To choose the right recovery strategy and optimize field-scale efficiency, reservoir engineers need to develop a qualitative understanding of the reservoir, explore alternative well placements, and consider different techniques for mobilizing immobile oil and improving sweep efficiency, etc. The traditional approach has been to use reservoir simulators. These have high computational costs, and this limits the engineers' ability to build cause-and-effect knowledge for the field by systematically exploring alternative model assumptions and different production strategies.
In the project, we have continued the development of flow-diagnostics methods that improve the geomodeller's understanding of flow patterns, highlight and quantify sweep and displacement efficiency, and provide measures on how geological uncertainty affects the displacement of fluids. Applying this to visualize traditional multiphase reservoir simulations simplifies analysis and interpretation. Equally important, however, one can also use the new methods to study dynamic reservoir responses without the need for expensive ?full-physics? simulations, or to limit the use of such simulations. One such example is the use of flow diagnostics to rank ensemble models, identify outliers, and select a smaller, representative subset that can be brought forward for more extensive (simulation) studies.
An attractive feature of flow diagnostics is their low computational costs and visually intuitive presentations of the interwell communication, which combined enable the reservoir engineers to perform insightful what-if analysis, rank ensembles of equiprobable geological models, and optimize recovery strategies under uncertainty. Examples of such analysis include interactive well placement optimization, interactive study of fault sealing affects well allocation, etc.
Data-driven methods have received a lot of attention lately, and particularly methods based on various forms of machine learning. Pure input-output models usually lack generality, this has led to a shift in focus toward physics-informed methods that maintain predictive power away from the data points used to train the model. In the project, we have studied a new approach that uses a traditional reservoir simulator, applied to a coarse graph-representation of the reservoir, as a basis for data-driven models, calibrated to measured data, or reduced models, calibrated to simulated data. Unlike typical neural-network type models, these graph-based methods respect basic flow physics. We have studied and improved existing methods and developed a new formulation, called CGNet, which has proved to be particularly promising.
All new methods developed in the project are available in SINTEF's open-source MRST software, which is widely used by scientists, engineers, and students all over the world, in both academia and industry. In addition to the results already discussed, this also includes new modules for ensemble simulation and adjoint- and ensemble-based optimization.
Prosjektet har gjort det mulig å bringe ideer vi har forsket på i snart to tiår opp til et TRL-nivå hvor de nå kan brukes i industrien. Noen resultater er allerede i bruk, mens andre forventes å bidra til å forenkle modellering og fortolkning av dynamiske simuleringer i årene framover.
Resultater fra prosjektet er synliggjort bredt, både nasjonalt og internasjonalt, og viktige kontakter er knyttet som på sikt vil føre til økt internasjonalt forskningssamarbeid.
Som et ledd i prosjekt har vi utvikle robust og veldokumentert åpen kildekode som gjør flytdiagnostikk allment tilgjengelig, og vi har allerede sett flere eksempler på at andre forskningsmiljø har plukket opp og brukt metodene vi har utviklet.
Siste, men ikke minst, har vi økt vår egen kompetanse innen fysikk-basert maskinlæring og utviklet en ny angrepsmåte for datadreven modellering som ser svært lovende ut og vil danne grunnlag for mye ny forskning.
Simulating multiphase flow is time-consuming using standard reservoir simulators. This limits the ability to build cause and effect knowledge for the field by systematically exploring alternative model assumptions and different production/injection strategies.
We will develop highly efficient computational methods to help engineers evaluate the effect of geological heterogeneity, locate regions likely to remain unswept, and provide estimates of sweep and displacement efficiency. The new tools will present how injection and production wells communicate in a visually intuitive way, and can be used to postprocess simulation results or as standalone tools for performing insightful what-if analysis, ranking ensembles of equiprobable geological models, and optimizing recovery strategies under uncertainty.
The most critical R&D challenges include: 1) Develop new methods for measuring the effect of heterogeneity and forecasting recovery efficiency orders-of-magnitude faster than conventional reservoir simulation. This includes both model-based and combined model/data-based approaches. 2) Extend these methods to incorporate uncertainty, e.g., given in the form of ensemble models. 3) Demonstrate how the methods can be used to optimize injection plans.
The project will provide reservoir engineers with new and intuitive computational tools that will enhance their understanding of reservoir models and simulated data, reduce the turnaround time for modeling, and simplify the process of working with ensemble models and finding optimal recovery strategies.