Integrated Reservoir Management: Decisions, Data, and Discipline—Episode 3

Episode 3: Building the Closed-Loop IRM System

In February 2006, Saudi Aramco brought Haradh III, the southernmost increment of the Ghawar field, onstream ahead of schedule and at its planned production capacity of 300,000 B/D. The development was built on 32 multilateral maximum-reservoir-contact (MRC) wells, 28 smart completions with downhole inflow control valves across 113 laterals, real-time geosteering, and an integrated field management system that linked downhole monitoring to surface control in real time.

SPE 105187 documented the outcome: smart MRC wells delayed water encroachment, improved flood-front conformance, and reduced water production relative to conventional well designs in the same reservoir. The peripheral water-injection program preceded crude production, by design, as part of a planned pressure-maintenance commitment that was built into the development plan rather than deferred after first oil.

Haradh III is cited frequently as a technology showcase. It is also an integrated reservoir management (IRM) discipline showcase, less frequently recognized as such. The peripheral injection preceded production. The surveillance architecture was designed before the wells were drilled. The decision to use smart completions was driven by a known water-encroachment risk in the Haradh carbonate that had been characterized across earlier Haradh increments. The loop was designed in from the beginning, not retrofitted after performance diverged from expectation.

That is what a functioning IRM system looks like in practice. The two preceding articles in this series examined why it so rarely does: decisions arriving late, confidence misplaced in models and data, and correct recommendations failing to reach the field. This final episode is about building the discipline that closes those gaps and keeps them closed.

Most reservoir management processes run in one direction. Data comes in. Models are updated. Analyses are completed. Recommendations are issued. Some fraction become field actions. But the outcome of those actions rarely feeds back systematically into the process that generated them.

The forecast that proved too narrow does not change the next forecast methodology. The threshold that proved too conservative is not recalibrated against what actually happened. SPE 119098 demonstrated formally that the value of closed-loop management over open-loop management comes not from any single optimization step but from the iteration. Each cycle builds on what the previous one revealed. That principle applies with equal force at the organizational level.

Before describing how to build that loop, it is worth addressing the argument that the industry's current investment in digital technology has already solved this problem.

The past decade has seen significant investment in artificial intelligence (AI)-assisted surveillance, real-time production optimization, and digital twin implementations. The intent is correct. The results have been uneven, and not for the reasons most often cited. The constraint is not the algorithm. It is the data the algorithm runs on.

The data challenge is not abstract. Allocation in multilateral wells completed across multiple zones is derived rather than directly measured and carries systematic uncertainty that is acknowledged at completion design and rarely propagated into the models that consume it downstream.

Reservoir characterization inputs, whether from core, image logs, or well tests, are expensive, infrequent, and not always representative of interwell behavior at the scale the simulation model requires.

The contextual information that would allow a surveillance algorithm to distinguish a genuine reservoir signal from an operational artifact, the separator turnaround that affected 2 weeks of allocation, the rate change that coincided with the pressure transient, exists in field records and individual memories rather than in the structured data record that AI systems consume.

Allocation data carrying 20% systematic uncertainty, fed into a machine learning surveillance model, does not produce better decisions. It produces faster wrong answers with higher confidence. A digital twin built on a dual-porosity history-matched model whose fracture transfer function uncertainty has never been formally quantified does not resolve the confidence problem described in Episode 2. It automates it, continuously, at scale.

This is not an argument against digital tools. It is an argument for sequencing. The closed-loop discipline this series proposes generates, as a direct byproduct, the data quality that AI systems require.

Teams that write down evidence thresholds clarify which measurements matter and how reliable they need to be. Teams that run forecast calibration reviews discover which data inputs carried hidden systematic error. Teams that document decision outcomes with stated expectations generate labelled, contextualized field response data that supervised learning needs to function as intended. The discipline creates the data readiness. The data readiness enables the AI. In that sequence, and only in that sequence, does the technology deliver what it promises.

Digital IRM should therefore be treated as an enabler of the decision system, not a substitute for it. Technology becomes most valuable when it accelerates a disciplined loop that already connects data, models, decisions, actions, and verification.

What does running the loop require?

Before any material field action, state explicitly what reservoir response is expected and over what timeframe. After that timeframe elapses, compare what happened with what was expected. If the response falls outside the predicted range, the discrepancy itself becomes information: about the geological concept, the threshold, or a mechanism that was not adequately represented. Something should change in the next cycle as a result. Not who was responsible, but wWhat the team now understands that it did not understand before.

Three indicators make the health of the system visible. Decision-to-action lag time, tracked for every significant decision, reveals whether the execution system described in Episode 3 is functioning. Forecast calibration score, the proportion of 12-month forecasts where the actual outcome falls within the stated uncertainty range, directly measures whether the overconfidence problem as documented in SPE 195914 across 56 Norwegian Continental Shelf fields is being corrected in this asset. Verify completion rate, the proportion of executed field actions with a documented outcome review, measures whether the learning loop is closing.

Haradh III worked because the loop was closed by design: injection preceded production, surveillance was built into the completion, water encroachment was managed in real time through downhole valve adjustment, and the learning from earlier Haradh increments was explicitly carried into the development concept for the next one.

That is not a technology story. It is a discipline story that technology enabled.

The fields that deliver the most from their resource base are not always the ones with the most advanced technology. They are the ones where the loop runs cleanly. The discipline must be built first. Then the technology can help.

For Further Reading

SPE 105187 Haradh-III: Industry's Largest Field Development With MRC Wells, Smart Well Completions, and the iField Concept by A.O. Al-Kaabi.

SPE 119098 Closed-Loop Reservoir Management by J.D. Jansen, et al.

SPE 195914 Production Forecasting: Optimistic and Overconfident, Over and Over Again by R.B. Bratvold, A. Abellan.

SPE Textbook Series: Making Good Decisions by R.B. Bratvold and S.H. Begg.