[Editor's Note: Dimitrios Voulanas is a member of the TWA Editorial Board and is the author of previous TWA articles.]
A reservoir simulator is designed to answer a familiar question: What happens if we run this operating plan? Reservoir management, on the other hand, asks a harder question: Considering all the plans available, which one should we choose?
The difference between the two seems small, but it changes the computational problem. One forecast may require one carefully prepared simulation. A decision may require hundreds or thousands of evaluations as injection rates, well constraints, cycle timing, and operating limits are varied.
The industry needs proxy models not simply because full-physics simulation can be expensive, but because many decisions require a broader set of credible alternatives than full-physics simulation alone can assess within the available time.
A proxy model, also called a surrogate or reduced-order model, is a faster representation of a computationally expensive system. Its value is not the speedup itself. Rather, it is the decision workflow that speed makes possible. A proxy can screen more alternatives, reject infeasible cases earlier, rank tradeoffs, and reserve high-fidelity simulation for the candidates that matter most.
The Bottleneck Has Moved From Forecasting to Search
Reservoir simulation remains indispensable for detailed forecasting. It encodes geology, fluid behavior, wells, controls, and operating history in a form that engineers can interrogate. However, a simulator is typically used to evaluate a defined case. A decision workflow must do more. It must compare alternatives, enforce constraints, expose tradeoffs, and determine which cases deserve deeper analysis. This changes what “good accuracy” means.
A proxy may reproduce a fieldwide pressure trend or a familiar production curve and still be unsuitable for optimization. If it reacts incorrectly when an injection rate changes, a producer is constrained, or a well is shut in, the optimizer can rank the wrong plan even when the average prediction error appears acceptable.
Fig. 1 shows the required division of labor. The simulator provides the reference physics and the data used to construct and validate the proxy. The proxy provides the speed needed for broad search. The optimizer proposes candidates, but feasible finalists return to the simulator before an operating recommendation is adopted. Those verification results can then guide improvements to the proxy.
The Decision Determines What Must Be Accurate
CO2-EOR makes the distinction concrete. CO2 has been injected into the SACROC Unit since 1972 and remains an important field setting for studying long-term injection, recovery, and storage-relevant behavior. An operating study may vary water-alternating-gas (WAG) timing, injection allocation, producer limits, bottomhole-pressure constraints, and shut-ins. The objective may balance oil recovery, CO2 utilization, and CO2 retention, while feasibility depends on pressure, saturation movement, breakthrough, and facility capacity.
A proxy designed for that decision must respond correctly to those controls and preserve those constraints. A model that predicts one baseline production profile accurately is not automatically qualified to compare hundreds of alternative WAG schedules. The Department of Energy (DOE)/National Energy Technology Laboratory (NETL) SACROC project frames the field problem as an integrated reservoir-management challenge involving data integration, reservoir characterization, CO2-EOR, and sequestration. A useful proxy must preserve the responses that affect those operating choices.
CO2 storage presents a related but different problem. At Equinor's Sleipner area, CO2 has been separated from produced gas and stored in the Utsira Formation since 1996. The company reports that more than 19 million tonnes had been stored by the end of 2020. The question is not only where the plume may migrate. It is how much CO2 can be injected, how quickly, through which wells, and with what pressure and containment margin.
A storage proxy must therefore preserve pressure buildup and plume-relevant state behavior under changing injection controls. This is not merely a modeling preference. EPA's Class VI requirements call for computational modeling of the injected CO2 plume and associated pressure front. DOE/NETL guidance on risk management and simulation also treats modeling as part of a broader assurance process that includes monitoring and verification. A proxy can accelerate scenario screening, but the regulatory quality model must remain the basis for containment, pressure, and operating decisions.
The same principle extends to underground gas storage and geothermal operations. The US Energy Information Administration describes underground gas storage in terms of working gas, base gas, injection capacity, withdrawal, and deliverability. Geothermal decisions center on circulation rate, pressure difference, temperature decline, pumping requirements, and sustained heat recovery. The applications may use similar modeling architectures, but they do not ask the model to preserve the same behavior.
Fig. 2 is therefore more than a comparison of applications. It is a design rule. Begin with the operating decision, then identify the controls the engineer can change, the objective being pursued, and the variables that determine whether the result is feasible. The DOE describes FORGE as a dedicated field site for developing, testing, and accelerating enhanced geothermal systems technologies. The project also makes its data publicly available. That combination of field testing and open data illustrates why model architecture, training data, and validation tests should follow the decision rather than precede it.
A Proxy Can Be Accurate and Still Recommend Wrong Action
Optimization changes the risk profile of a proxy model. A conventional validation exercise asks whether the model accurately predicts cases selected by the engineer. An optimizer on the other hand is not passive. It searches aggressively for combinations that improve the objective, often pushing toward the edge of the training domain where the proxy is least certain. This is why the approved operating range must be explicit and why performance should be revisited as new simulation, monitoring, and operational information becomes available.
That creates failure modes that average error can hide. A pressure model may reproduce fieldwide trends while missing a local pressure limit. A well-response model may match historical rates but react incorrectly to a new bottomhole-pressure schedule. A saturation model may look acceptable in aggregate while misplacing breakthrough. In each case, the proxy can be numerically fast and statistically respectable yet still recommend an unusable plan.
Decision-centered validation must therefore test more than overall fit. It should examine control perturbations, multistep forecasts, constraint variables, boundary cases, and whether the proxy preserves the ranking and feasibility of competing candidates. The practical question is not only, “Does the model predict the state?” It is also, “Does it preserve the candidate ranking and feasibility judgments that the reference model would support?”
Fig. 3 turns that principle into an operating rule. The proxy can screen and rank alternatives only within the validated operating envelope. Combinations near or beyond that envelope should return to the high-fidelity simulator rather than be accepted through silent extrapolation.
The envelope should also evolve. When high-fidelity verification identifies an informative boundary case, that result can be used to update the proxy and to design new, independent validation tests. This creates a disciplined feedback loop in which the supported domain expands where the decision process actually needs more coverage, rather than accumulating simulations without a clear purpose.
What Reservoir Teams Should Do Differently
A decision-ready proxy does not begin with a preferred algorithm. It begins with a clear statement of the engineering choice. Four practices make that distinction operational.
- Define the decision first. Specify the controls that can be changed, the objective to be improved, the constraints that cannot be violated, the decision horizon, and the consequence of a wrong recommendation. This prevents a technically impressive model from being trained for the wrong task.
- Validate feasibility, not only averages. Test the variables that determine whether a candidate can be used. If pressure, saturation, temperature, breakthrough timing, deliverability, or facility capacity governs the decision, those quantities need direct validation and explicit acceptance thresholds.
- Make the operating envelope visible. Define the range of controls and states for which the proxy is approved. Monitor how far each optimization candidate lies from the conditions represented in training and validation. Treat extrapolation as a managed exception that triggers review, not as an invisible assumption.
- Keep full physics in the loop. Re-simulate finalists with the high-fidelity simulator, compare both objectives and constraints, and feed informative discrepancies back into model development. The simulator remains the source of physical authority; the proxy supplies the search speed.
For young professionals, this workflow points to a broader skill set. The strongest contributors will combine three perspectives.
- Excellent reservoir physics to identify what must be preserved
- Enough computational modeling to recognize where a proxy can fail
- Sufficient optimization to connect predictions to operating choices.
Conclusion
A proxy model should not be judged only by how much faster it runs or how closely it matches a familiar curve. It should be judged by the decisions it supports, the constraints it preserves, and the discipline with which uncertain cases are returned to high-fidelity simulation.
Full-physics simulation remains essential for detailed forecasting, technical assurance, and final validation. Proxy models extend its reach by making repeated screening, constrained optimization, and rapid comparison practical. They allow reservoir teams to explore a larger decision space without pretending that the reference physics no longer matters.
The goal is not a faster answer to one scenario. It is a better decision among many.
