ISI Logo

Automated History Matching

Comparing Traditional Numerical Reservoir Simulation with Top-Down Modeling

Shahab D. Mohaghegh - January 2019

In this document/Blog, “History Matching” process is compared in the context of traditional numerical reservoir simulation and data-driven reservoir modeling, also known as Top-Down Modeling (TDM). Furthermore, the comparison will include application of these technologies to both conventional and unconventional resources. For the unconventional resources, the comparison will be short and not detailed in this document/Blog. More details about the applications of these technologies to unconventional resources will be discussed in a separate document/Blog.

Conventional Resources

In order to forecast well behavior in a field, the reservoir model must be able to match the history of all the wells in the field with reasonable accuracy. Matching the production history (dynamic variables) of a reservoir is necessary but not sufficient in order to use the model for the purposes such as forecasting the reservoir behavior, optimizing its production and recovery, and identifying the location of the new infill wells.

This short document/Blog summarizes the contribution of a new technology (new, compared to traditional numerical reservoir simulation) that incorporates Artificial Intelligence and Machine Learning to perform automated history matching for data-driven reservoir modeling in conventional plays. History matching is defined as “the process by which the input data to a reservoir simulation model (geological description, fluid properties, relative permeability, etc.) are altered in such a way to match the recorded data (fluid rates, pressures, tracers, temperatures, etc.) [Reservoir Simulation: History Matching and Forecasting, James R. Gilman and Chet Ozgen, Society of Petroleum Engineers (SPE), 2013, ISBN 978-1-61399-292-0] ”.

Currently there are two types of reservoir modeling:

  1. Traditional Numerical Reservoir Simulation and Modeling: These are reservoir models developed based on our geological interpretations and current understanding of the general physics of fluid flow in the porous media. Depending on who is involved in the development of such reservoir models, it might include large amounts of biases, predigests, preconceived notions, as well as perceptions, interpretations, and beliefs,
  2. Data-Driven Reservoir Simulation and Modeling: These are reservoir models developed purely and completely based on facts, reality, and field measurements using pattern recognition capabilities of Artificial Intelligence and Machine Learning. Development of Data-Driven Reservoir Models require domain expertise (data-knowledge fusion).

During the life of an oilfield two sets of measurements are performed. One set of the field measurements are associated with drilling, completion, reservoir (rock and fluid) characteristics, while the other set of field measurements are associated with operations and production including well test and workovers. In the context of this document/Blog, we will call these measurements, two sets of realities. First set realities are associated with the reservoir and the second set of realities are associated with production. The objective of reservoir simulation and modeling is to relate (correlate) these two sets of measurements to one another.

History Matching with Numerical Simulation

In the context of history matching, in traditional numerical reservoir simulation, the “base model” is developed initially based on the first set of realities (measured reservoir related data). The fact is that the “base model” never matches the second set of realities (measured production from the field). As mentioned before, since two sets of realities (field measurements) are being considered, we need to either modify them both or select one of the sets for modification in order to accomplish a match for the second set. During the history matching process of the traditional numerical reservoir simulation, the first set of "realities" are modified in order for the model to match the second set of realities.

Historically, in our industry, we have incorporated a common approach in addressing the problem of the “base model” not matching the production related measurements from the field. Since reservoir engineers and reservoir modelers hardly ever make mistakes, if there is something wrong with the “base model” that causes it not to match the historically measured dynamic variables from the field, it must be the fault of the geoscientists. The geoscientist have built and then up-scaled the geological model that is being used as the foundation of the numerical reservoir simulator (the base model). Therefore, in order to achieve a history match of multiple dynamic variables (Oil, GOR, WC, etc.) for every single well and for the entire field, we (as reservoir engineers and reservoir modelers) have no other option but to modify the geological model that has been developed and presented by the geo-scientists. The modifications made to the geological model to achieve history match usually requires justifications. Years of such practices have resulted in an art of justification (a very good example of such justifications is the methods used for blaming it all on upscaling) that is completely intolerant of any new technology that might challenge the traditional approaches.

Even if the technology that is being used as a new alternative to the numerical reservoir simulation is one that is changing the entire world that we live in i.e. Artificial Intelligence and Machine Learning, it is not allowed to be applied to hydrocarbon reservoirs. From the point of view of the traditionalists in our industry Artificial Intelligence and Machine Learning are allowed to be used for anything other than modeling a reservoir. AI&ML are used to build autonomous, driverless cars, recognize text and voice and translate them to multiple languages in real time, protect the security of our airports, and in the near future, maybe even change the world and guide human evolution in directions that may completely modify the future of our species. Nevertheless, such technology should never be applied to what we have done as reservoir modelers for the past several decades. To some individuals, the traditional approach to modeling hydrocarbon reservoirs has turned into a religion that must never be challenged.

Let us take a look back at our current practices in reservoir modeling and specifically the process of history matching. How long does it take to history match a model for a highly complex multi-layer reservoir? Well, it depends on the project deadline. It can take a few weeks, a few months, or even more than a year. The final history matched model is achieved very close to the project deadline. If the deadline is extended (regardless of the original length of time) the history matching project can and will continue until the new deadline is met.

When it comes to traditional approaches to history matching a numerical reservoir simulation model, there is ALWAYS room for improvement. This is because the experience and the type of prior reservoir complexities that the reservoir engineer/reservoir modelers have been exposed to, will guide them on the history matching process of the new reservoir. When and if you bring in a new reservoir engineer/reservoir modeler into an ongoing project, based on her/his experience, the model can be modified, sometimes extensively. This simply means that the degree of success, the accomplishment, and the quality of the history match for a complex numerical reservoir simulation model is highly dependent on the professional (or the team of professionals) that is in charge of the project.

In traditional approach to history matching a numerical reservoir simulation model the modifications made to the geological model in order to achieve a history match can be classified into three major categories:

  • Global modifications of the geological model,
  • Regional modifications of the geological model, and
  • Local modifications of the geological model.

Global modifications of the geological model usually attempt to modify general structures of the hydrocarbon reservoirs such as depositional environments or other characteristics that would affect the overall behavior of the reservoir. Sometimes it is possible that such modifications move the entire reservoir model to a closer match than the model's behavior prior to the global modifications, however, it hardly ever (actually never) solves the history matching problem for all the individual wells in the field. As the number of the wells, the age of the field development (maturity of the field), and the complexity of the reservoir (multi-layer, highly heterogeneous) increases, so does the chances that the contribution of the global modifications prove not to be very helpful.

The second option is the Regional modifications. Regional modifications are global modifications that are applied to parts of the reservoir that sometimes may have been compartmentalized based on existing structural characteristics of the field/reservoir. Impact of Regional modifications are much the same as the Global modifications that only influence a specific sector of the field/reservoir.

Since Regional and Global modifications usually do not solve the problems (help the reservoir modelers to achieve history matching), reservoir engineers and reservoir modelers are forced to move on to the third option that is the Local modifications (this is ALWAYS the case). In Local modifications, geological model’s characteristics around each individual well that is not being correctly or closely history matched, are modified in order to achieve history matching. Sometimes direct modifications of the reservoir characteristics around the wellbore are attempted while some other times, in order to make the modifications indirect, the transmissibility of the grid blocks near the wellbore are modified. Since numerical reservoir simulation models provide a very large number of knobs to be played with for the accomplishment of a history match, the local modifications to achieve history match can become quite unreasonable. Other modifications that are sometimes used to achieve history match at the individual well levels include wellbore and completion related modifications such as “skin”.

Another serious challenge that usually undermines the success of history matching of traditional numerical reservoir simulation models is the effort to simultaneously history match multiple dynamic variables, such as Gas Oil Ratio, Water Cut, Reservoir Pressure and Water Saturation. Many of the mature fields that have a significant amount of such measurements, turn the history matching of such field into a nightmare. The problem that is usually encountered in such cases, is that while certain modification of the reservoir realities help history matching some of the dynamic variables, they end up destroying the history match of other dynamic variables.

Furthermore, following two notes need to mentioned:

  1. Most of the items mentioned above regarding the history matching of the traditional numerical reservoir simulations will make sense to professionals that have been involved with performing simulation and modeling of real-world complex mature field. Most of the academic exercises are designed to avoid such complexities that are encountered in the real life.
  2. One may say that in the traditional numerical simulation the first set of realities are not modified, but the interpretation of the first set of realities is modified in order to achieve history match. First of all, that is not correct especially if and when “Local Modifications” are used in order to achieve history matching. This is due to the fact that when transmissibility modifiers are used as modification while the actual measurements are not modified, people seem to forget that the transmissibility is a function of the measured values. So when they are modified, it means that the measured values are being modified, however, indirectly. Second, even if one accepts that it is only the interpretations that are modified (and not the measurements), then the interpretations are modified enough to match the history and it means that in order for them to match the future behavior, may be more modifications will be needed. This can actually be proven to be correct if and when blind validation in time is used. A process that reservoir engineers no longer use during the history matching of traditional numerical reservoir simulation models.

History Matching with Top-Down Modeling

Modeling production and recovery from hydrocarbon reservoirs using a completely data-driven approach is philosophically different from the traditional numerical simulation. Top-Down Modeling, which is a data-driven reservoir modeling technology, focuses on finding the most appropriate set of patterns in a large set of combined static and dynamic measured data from a given field with tens or hundreds of wells. In the context of Top-Down Modeling, all field measurements that contribute to production and recovery (well construction, completion, reservoir characteristics, fluid characteristics, operational conditions, etc.) are included in the construction, training, calibration, validation and forecast of the Top-Down Model.

The static data is used to build a simple geological model of the reservoir that serves as the foundation of large amount of complex behavior that are represented by the dynamic set of measurements. The connection and the interaction between the static variables with massive amounts of dynamic variables that are segmented into time-steps represent a highly complex and unconventional time series that cannot be treated as conventional time series data for predictive purposes. This will be explained in more details later. .

Given the fact that reservoir engineers and reservoir modelers were the original developers of Top-Down Modeling and not statisticians, make this technology a reservoir engineering related technology that uses AI&ML as a set of tools to make life easier for reservoir engineers, reservoir modelers, and reservoir managers in our industry. At the same time, Top-Down Modeling is completely different from traditional time series approaches that do not include all other measured data in the process. That is the main reason for Top-Down Modeling’s success in the real world (with so many use cases all around the world) when it is used as a reservoir management tool to identify best infill location and choke setting (production) optimization of the existing wells, etc.

One of the major differences between history matching using the Top-Down Modeling and the Numerical Reservoir Simulation is the dynamic variables that are used as input and output. In Numerical Reservoir Simulation you must first identify the well type. The well type determines which dynamic variable is used as the input. For example if the well is identified as an Oil Well, then oil rate becomes an input to the system and dynamic variables that are treated as output become GOR, WC, and FBHP, noting that FBHP is usually not measured but calculated using empirical relationships that include many modifiable coefficients. This approach significantly contributes to the degree of control that is imposed by the practitioner on the history matching process of Numerical Reservoir Simulation. As the amount of produced gas increases, the control of GOR becomes tough, therefore, if the practitioner changes the well type to a Gas Producer, then Gas production gets to be the input to the system and the easier dynamic variable that now is the ratio of the liquid hydrocarbon, becomes the dynamic output, etc. Now you can see why automating the entire history matching process in the context of traditional numerical reservoir simulation becomes a serious hurdle (this is above and beyond the issues associated with geological modeling’s contribution to the computer assisted History Matching).

In Top-Down Modeling, such options of selecting and then modifying the dynamic variables as input and output is completely removed. Since Top-Down Model is a coupled reservoir-wellbore model, rather than just a reservoir model, it works with the wellhead pressure (WHP) and choke settings and not the flowing bottom-hole pressure (FBHP). In reality, having measurements of wellhead pressure and the choke setting is far more realistic than having continuous FBHP measurements. Furthermore, in the Top-Down Modeling the starting point of the analysis is the wellhead and not the bottom-hole interface between well and the reservoir. This results in the inclusion and integration of any type of artificial lift into the reservoir simulation and modeling process.

In Top-Down Modeling Artificial Lift becomes part of the full field and reservoir modeling, analysis and optimization and the need to perform a completely different set of analyses and developing a completely separate set of models for the analysis and optimization of Artificial Lift is avoided. This would be of much interest to those that are familiar with the physics and the mathematics of analysis, modeling, and optimization of Artificial Lift. This is due to the fact that coupling of the reservoir and wellbore in the modeling process introduces very realistic boundary conditions that significantly increases the accuracy of analysis, modeling, and optimization of Artificial Lift process.

If the scientific paradigm shift and the philosophical contrast of Top-Down Modeling with traditional numerical simulation is not well understood and subscribed to, success of performing such simulation and modeling of highly complex mature oilfield will be seriously compromised. Those that are engaging in such efforts that treat Artificial Intelligence and Machine Learning as another set of statistical or mathematical algorithms for regression purposes, are wasting their time and their company’s resources because the results of their efforts at the very best will be a set of mediocre correlations without honoring any physics and causations. This is true even for those that try to incorporate physics in their superficial AI&ML approaches. Details regarding the correct and incorrect use of physics and geology in the context of data-driven reservoir modeling (or any other engineering related applications) will be discussed in a separate article.

Automated TDM History Matching

As far as the philosophy and the solution approach are concerned, the substantial differences that exist between Numerical Reservoir Simulation (NRS) and Top-Down Modeling (TDM) results in a completely different approaches to history matching between these two reservoir modeling technologies. While the history matching of the Numerical Reservoir Simulation can only be “assisted” by automated algorithms, the history matching of Top-Down Modeling is a completely automated process. In other words, in order to accomplish a history match for a Top-Down Model, all that needs to be done is to click a few buttons after set up, design and training of a Top-Down Model. This usually is a very short process when it is compared with the time that is required to history match a Numerical Reservoir Simulator.

Another major and important difference between history matching a Numerical Reservoir Simulation (NRS) model and a Top-Down Model (TDM) is the amount of modifications that need to be made to the base model. This fact significantly contributes to the complete automation of the TDM history matching process. In order to accomplish a history match, unlike Numerical Reservoir Simulation, in Top-Down Modeling no modifications are made at the “Well” level. Global modification to the base model is the only type of modification that is made to a Top-Down Model in order to accomplish a history match. Unlike Numerical Reservoir Simulation (NRS) model in Top-Down Model (TDM) the “base model” (the first model that is built with complete automation based on the first set of measured realities – reservoir characteristics) usually provides a good match for a significant percent of all the wells in the field. This fact makes the modifications that are mentioned below an easy to accomplish process.

As was mentioned before, no “Local (well-based) Modifications” are needed in order to achieve history matching in the context of Top-Down Modeling. All modifications required in order to enhance the quality of history matching in Top-Down modeling are “Global Modifications”. Following is a list of the global modifications that are made to a Top-Down Model in order to accomplish a history match. The experience has shown that history match of a Top-Down Model is a function of one or more of the following characteristics of the TDM (details of all the items that are mentioned below has been explained in detail in the book published by SPE [Data-Driven Reservoir Modeling]):

  1. The architecture of the TDM design,
  2. The architecture and/or the number of the data-driven models that are part of the architecture of the TDM design,
  3. The combination and the values associated with the hyper parameters that control the training, calibration and validation of the data-driven models, and finally
  4. The combination of the static and the dynamic variables that are used during the training, calibration and validation of the data-driven models.

When performing history matching for a given field using Top-Down Modeling, the possibilities are (a) the history match of the entire field is not good/acceptable, (b) the history match of certain region/s in the field is not good/acceptable, or (c) the history match of a certain number of wells are not good/acceptable. Based on previous experience, unlike the Numerical Reservoir Simulation the case (a) above hardly ever happens. Nevertheless, if either case (a) or case (b) as mentioned above take place, then it is suggested to examine and possibly modify the combined set of 4 items above in order to enhance the history match results. However, if the case (c) as mentioned above takes place (meaning that the majority of the wells in the field have a good/acceptable history match), then it is suggested to check the data of the wells with the poor history match and make sure that their quality is as good as expected, instead of modifying anything about the TDM. The TDM itself sometimes work as a fantastic tool for data QC and QA. TDM allows you learn so much about your field and the data that has been collected.

Final Note:

Unfortunately, there are many misuse of Artificial Intelligence and Machine Learning in our industry, specifically when it comes to reservoir modeling. Such misuses make significant contribution to the traditionalists in our industry that are very much against using AI&ML in reservoir engineering and sometime in any petroleum engineering related problems. I will address some of these misuses that include (a) lack of understanding of the existing differences between statistics and machine learning, (b) treating AI&ML as just another regression algorithm, (c) use of physics within context of AI&ML by using terminologies such as “data-physics”, or “physics-based data-driven” for modeling purposes, (d) treating production data as a conventional time-series problem, etc. in a separate document/Blog. Also, application of Automated History Matching to Unconventional Resources in order to address issues such as Frac-Hit will be discussed in another document/Blog.