In this work, UNISIM-III, a giant-field benchmark case study was used. This synthetic reservoir was created to study and applied on Brazilian pre-salt fields ^[13-14]. The giant field consists of karsts and volcanic rocks, within a carbonate-depositional environment. The benchmark case study includes an ensemble of simulation models (UNISIM-III-2022) to capture uncertainty; and a reference case (UNISIM-III-R) which emulates the “true field”. For a detailed explanation of the simulation and reference models, we refer the readers to Reference [14].

This giant field is divided into four sectors (Fig. 2a), with the sectors sequentially developed. The sectors communicate with each other and each one corresponds to a production platform. For sustainable development, 100% of the gas produced is re-injected into the reservoir. The authors also found that applying water-alternating gas maximizes oil recovery. Fig. 2a shows an FDP in UNISIM-III-2022 with 33 producers and 32 injectors scattered across the field. 14 of these wells (Fig. 2b) were drilled within the first 1219 d of field development.

In this work, we only focus on optimizing the field’s net present value (NPV) using the decision variables from Sector-2 (S2): (1) Thirteen out of seventeen development wells of Sector-1 are already operational before the first cycle of CLFD. (2) Despite working with limited decision variables from only a quarter of the field (i.e., S2), the full-field models were adopted to include the impact of neighboring sectors for making a well-informed decision ^[10]. One must note that the locations of the wells of other sectors are considered fixed for all three cycles. Only to-be-drilled wells of S2 are revised using the accrued information over time. (3) Although the decision variables of S2 were optimized, we assimilated newly acquired information (well-logs, production data, etc.) from all four sectors. (4) Since our study target is the entire field, taking one or all sectors employing CLFD has little difference in execution time. (5) Sector-2 is with the maximum influence from neighboring sectors (Fig. 2a). Thus, it is a good sector for testing our predictive analytics-based optimization process.

Loomba et al. ^[11] assimilated data in 1219 d to optimize the NPV of the “true field” by 13%, using undrilled wells (of all four sectors) as the decision variables. We used their optimized FDP as our initial strategy. Table 1 chronologically summarizes the major activities in the field with additional information on the execution of each cycle of CLFD. We considered a period of 3 months to drill, complete, and produce for each well. Only one well per sector is drilled at a given time. At 1219^th day, Cycle 0 (pre-CLFD) was executed; at 1978^th day, Cycle 1 was executed to revise the blocks of 7 to-be-drilled wells in S2; at 2404^th day, Cycle 2 was executed to revise the location of 4 to-be-drilled wells in S2 and finally, at 3531^th day, Cycle 3 was executed to revise the blocks of 4 to-be-drilled wells in S2.

The first cycle of CLFD uses the production information of 0-1978 d. By using the ensemble-smoother with multiple data assimilation (ES-MDA) ^[15], we limited our work to 100 scenarios. The steps are provided below: (1) Generate 100 images of porosity, permeability and rock-types using 27 well-logs (including 13 new logs from wells drilled after 1219 d); (2) Using noisy production data over 5.4 years (1978 d), the data assimilation is performed for 100 scenarios, combining uncertainty attributes and geostatistical images; (3) Approve scenarios from the posterior ensemble of scenarios (only 44 were approved in this cycle); (4) Use the proposed workflow to efficiently revise the FDP with the accrued information; (5) Execute the optimization process using multiple RMs over ten iterations to improve all 44 simulation scenarios by a median value of 2.3%.

By implementing the optimized strategy in the reference case, we improved the NPV by 0.8%. Major improvement in oil recovery factor comes from S3 (1.0 percentage points) and S2 (0.5 percentage points). A larger change in recovery factor of S3 demonstrates how the revised decision in S2 helped the neighboring sector. Meanwhile, it stresses the importance of working with the entire field rather than isolated models when making a holistic decision for the complete field. The percentage change in oil recovery (at the end of Cycle 1) for individual wells is shown in Fig. 3. A similar trend between expected and actual percentage change in cumulative production is observed, with some exceptions (P45, P16, P42, P47 and P25). Such differences may occur because of various factors, including the uncertainty attributes, like faults close to such wells. It is intriguing to observe how even limited changes made in S2 can affect the wells in S4 (several kilometers away).

It is also vital to maximize the likelihood of success and measure it with the number of improved scenarios, especially when working with a huge uncertainty. In Cycle 1, we improved all the scenarios. To supplement this observation, we also calculated the probability of success of the selected FDP using Z-transformation (a kind of mathematical transformation of a discrete series) of the percentage change in results for each RM. The probability of success (or percentage change greater than 0) of the selected FDP in the true field was observed to be 99.5% in Cycle 1.

The efficiency of this cycle was benchmarked against the first cycle of Loomba et al. ^[1] (Table 2). It can be seen that the efficiency of method in this work has improved by 89%.

Like Cycle 1, we implemented CLFD workflow on subsequent cycles 2 and 3. In Cycle 3, the field had been developed for 9.7 years (3531 d), as about 32% of the field’s life. Unlike previous cycles, we assumed that fault transmissibility uncertainty is reduced by Cycle 3. We therefore used a range of 0 to 0.5 for faults transmissibility across the field (rather than 0 to 1).

However, we came across a critical problem in Cycle 3, which resembles practical concerns in the real field development. An extremely high gas-oil ratio (GOR) was observed in producer P17 (well from S1). As the well had been active for a considerable time, we did not consider this observation as an outlier. Given that P17 is drilled in proximity to S2, it is vital to ensure that our models depict this behavior. However, the simulated GOR of P17 was much lower than this observed GOR (Fig. 4) even after assimilating production data with ES-MDA. Assuming good connectivity between injectors and P17 as a solution to this problem, we proposed a channel connecting producer P17 and injector I17 to find a workable solution (Fig. 5a). In Fig. 5b, one can observe that the gap between the simulated and actual GORs of P17 is narrowed greatly after revising data assimilation.

Poorly history-matched models constrained us to repeat the data assimilation process in Cycle 3, leading to extra time. To accommodate these necessary changes, time spent on the data assimilation process was commensurate to the optimization process. As such, Cycle 3 took the longest time among the three cycles. In Fig. 6, one can also observe the corresponding changes. Well P27, placed in proximity to the unbeknownst channel in Cycle 2, was moved further away by the end of Cycle 3.

Table 3 presents detailed results of all three CLFD cycles. We exhibit the best FDPs obtained at the end of each cycle in Fig. 6.

Working with only limited decision variables from a quarter of this giant field, we were able to improve the field’s NPV by USD 2.7×10⁸ (increase rate 2.5%) which emphasizes the merits of implementing an efficient CLFD workflow. Fig. 7 shows the evolving NPV over cycles. Very high uncertainty in the models can be observed at the initial phase of field development. The degree of uncertainty is decreased with the evolving ensemble during development. Growing optimism in the evolving ensemble implied by a sharply increasing median NPV is obtained compared to the true field’s NPV. Assimilating newly accrued information over multiple phases of field development can be beneficial to decrease uncertainty and improve predictive capability.

Table 4 depicts the predicted recovery factor at the end of all cycles. As limited variables were modified to optimize the field’s NPV, we can observe limited improvements of recovery factor in individual sectors.

In a CLFD workflow, results propagate from one step to another. One must be mindful of the quality of these results to avoid failure of the complete workflow ^[6]. Hence, we presented a holistic solution to make the efficiency of the workflow improved by 85%. Using this efficient workflow, we recurrently revised the FDP of an extensively time-consuming model in about 30 d, a much acceptable timeframe.

The quality of being risk-informed is equally critical in a CLFD workflow. Adding to the discussion of Loomba et al. ^[1], we present new insights for practical implementation of a risk-informed CLFD workflow:

(1) Action. The depositional sequence plays a dominant role in the recovery process. Hence, one should construct the ensemble of geostatistical images using multiple geologically consistent scenarios. Considering efficiency, the number of geostatistical realizations generated and used for subsequent steps is another concern. Loomba et al. ^[1] used 500 realizations in their work, asserting that a large ensemble is necessary to capture the geological uncertainty. However, using as many as 500 realizations to develop a time-consuming model can be impractical. As such, this work uses only an ensemble of 100 realizations to work with time-consuming models. Despite using an 80% smaller ensemble on a highly heterogeneous and giant field fraught with various complexities, the revised workflow worked. This fact raises the question regarding the apt size of the ensemble for a successful understanding of geological uncertainty.

(2) Update inputs. In this work, we revised the FDP without including any direct information (i.e. known and identified information, without any uncertainty) ^[1] to test the workflow in a challenging environment. Regardless, this step is key to reducing uncertainty during the field's development phase. One should estimate the impact of all unknown uncertainties which can be used to make an informed decision for subsequent steps. For example, one should focus to approve the scenarios that include more impactful uncertainty attributes. Such a risk-informed step bolsters the likelihood of success of the outcome, particularly when one improves all scenarios.

(3) Data assimilation. Several processing activities are involved to separate the extracted liquids. While the total hydrocarbon extracted from the field is measured more accurately after such processes, the well production data is bound to suffer with higher noise. Also, depending on the availability of gauges, only limited well parameters (for example, total liquid rate and bottomhole pressure) may be available to perform data assimilation. It is therefore important to include more subjective knowledge of the field to improve the data assimilation process under such complications. In practice, the true reservoir characteristic of an oil and gas field is unknown. But using a benchmark case gives us the advantage of studying the "true answer" after executing all cycles of CLFD. We capitalized on this fact and revised the DA process in Cycle 3. Re-evaluating the uncertainty parameters, we realized that fractures running along the fault were the primary reason for the observed discrepancy. This observation highlights three aspects. First, unknown uncertainty attributes can cause mismatched production data. Second, as the reference case was built independently of the simulation model, it is common to encounter differences. Such differences helped us demonstrate problems from real situations. Third, such fractures only distributed in proximity of the fault. However, as the scale of the reference and simulation models is quite different, simulation models are not able to capture this behavior of fractures, even after DA. With the large differences in the scale, it is quite unrealistic to portray this behavior accurately as the DA cannot identify such instances. We can affirm this further based on the results obtained using channel to simulate the faults. In summary, a better DA technique is also required to work with a much larger noise and, without any huge uncertainty attributes. To improve DA in such problematic instances, we could also use local grid refinement in addition to new uncertainty parameterization to understand the mismatched area better.

(4) Approving scenarios to select and optimize RMs. In this work, we used reservoir engineering insights alongside the CLEO algorithm. One of its benefits in the CLFD workflow is that it helps prevent drilling a well in an unproductive region by using the data of the approved scenarios. This also helps mitigate the need for correctional steps, like flexibility of drilling (FoD) ^[1]. Meanwhile, to improve reservoir engineering insights, we need better ways to include the impact of different factors, like faults. For example, if there is a sealing fault close to the well, we need better techniques to incorporate this information and use zero oil flux from that direction. We also need to maximize the likelihood of success over the approved scenarios to ensure success in the "true field". Maximizing the expected monetary value (EMV) of the RMs alone does not ensure its success in the "true field". One must maximize the percentage improvement in the EMV as well as the scenarios improved. Using a growing set of RMs in conjunction with a risk-averse objective function is a good way to efficiently reach that objective. At the same time, a higher probability of success in the true field does not assure a lower mismatch between the expected and actual value of CL. There are perceived similarities between the concepts of (1) growing RMs with iteration and (2) propagation of best experiment (PBE). For example, both of them can improve efficiency of the optimization process, work with multiple realizations. And for a given subproblem, the total number of realizations is fixed. Nonetheless, they have some dissimilarities. The former solves a series of subproblems with increasing RMs, while the latter works with only one optimization subproblem. For growing RMs with iteration, the number of realizations is increased as we advance from one subproblem to the next; while for PBE, the number of realizations is increased within one subproblem only. The former was designed to handle as many realizations as possible, irrespective of how many of them are actually being improved, while PBE was developed to obtain a solution which improves all realizations within the subproblem. The former runs all simulations within a subproblem; thus, it does not improve efficiency within a subproblem. In contrast, PBE runs only a selected simulation; thus, it improves efficiency within a subproblem. The former has a high overall efficiency as it controls all subproblems, while the efficiency of PBE (as a standalone) is comparatively lower as it only acts a supplementary tool within a subproblem. In short, both frameworks supplement rather than substitute each other. As Loomba et al. ^[11] highlighted, we can achieve a globally optimal solution with simulation models. But, as model error always exists, the solution may be far from a globally optimum solution when applied in a real case. As such, we suggest using predictive analytics to make the optimization process faster. For a better prediction model, we could also use a more detailed feature matrix and effective learning algorithms. Learning feature importance to pick the best features for the feature matrix is another important step to improve the prediction model.

Depending on the size of the field, number of processors, average simulation time per model, and other computational capabilities, it can take weeks to months to revise the FDP. Nevertheless, we must ensure that FDPs can be revised within a realistic time frame. For a practical implementation of the workflow, it is also vital to include decision-making time as delays in oil production can directly impact the NPV and, at times, render the changes in FDP useless. As none of the previous works considered this decision-making period, we included it to exhibit the practicality of our workflow. While roughly 30 d were required by all the cycles for revising FDP, we added an extra 30 d to include the unforeseen managerial and engineering decisions. This extra buffer period also explains the benefit of CLFD workflow in case of additional unexpected delays.

The execution time of the workflow also raises the concern of it being a continuous process. Assuming it takes 30 days to simulate the whole CLFD workflow, the only additional information at this point, in most of the cases, would be a single noisy production data per well. As this additional information may not be value-additive, CLFD requires deliberate planning to maximize its value.

In addition, the number and size of cycles are two independent variables of CLFD. All authors have highlighted the benefit of increasing the number of cycles in CLFD. However, the number of cycles should be increased cautiously to avoid the disadvantageous in certain cases. For example, in the presented example, no new well logs are acquired between 1219-1756 d. In other words, no new spatial and temporal information is acquired for the to-be-developed sectors 2, 3 and 4, so ill-informed decisions may be taken, which can arbitrarily lead to success or failure. On the other hand, cycle size depends on the quantity of information it puts forward. So, for a successful implementation of CLFD (with adequate size and number of cycles), it is recommended to carefully recognize the amount of information, time consumed by CLFD workflow, objective function, and time of executing the workflow, among other things.

Updated ensemble of each cycle demonstrated similar characteristics (Fig. 8):

(1) Reduction in volatility. Decreasing coefficient of variation highlights the lower dispersion. For a given FDP, the sharply decreasing standard deviation can be observed in evolving data-assimilated ensemble.

(2) New information reduces uncertainty. Increasing minimum and decreasing maximum values led to a continually decreasing range of NPVs. Unrealistically optimistic and pessimistic values were ruled out from subsequent ensembles.

(3) Growing optimism. For a given FDP, the median NPV is always less than true value in the initial ensemble and ending up over the true value in the final ensemble, as shown in Fig. 8. This observation suggests that the ensembles are evolving to become more optimistic.

(4) Improved forecast. Despite growing optimism and large uncertainty reduction, the actual value of the field always falls within the prediction results of simulated models. This improves the predictive capability of evolving ensembles. At the same time, this does not necessarily imply that an individual cycle would always induce a positive change as it depends on the percentage of scenarios improved, among other things.

In this work, we only use a few wells (within a sector) as decision variables to optimize the objective function (NPV) of the entire field. Thus, it is understandable that the improvements are not significant as the decision variables in other sectors that play a critical role in the field behavior are not altered. In S2, however, the improvement of NPV is large, which exposes the benefit of CLFD in improving the field’s NPV. This setup also helps indicate the benefits of the ML-assisted CLFD workflow more clearly. The results also demonstrate the benefit of new information coming in the form of noisy production data. Even without updating new inputs in the simulation model during the earlier cycles, an ample change can also be observed.

Additionally, we mitigated bias related to the optimization process by starting with the initial strategy (that is, the optimized strategy proposed by Loomba et al. ^[11]) before executing Cycle 1. Given the great uncertainty and dearth of information, this FDP is not good enough for updated reservoir scenarios. At the same time, the optimized FDPs from the later cycles are also not amongst the best FDPs for the initial ensembles. These observations highlight two things: (1) the initial strategy is adequate for the uncertainty at the initial stage of field development, and (2) lack of information in the early phase of development misleads to suboptimal decisions.

With the implementation of each cycle, the new information is expected to reduce uncertainty with moderate degree rather than drastic change. For practical implementation of a CLFD workflow, we recommend using performance indices to ensure model variability. Descriptive statistics and image processing can be used to measure the variability at each phase of the cycle and maintain diversity in models. This step is essential to embrace key insights and differences from multiple scenarios. Maximizing the likelihood of success over such a variable ensemble would guide the geoscientists and engineers to select a more robust FDP.