either:

• build one from your report on the AI Engineer page with Build a model from data, which reads the measured data along with the composition, or
• open an existing model in PVT > Compositional (equation of state) and enter its lab data on Lab data & regression.

Quote pressures exactly as the report does. If the report says "2,435 psig", say psig: the AI Engineer converts gauge values to absolute before any comparison. Results are always reported in psia.

#Ask for a tune

Ask in plain words, and name the targets you care about. For example:

Tune the active model to its lab data. Match the bubble point and the separator GOR and API.

The DL oil viscosity is off. Match it without changing the phase behaviour.

Which parameters should I tune for this condensate? Don't change the model yet.

A target you name is a requirement. The AI Engineer must include it in the regression, tune parameters that move it, and report its error after tuning.

#What it does

The AI Engineer follows standard practice (Whitson and Brulé; Coats and Smart):

1. Checks the data. It can run quality checks on the raw composition, sample validity, CCE, differential liberation, CVD and separator data, and it compares the model with the lab data point by point.
2. Checks the characterisation. It looks at the plus-fraction split and the plus-fraction MW and SG first, because plus-fraction MW is the most uncertain lab number.
3. Picks parameters. It measures how much each candidate parameter moves the errors and drops parameters that do the same job. It aims for one to three parameters.
4. Regresses. For a standard study it prefers a staged fit: saturation pressure by plus-fraction MW, then all phase behaviour, then viscosity. Otherwise it runs a regression on the chosen parameters within their bounds.
5. Checks the result. It compares the tuned model with the lab data again and can run sanity checks on the tuned model: heavy-end property trends, K-value crossing, viscosity trend, and volume-shift and interaction-coefficient ranges.
6. Publishes one model. Every attempt is a draft. It publishes only the tuned model it recommends to your library, and tells you which one and why.

#What it adjusts

Heavy-end Tc, Pc and acentric factor are kept monotonic with molecular weight by default. Interaction coefficients can be kept at zero or above.

#What it matches

Saturation pressure, separator GOR, stock-tank API, oil density at the saturation pressure, CCE relative volume and liquid dropout, differential-liberation Bod, Rsd, oil density and oil viscosity, CVD liquid dropout, cumulative production and gas Z-factor, and swelling-test data when you provide it.

#Reading the result

The AI Engineer reports, for the tuned model:

• the error on each measurement before and after, in %,
• the fitted parameter values,
• any parameter that stopped at its bound,
• any target still outside a normal match tolerance: 1 % for saturation pressure and 2 % for the other quantities.

If a target you asked for is still outside tolerance, it must say so plainly and try one more sensible parameter set. It may not report the job as done while a requested target is off.

A parameter at its bound usually means the data are inconsistent. The AI Engineer says so rather than forcing a fit.

#Excluded points

The AI Engineer may exclude a lab point only with an engineering reason, such as a contaminated sample, an obvious transcription error, or a CCE point that disagrees with its neighbours. It must tell you exactly which points it excluded and why. It may not exclude points just to lower the error.

#Check and accept

1. Press Open in PVT on the model card in the conversation.
2. Compare the tuned model with the lab data on Lab data & regression or Live tuning. The model library shows the tuned model as tuned from its parent; the original is unchanged.
3. Run Data QC and check the Tuned-model checks block.
4. If you want this model for the whole project, tick Use this compositional fluid for the whole project on Black-oil & export, or choose it in the Project fluid picker.

#Related

• AI Engineer overview
• Citations and fallback
• Compositional PVT: tuning to lab data
• Project fluid