04 / From measurements to investigation
What happens at an injector can influence producers elsewhere, after a delay. Explore the evidence, change a scenario, and decide what deserves a closer look.
Learning mode · synthetic field. All wells, measurements and responses below are invented. This transparent teaching model uses preset relationships and delays; it is not a trained Graph Neural Network (GNN) or Gated Recurrent Unit (GRU), and does not predict a real field.
Two injectors. Four producers. One reservoir unit, with a preferred flow path and a partial barrier. Select an injector and compare the three kinds of relationship evidence.
A GNN can learn relationships across a well network; a GRU can represent how behaviour develops over time. Inputs can include injection, production, water cut, well locations and reservoir units. Real deployment also needs data preparation, validated performance on unseen periods, and engineering review.
This lab instead uses visible, fixed coefficients. Its attention weights are invented examples, not neural-network outputs. It demonstrates the questions and workflow without claiming model accuracy.
A convincing curve is not enough. Check operating time, missing data and performance on dates the model did not learn from.
An uptime trap
A well produces 1,000 barrels of liquid per operating day. Change the days it runs in a 30-day month.
The online rate remains 1,000 barrels/day. Lower monthly volume alone does not prove reservoir decline.
For this exercise the online rate is a fixed assumption. At zero uptime, no online performance can be measured from that month.
Synthetic performance example
| Month | Observed | Modelled | Period |
|---|
These twelve points illustrate the distinction only. They are not evidence that a trained model works. Field totals must also be checked at well level.
| Producer | History available | Operating context | Review implication |
|---|---|---|---|
| P-01 | 12 / 12 months | Stable uptime | Useful reference case |
| P-02 | 12 / 12 months | Lift setting changed recently | Separate lift effects from injection response |
| P-03 | 12 / 12 months | Recent shutdown | Compare online rate with calendar-day volume |
| P-04 | 8 / 12 months | Missing injection-response observations | Insufficient evidence to confirm a weak connection |
Change one injector relative to its base rate of 1,000 barrels/day. Compare the four producers at a chosen horizon. The other injector stays at its base rate.
Read the result carefully
Compare the oil benefit with extra water handling, injection demand and facility limits. Endpoint rates are not cumulative volumes.
This illustration omits pressure limits, changing fluid properties and coupled network constraints. No uncertainty interval is claimed.
| Producer | Base oil | Scenario oil | Δ oil | Δ water | Δ liquid | Delay |
|---|
At time t, the liquid-rate change is injection-rate change × pair coefficient × response fraction. The response fraction is zero before the preset delay, then grows as 1 − exp(−(t − delay) / 45). Oil and water split using each producer's fixed incremental oil fraction.
Pair coefficients and delays are assigned for teaching, not learned. Rates use an illustrative common surface-volume basis; this is not a reservoir material-balance calculation. The model has no forecast uncertainty estimate and does not establish causality.
| Producer | Liquid-response coefficient | Delay, days | Incremental oil fraction |
|---|---|---|---|
| P-01 | 0.60 / 0.08 | 10 / 65 | 60% |
| P-02 | 0.25 / 0.30 | 25 / 30 | 35% |
| P-03 | 0.10 / 0.55 | 50 / 15 | 25% |
| P-04 | 0.02 / 0.12 | 80 / 45 | 15% |
A candidate is a place to ask a better question. It is not proof of bypassed oil, a confirmed flow path or an instruction to change injection.
Includes scenario assumptions, evidence limitations and proposed next checks.
In the constructed case, I-01 has a strong path to P-01 and weak communication across the partial barrier to P-04. I-02 most strongly affects P-03. Compare positive and negative injection scenarios, including a horizon shorter than the expected delay.
Here those relationships are built into the calculation, so agreement is only a teaching demonstration. A real validation must test a trained model against independent synthetic truth and unseen data, then assess transfer to a new field.
Before applying DDM to a real asset, agree on the decision to support, forecast horizon, trusted data sources, reservoir layers and pressure regime. Establish injector history, downtime and known candidate areas with the engineering team.
← Revisit production surveillance Explore the engineering processes →