AI ENGINEER
AI Engineer overview
What the AI Engineer does, where to find it, the calculations it can run on your project, and what it will not do.
The AI Engineer reads your lab reports, builds and tunes compositional fluid models with the same equation-of-state engine as the PVT page, and answers questions about the whole project by running Converge calculations. Every number it gives you comes from a calculation it ran in your project, and nothing it builds replaces your published models until you accept it.
Where to find it
You can reach the AI Engineer in two places:
- The AI Engineer page. Choose AI Engineer in the sidebar. The left side has the AI engine panel (provider and model) and Build a model from data. The right side holds the Conversation.
- The side panel. Press AI Engineer in the top bar on any page to open a chat panel next to your work. Open the AI Engineer page switches to the full page; New conversation starts over.
Type in the message box and press Enter to send (Shift+Enter for a new line). A progress line shows the number of steps and the seconds elapsed. Press Stop to cancel at any time.
What it can do
| Task | What happens |
|---|---|
| Build a fluid model from a report | Drop a PVT report or paste a composition. It extracts the composition, plus-fraction properties, reservoir temperature, separator conditions and every lab experiment, creates a model and compares it with the measured values. |
| Tune a model to lab data | It regresses equation-of-state parameters to your lab data and reports the error before and after. See Fluid tuning. |
| Check lab data | It runs quality checks on composition, sample validity, CCE, differential liberation, CVD and separator data before tuning. |
| Run PVT calculations | Saturation pressure, phase envelope, flash, CCE, differential liberation, CVD, separator test, swelling test, compositional grading with depth, minimum miscibility pressure, viscosity model comparison and black-oil tables. |
| Compare models | Side by side for two to six models: saturation pressure, fluid type, GOR, API, CGR and lab error. |
| Write simulator input | Black-oil tables and compositional decks in the same formats as the export panel. The full file downloads from PVT > Black-oil & export. |
| Answer about the project fluid | It starts from the single project fluid, black-oil or compositional, including the bubble point from the PVT page. |
| Answer about other modules | It reads the saved data of material balance, decline, well performance, well test, simulation, network, integrated forecast, economics and production data, read-only. |
| Run any engine calculation | Material balance, decline fits and forecasts, IPR, VLP and nodal, well-test analysis, simulation, network, integrated forecast, economics, black-oil, gas and water properties, data parsing. These are the same functions the scripting editor and the REST API use. |
The side panel offers starter questions such as "Why is the bubble point of the active model where it is?" and "Compare my models and tell me which one matches the lab data best."
Files it reads
Build a model from data accepts CSV, TXT, PDF, spreadsheets (xls, xlsx), documents (docx) or a screenshot, up to 15 MB each. Spreadsheets are converted to one CSV block per sheet and docx files to text before the model reads them. You can also paste data under Or paste data and add Instructions (optional), for example "Use PR78 and match the bubble point of 2,950 psia at 212 °F." Press Build model.
What you see
When the AI Engineer creates or revises a model, a card appears in the conversation with the model name, its origin badge, New model or Revised model, and the buttons Open in PVT and Make active. When it recommends tuning changes, each recommendation card shows the parameter, Increase, Decrease or Keep, a confidence level, the suggested value and the expected effect.
Rules it follows
- Numbers come from calculations. It may not estimate, recall or invent a model value. If a calculation fails, it says so and says what it needs. See Citations and fallback.
- Lab values stay labelled. Numbers you type or attach are treated as measured data and shown under a "Lab" or "Measured" heading.
- Gauge pressures are converted. Every calculation that takes a pressure needs its basis (psia, psig, bara, barg, kPa or MPa). Gauge values are converted to absolute before any comparison, and results are reported in psia. When the lab quoted psig, it gives both, for example "2,449.7 psia (2,435 psig)".
- Isomers stay separate. iC4 and nC4, and iC5 and nC5, stay as reported. A lumped "C4" or "C5" is refused with a request for the isomers. It reports how many components it read against how many it modelled.
- Assumptions are stated. When data are missing, it says what it assumed, for example a C7+ SG estimated from its MW, or default separator conditions.
- Models are never edited in place. It creates a new model, or a draft revision linked to the original. It publishes only the final model it recommends to your library and tells you which one and why.
What it cannot do
- Overwrite a model. Drafts stay out of your library until it publishes its final recommendation, and the original model is always kept. A published model may be made the active model in the PVT workspace; the card then says Made active, and you can switch back from the library. Check the Project fluid picker on the PVT page afterwards, because that is the fluid every other module reads.
- Write to other modules. It reads material balance, decline, well, simulation and economics data but does not save changes to them.
- Tune black-oil correlations. It tunes compositional models. To tune black-oil correlations, use Tune to lab points on the Lab match tab. See Black-oil PVT.
- Reach outside Converge. It has no file or system access and no web search. It works only with your project, your attachments and the Converge calculations.
- Read very large inputs whole. Each other module's saved data is capped at 60,000 characters and all modules together at 250,000 characters. A text attachment is capped at 60,000 characters.
- Run without limit. A turn stops after 4 minutes on the server. The app also stops waiting after 90 seconds with no response and after 5 minutes in total. If your company's monthly AI request allowance is used up, requests are refused until the next month.
- Replace sign-off. An engineer still has to review and sign off a PVT model. See Review and sign-off.