Accelerating field development and target selection using AI and automation

Accurate reservoir modeling during the field development phase is a principal factor in defining between poor and excellent economic performance of an asset. However, traditional methods of analysis constrain any possibility to optimize the process, because it is too time-consuming to analyze the vast quantities of data required to create meaningful scenarios—producing just one or two modeling scenarios within the project timeframe has been the norm.

An efficient, semi-automatic well location optimization approach has shown significant speed-up in well target selection.

Today, cloud-based solutions provide us with a viable means to overcome this challenge. The use of artificial intelligence (AI) and automation technologies, powered by the virtually unlimited computing capacity of the cloud, are enabling operators to explore hundreds of scenarios in a fraction of the time. Simulating scenarios and returning detailed results are now achieved on the same day as the analysis began, as operators aim to improve decision-making by describing the entire span of uncertainty to maximize the value of offshore asset portfolios.

Leveraging machine learning, automation and advanced data analytics solutions, enabled via a cloud-based platform, Petoro and Schlumberger collaborated to develop a well placement solution that can automate the increasingly challenging task of analyzing ensemble-based reservoir modeling data.

WELL PLACEMENT SOLUTION

Reservoir modeling and simulation support reservoir management in a series of decision gates to mature projects, from opportunity evaluations to field development plan submissions. One, two or sometimes multiple technical and financially feasible concepts are studied before deciding whether to submit a field development and operation plan.

Field development optimization under subsurface uncertainty is designed to provide robust decision support along all phases of project maturation. Ensemble-based workflows have been introduced in recent years to capture subsurface uncertainty and to deliver a distribution of reservoir performance rather than single deterministic results.

Generating hundreds or thousands of realizations is becoming the norm, as it can now be done in a fraction of the time, using cloud-based technology. This is such an on-demand reservoir simulation service, that the challenge shifts toward the analyzing and interpreting of results, which can generate huge bottlenecks in analyses. To overcome these bottlenecks, Schlumberger and Petoro developed a solution that uses AI, enabled by the DELFI* cognitive E&P environment to accelerate the process.

The new approach relies on a self-supervised, deep learning model to categorize, cluster and rank 3D-structural features, based on the size and structure of connected volumes, to propose new well locations. This article focuses on well location optimization under subsurface uncertainty. Examples of how the well placement solutions have been deployed include greenfield and mature field scenarios.

Probabilistic well ranking. To characterize well performance, a simplified net present value (NPV) model was defined, which accounts for revenue from well oil production, and costs from water injection and water production. Well NPV is used to compare the well-specific economic demand, including CAPEX of the producer and an average CAPEX of platform and injectors, divided by the number of producers.

Two characteristic values are reported in the probabilistic well ranking: the well economic demand and the percentage of realizations for which the calculated well NPV delivers above the economic demand. Wells that may deliver a predefined economic demand on a few selected realizations, but fail overall, may be eliminated or replaced. The probabilistic well ranking delivers a statistical well delivery statement that can be compared to a target requirement, e.g., 90% of all wells deliver the economic demand at an 80% level or higher.

Probability maps. Uncertainty analyses in most reservoir modeling and simulation workflows are built on ensemble-based methods. Cases, included in the ensembles, are equiprobable or linked to a weighted probability. Probability maps are generated for efficient spatial evaluation of reservoir performance over time, synthetizing information for a given petrophysical property across large amounts of model realizations. Each grid cell of the model is represented by a probability of a given physical property to exceed a predefined threshold value.

Expert-guided machine learning (ML). For this project, drilling targets were selected, based on a reservoir opportunity index coupled with expert knowledge. The opportunity index was a rock quality metric calculated for every grid cell, and typically included a combination of porosity, permeability and mobile oil saturation. The application of a threshold criteria for acceptable values of an opportunity index generated potential target regions. As a refinement step, connected volumes were calculated for each region, and these were again subject to a threshold criterion. All remaining reservoir units were high-quality targets, referred to as “hotspots,” with enough drainable volumes to meet the economic demand of a well.

In this process, an expert labeled hotspots (based on structures, locations and potential drilling constrains) as “target” or “not-target.” A 3D convolutional neural network (3D CNN) ML model, running in the DELFI environment, was trained to quantify similarities between subsurface structures, further classifying subsequent hotspots as “target” or “not-target.” Hotspot regions qualified as target were included in a new probability map calculation. The resulting probability map was used for locating new well target regions.

CASE STUDIES

To test the efficacy and accuracy, Petoro and Schlumberger deployed the well placement solution in two separate case studies. The first covers a semi-
synthetic reservoir model, while the second case study details a real application scenario.

Fig. 1. Probability maps of the 18-well base case scenario with no reactive control for upper (top) and lower formation (bottom) at the beginning (left) and at the end of production (right). The color coding of the maps highlights the probability (P) of finding oil saturation (SOIL) above a threshold of 70%. The legend shows P(SOIL ≥ 0.7) separately for upper and lower formations. Red colored areas in the upper formation show a 60% probability finding SOIL above 70%, red colored areas in the lower formation show a 100% probability finding SOIL above 70%.

Case study 1: The Olympus challenge—open benchmark. The “Olympus challenge” has been defined as an open benchmark study on field development optimization under geological uncertainty. The object of the study was a semi-synthetic reservoir model, which was primarily designed for technology and workflow verification.

The geological model had no structural complexity, and different facies characterized upper and lower formations. An ensemble of 50 equiprobable realizations was provided to assess the impact of subsurface uncertainty on field performance objectives. A water injection strategy for pressure support was defined for a 20-year production forecast scenario, including seven injection and 11 production wells.

The objective of the case study was to deliver a robust field development plan, measured by the 50-case ensemble. For that purpose, Fig. 1 provides a probabilistic view on the depletion efficiency for the upper and lower formation over a 20-year production period, including all 50 realizations. Shown is the probability to find saturation oil above a predefined threshold. Maps for upper and lower formation, at the end of production, show large patches of remaining oil, which suggests a significant improvement potential of the field development plan.

The economic success criterion for this work was to deliver a well location plan, including seven injectors and 11 producers, with “a minimum of 90% of all producers delivering the economic demand at an 80% probability level or higher.” The challenge for this work was to relocate wells that delivered the economic demand below the 80% probability level, i.e., producers 3, 6, 7, 10 and 11.

Fig. 2. Workflow design.

Workflow design. A structured workflow design for well location optimization under subsurface uncertainty is suggested and summarized in Fig. 2. All workflow steps focus on the ensemble performance. Firstly (step 1), a well location design is simulated and well performance delivery for all producing wells is analyzed using a probabilistic well ranking (step 2). In case of a suboptimal well location design, low-performing wells are removed (step 3), and the reservoir performance of the revised model is analyzed (step 4). Finally (step 5), resulting information is combined with the supervised ML and modeling approach to identify hot spot regions for improved well performance delivery. The workflow is applicable to single infill wells, as well as full-field well placement designs. Multiple wells can be optimized sequentially or in groups.

Results. The qualification of well location designs in two consecutive optimization steps following the workflow is highlighted in Fig. 2 and summarized in Fig. 3. Lowest-performing wells (producer 6 and 7) are identified in the base case scenario. After relocation, both wells deliver the economic demand at an 80% probability level, shown on the left side of Fig. 3. In a second iteration, producers 3, 10 and 11 are relocated. The probabilistic well ranking on the right side of Fig. 3 shows that all except producer 3 deliver the economic demand at an 80% probability level or higher. This fully satisfied the predefined economic success criterion for the well location optimization project under uncertainty.

Fig. 3. Probabilistic well ranking calculated after the first optimization step (left) and the final well location design (right). Each dot shows the probability of a producer to deliver the economic demand after 20 years production.

The analysis of NPV histograms across ensembles for a period of 20 years of production, or more, specifically for the P10, P50 and P90 percentiles, demonstrates significant economic gain compared to the base case scenario. A robust field development design, however, is required to demonstrate consistent economic indicators outperforming across all ensemble cases, in comparison to a reference scenario, which can be measured by an offset-distribution. The final optimized well location design shows a consistent incremental increase of NPV for all realizations.

Case study 2: North Sea reservoir. Petoro’s motivations were to identify opportunities to increase the value of future production across its large portfolio on the Norwegian Continental Shelf. An important element is the identification and localization of future infill well targets. In the deterministic approach, targets were identified and optimized on the base case model. This was guided by the engineer’s knowledge and understanding gained during the history matching process. This workflow did not transfer easily to an ensemble model. Further, there was no guarantee that a single realization selected for having a production profile closest to the mean of the ensemble profiles would contain hotspots that were congruent across most of the members.

Fig. 4. Maps of remaining mobile oil volume in one target formation. Shown is the average map covering all equiprobable ensemble members (left) versus a randomly selected single member (right). The color coding follows a normalized scale between low (purple) and high values (red).

Figure 4 shows a map of remaining mobile oil volume in the northern part of a reservoir within one of Petoro’s North Sea fields. Hot spots brighten and diminish as the ensemble members are cycled through. The opportunity index approach described earlier helps position possible infill wells, based on a combination of drainable volumes of hydrocarbons and a predetermined economic cut-off, utilizing information across the ensemble.

Figure 5 shows a map of the probability of having an opportunity index exceeding a user-chosen threshold. This has been used, together with a map of average remaining mobile oil volume, to identify oil producer candidate locations. This workflow enables optimization of a field development plan, using the entire ensemble of subsurface realizations. This increases the robustness of a target design, compared to the alternative of optimizing target objectives, based on one or a few selected realizations.

Fig. 5. Probability map of a normalized opportunity index > 0.7 (left) compared to an average map of the opportunity index, including all ensemble members. The color coding follows a normalized scale between low (purple) and high values (red).

CONCLUSION

The methodology for well location optimization under uncertainty, coupled to a deep learning predictive modeling approach for assisting hot spot target identification, has proven to be effective. The iterative workflow design is built on a probabilistic case analysis and basic reservoir engineering principles. The machine learning model approach was introduced to capture subsurface features, including structure and property similarity for classification and selection of potential well target regions.

With application to the Olympus case study, a base line well location design was significantly improved in a few steps. The base line location design was underperforming, due to subsurface uncertainty; close to 50% of all producing wells were delivering significantly below target expectation. The well location design was improved in two optimization steps, which enabled more than 90% of all producing wells to deliver the economic demand at an 80% probability level or higher.

The expert-guided machine learning approach proves to be helpful for fast well target selection. However, the parameterization of the expert-driven input remains a challenge for generalization. Alternative training objectives for the machine learning approach focus on expert defined reference wells as a template for potential target wells.

In conclusion, an efficient, semi-automatic well location optimization approach has shown significant speed-up in the well target selection. The workflow design has been applied successfully to a semi-synthetic benchmark study, as well as to real field scenarios.

ACKNOWLEDGEMENT

This article was adapted from SPE paper 202660, “Ensemble-based Well Location Optimization Under Subsurface Uncertainty Guided by Deep-learning Approach To 3D Geological Feature Classification,” which was prepared for presentation at the Abu Dhabi International Petroleum Exhibition & Conference Virtual, Abu Dhabi, UAE, Nov. 9–12, 2020.