ISI Logo

Surrogate Reservoir Models: Frequently Asked Questions (FAQ)

FAQ SRM

Please communicate ANY questions that you might have on our technology and you will receive a prompt reply from ISI.

Q: What is the definition of SRM?

A: Surrogate Reservoir Model (SRM) is defined as replica of a reservoir simulation model that runs in real-time. SRMs are ensemble of multiple, interconnected neuro-fuzzy systems that are trained to adaptively learn the fluid flow behavior from a multi-well, multi-layer reservoir simulation model, such that they can reproduce results similar to those of the reservoir simulation model (with high accuracy) in real-time.

Q: Can SRM be developed for compositional simulation models?

A: YES. SRM can be developed for black oil models, compositional model, dual porosity models, etc.

Q: How fast is a SRM run?

A: Each SRM runs take a fraction of a second while reservoir simulation model runs may take hours or days.

Q: What are inputs to SRM?

A: There are two sets of inputs to SRM, static and dynamic. These inputs are combined to form an integrated, uniquely structured, spatio-temporal database that is used as the basis for the SRM. The static inputs to SRM are reservoir characteristics from all the grid blocks in the simulation model (the geo-cellular model), boundary conditions, and well configurations. The dynamic inputs are extracted from a limited number of simulation model runs as well as the operational constraints that are imposed on each run.

Q: What are the outputs of SRM?

A: Output of SRM is very much the same that are expected from a reservoir simulation model. They are pressure or rate profile at the well (depending on the operational constraints) or pressure and saturation distribution as a function of time (each time step) at each grid blocks in the simulation model.

Q: Has SRM been tested with commercial and in-house (operator's) reservoir simulations that are well known in the industry?

A: YES. SRMs have been tested successfully on ECLIPSE™ (Schlumberger), IMEX™ and GEM™ (Computer Modeling Group) and POWERS™ (Saudi Aramco) reservoir simulation models. They have been tested with models with up to 6.5 million grid blocks. There are no limitations on the size of the model that can be represented by SRM.

Q: Is SRM the same as response surface? What is the difference?

A: NO. SRM is completely different from Response Surface. Response surfaces are statistical representation of a large number (hundreds) of reservoir simulation runs. As such, they have no generalization and abstraction capabilities. Being statistical tools, response surfaces are bound by predetermined functional forms (linear, polynomial, power, etc.) that severely limit their ability. SRMs are adaptive AI-based systems that observe and learn the fluid flow behavior by rigorously training on the input-output data pairs prepared from a given reservoir simulation model.

Q: How many simulation runs are needed to develop a SRM?

A: Major innovations in spatio-temporal data curation and management that has been inspired by physics of fluid flow in porous media have resulted in a significant reduction in simulation run requirements for the development of SRM. Most SRMs can be developed by less than 15 simulation runs.

Q: How accurate are the SRM results when compared to the simulation runs?

A: SRM results are highly accurate. They can replicate reservoir simulation model's result with higher than 90% accuracy (sometimes up to 99% accuracy).

Q: What is the science and technology behind SRM?

A: Artificial Intelligence and Data Mining is the technology behind SRM. Specifically; the main science behind SRM is a proprietary implementation and integration of fuzzy set theory and artificial neural networks that resulted from more than two decades of research and development.

Q: What are some of the applications of SRM?

A: Surrogate Reservoir Model (SRM) provides the means for taking maximum advantage of simulation and modeling investments. Thousands of man-hours and millions of dollars are invested in developing sophisticated reservoir models. Long run-time of reservoir simulation models contributes to the inevitable fact that the full potential of the simulation models remains untapped. Some of the utilities of Surrogate Reservoir Models are:

  • Comprehensive Reservoir Analysis for designing field development plans.
  • Quantification of Uncertainties associated with the geo-cellular model.
  • Solving Inverse Problems for optimization purposes.
  • Effective and fast track AI-Assisted History Matching.
  • Real-Time reservoir management for Smart Fields.

Q: How can SRM be trained with such a small number of simulation runs?

A: If we treat data and information that results from a simulation run the same way that it is treated during the development of response surfaces, then the notion that "15 runs are too few to capture the fluid flow behavior in a reservoir" is absolutely valid. In order to be successful in any AI related project (development of SRM being one) you must completely revisit and revise your conventional view of data. I will talk about this fundamental, very important and crucial issue at a later time.

For the purposes of addressing this question let us divide the complete reservoir simulation and modeling process into two stages. Furthermore, let us define "Stage One" as the stage during which the initial (base) model is constructed and is modified to achieve a history match, and "Stage Two" as the stage during which analysis of the history matched model is performed in order to accomplish reservoir management objectives.

Normally during Stage One we change the geologic parameters of the base model to reach a history match while keeping the operational constraints constant (same as those in the field). During the second stage, we assume that the history matched model represents a reasonable geological realization and to perform analysis we change the operational constraints.

To minimize confusion, let us address the issue of number of runs that is required to build an SRM, in each of the above stages separately.

Stage One:

Let us consider that from the point of view of the geological model, a given geological realization is synonymous with one simulation run. Consider a heterogeneous reservoir with 100 wells that are reasonably distributed throughout the reservoir. From a geostatistical point of view a single run of this model provides "one set" of data and information about the state of this reservoir. Therefore, to collect 100 run "worth of data and information" you must generate 100 geological realizations and run them each, at least, once. But what if you could extract 100 run "worth of data and information" from only one simulation run (we will demonstrate how this can be accomplished, later)? If that would be possible, then you would not need 500 simulation runs and only 5 simulation run would be enough (and equivalent) to provide the 500 run "worth of data and information" that is required to represent the fluid flow behavior in the reservoir. Therefore, it all comes down to "how the data and information is extracted, organized, treated, processed and compiled".

The notion that SRMs can only be used for simpler models is incorrect. Actually the opposite is true. Lack of heterogeneity in a reservoir is usually source of problems for developing an SRM. Once you realize how geological model is treated during preparation of the data and information for SRM you see that diversity of species (read reservoir heterogeneity) is an asset and not a liability.

Stage Two:

For a history matched simulation model, the operational constraints are modified for all 100 wells; each group of wells can have a separate set of constraints during a simulation run. Once the objective of the project is identified, a carefully designed and targeted data and information generation, results in significant reduction of simulation run requirements. There is no need for Design of Experiment (experimental design) or Latin Hyper Cube. Engineering judgment is all you need. When a teacher prepares his/her lesson plans to discuss a subject in a class, she never uses statistical tool in order to determine which subject should be discussed and for how long. Presenting data and information to the SRM is more like developing a lesson plan in teaching reservoir engineering than to performing long and tedious statistical analysis.

Keep in mind, if you apply the same operational constraints to wells that have reasonably similar behavior and are located in similar geological setting, you are providing redundant information. In other words, by cleverly designing simulation runs, much of such redundancies can be avoided.

Once the requirements for these two stages are well understood, then data needs to be extracted, organized, treated, processed and compiled into a well-defined spatio-temporal database that forms the foundation of the SRM. To address the compilation of the spatio-temporal database, one needs to develop a fundamentally strong understanding of how artificial intelligence works and how it learns. You may have heard, numerous times, that AI does not use physics during its development and learning and that it is a black box. Although these statements seem to be literary correct, they misrepresent how AI works and how it can be used as an engineering tool.

To be able to efficiently extract the required information from the simulation runs and present it to the AI (in both stages mentioned above) for learning purposes, following are necessary:

  1. The practitioner needs to completely understand the (physics) reservoir engineering aspects of fluid flow in the porous media, in other words, do not expect a computer scientist to do this for you. Domain expertise is the uncompromising requirement in any AI related problems in our industry. A fact that, unfortunately, many fail to realize.
  2. The practitioner needs to understand how AI learns. What are the distinguishing features of AI when it is compared to statistics? And how to communicate certain ideas to AI via data. How to debug a problem and how to measure success.

Q: What are other applications and uses of SRM?

A: Surrogate Reservoir Model (SRM) can serve as a comprehensive tool for data mining analysis of the numerical reservoir simulation models. Upon development of a numerical reservoir simulation model for an asset that thousands (if not millions) of man hour has been dedicated to, it is not hard to imagine that a wealth of information and knowledge is embedded in such a massive resource. SRM can help engineers and geo-scientists to unlock the mysteries that are associated with this valuable and often under-utilized resource. ISI tools such as IMprove (an Oilfield Data Mining tool) can be unleashed on the SRM in order to identify best practices in an asset based on the developed SRM.