Optimizing production in mature fields is a complex problem that is poorly addressed by the current spectra of subsurface modeling solutions. On one side, conventional 3-D numerical fluid simulators model the physics of multiphase porous media flow with a high degree of fidelity, offering the advantage of long-range forecasting  accuracy. However, intricate and time-consuming workflows and difficulty assimilating continuous data streams mean that the models and forecasts are immediately out of date. Furthermore, the computational cost of simulation means it is practically impossible to run the thousands of scenarios needed to quantify uncertainty.

Operators instead turn to analytical solutions and/or simplistic models such as material balance models and sector models for short-term forecasting needs that have the advantage of speed but lack the predictivity capabilities of full-field simulation. Argentine energy company YPF saw the business potential for new, datadriven technologies that combined the predictive power of a numerical simulator with the speed and computational requirements of an analytical solution, partnering with Tachyus, a provider of physics-based data-driven reservoir optimization, to test these new technologies on mature fields in Argentina.

Reservoir modeling technology
YPF used Tachyus’ Data Physics technology, which integrates physical governing equations into a datadriven machine learning model. Data Physics models are created by assimilating historical production and injection data to generate a set of petrophysical parameters that reproduce reservoir behavior over time. Unlike traditional data analytics techniques, the Data Physics model is permanently constrained by physical laws describing the movement of fluids in porous media—Darcy’s Law, Mass and Energy conservation and other constitutive relationships. Contrary to traditional reservoir simulators, Data Physics models can be created quickly once the initial data quality control and onboarding are completed. Understanding data quality and integrity proved pivotal to obtaining operationally relevant results.

The first step in the modeling process is data assimilation or training, using a modified Ensemble Kalman Filter (EnKF) that enables assimilation of data from thousands of wells and different data sources with a relatively small ensemble. Additionally, because the EnKF is known to underestimate uncertainty, statistical techniques are used to correct the uncertainty estimates to conform with empirical estimates.

The predictive capacity of the calibrated models for waterflood is demonstrated through a statistical backtest process, where the model is fitted to part of the historical data (training data), and the predictions from the model are statistically compared to the rest of the historical data (test data) at the field and well level. The predictive capability of the model is then assessed by using correlation coefficients between predicted and observed data. In this case, a minimum of 0.6 for both Spearman and Pearson correlation coefficients was deemed required to confirm predictivity, and values obtained were above this threshold in every case.

Once a predictive model is available for a specific field, the model can be used to predict future performance for different injection plans and to optimize for the best injection prescription. The technology uses evolutionary algorithms to handle the multi-objective optimization problem, which in this case involved several operational constraints, such as injection limits across different formations or wells. The result is a Pareto front (efficient frontier) that displays production values resulting from applying different injection regimes (Figure 1).

FIGURE 1. Each dot in the Pareto front represents a water-injection scenario. In this case, two and 15 years were analyzed, as shown by the bubble map. Each bubble is a graphic representation of the changes in water-injection volume recommended by the optimizer. (Source: YPF and Tachyus)

Scenarios on the Pareto front are used to make optimization decisions, such as reducing opex by maintaining production while slowing down water injection, increasing revenue by boosting production via water redistribution, reactivating injectors to boost production assessing infill drilling opportunities and many more.

Field selection and predictions
YPF selected three mature fields in each of the main producing basins in Argentina (Cuyo, Neuquén and Golfo San Jorge). A large number of wells and historical production and injection data at these fields indicated a high technical probability of training a predictive model. Each field has unique petrophysical characteristics. Two of the models were built at the level of individual production zones. With the other, in which back-allocation data were not readily available, the model was built at the well level. The optimization criteria were mainly production increases at constant injection within the limits of the existing facilities, though for one of the fields the technology was used to evaluate a 12-well injector reactivation campaign.

Evaluating the injector reactivation program
Along with increasing production, YPF aimed to reactivate 12 inactive injectors using a Data Physics reservoir model to find the best way to execute the reactivation program. Specifically, YPF needed to assess the optimum way to redistribute existing injection of 42,000 bbl/d of water and allocate an additional 9,000 bbl/d of water while reactivating the 12 inactive injectors.

A blind test was performed to increase confidence in the data-driven model in which the base-case (unoptimized) predictions for 15-year cumulative production (9,000 bbl of oil) were compared to a previously existing forecast obtained via a 3-D gridded simulation using a common commercial simulator (8,800 bbl of oil). This benchmark yielded agreement to within 2.5%, well within the individual forecast uncertainties, suggesting the data-driven model was at least as accurate as YPF’s previous model obtained by best-practice workflows.

Furthermore, this base case could be optimized by redistributing the same amount of water to add an additional 700,000 bbl of oil (+7.7%) over the 15-year period. After further modeling of efficient scenarios, YPF selected a target scenario for implementation in the field to increase injection from 42,000 bbl/d to 51,000 bbl/d of water. This scenario presents an optimized cumulative production increase of 1,500 bbl of oil (+16%) compared to the base scenario.

The injection reactivation analysis shows that nine is the optimum number of injectors to reactivate, leaving three injectors inactive. This turns into significant additional savings when compared to the analysis performed with the traditional commercial simulator. Additionally, the impact of each individual injector was assessed, producing an optimum reactivation schedule.

In this study, YPF has taken the first steps toward implementing new workflows for reservoir management using a data-driven approach grounded in the physics of flow. Applying the technology to injector reactivation demonstrated the flexibility of these new technologies, and YPF is investigating new applications and extensions, including infill drilling and real-time closedloop optimization.