Ivan Kozak
Pedro Alcala
Boston Consulting Group
Houston
Ferrante Benvenuti
Lorenzo Veronelli
Boston Consulting Group
Milan
An artificial intelligence-driven drilling approach allowed an operator to drill a full hole section in a geologically unpredicatable formations field with fewer bit trips and 30% higher rate of penetration (ROP) than historical performance. The number of parameter changes, meanwhile, was at least three times what would occur in a typical job.
AI drilling application
Artificial intelligence (AI) applied to well design and planning focuses on lithology, geomechanics, and well-event spatial correlations to optimize drilling. It accounts for expected performance and possible risks to minimize non-productive time (NPT). During well execution, AI optimizes performance across major drilling activities by providing real-time recommendations and alerts to the driller. In addition to eliminating NPT, AI minimizes wasted time from suboptimal execution of drilling activities. In a drilling-optimization well test, the rig crew was able to finish drilling within 50-70% of target time using the model’s real-time recommendations.
Drilling operations are subject to subsurface lithology, pressures, varying downhole equipment and rig equipment, shifting crews, and multiple well-design variables. Operations acquire a large amount of high-frequency data, from surface, downhole, measurement and logging while drilling (M/LWD), wireline, and reservoir sources. Operators often have several decades’ worth of field data at their disposal. But decisions made during drilling still frequently include a high degree of uncertainty because the engineer or drilling supervisor monitors a limited number of variables and performs only basic correlations and calculations. This results in suboptimal drilling and increased occurrence of major adverse events such as stuck pipe, influxes, and fluid losses.
AI solves complex problems by leveraging large amounts of data to perform instantaneous advanced analytical calculations. Results consistently retrain models for progressively better outcomes. Currently, many operators use software that visualizes key ratios and key performance indicators (KPI), which helps with decision making but fails to address complex multivariate elements to the problems. They centralize all data into one platform for cross-functional analysis, but do not apply machine learning for continuous performance optimization. AI delivers a step change both for optimizing performance and mitigating risks through two major drilling and completion applications:
- Well design and planning application (Fig. 1, left). Organizes spatial correlations of several variables (lithology, geomechanics, well events, etc.) to optimize planned well trajectory. The models predict drilling performance and risks to minimize NPT.
- Well execution application (Fig. 1, middle). Optimizes performance across major drilling activities by providing real-time recommendations and alerts during drilling and completion operations. In addition to further NPT elimination, the system minimizes invisible lost time (ILT), the differential between time to perform a specific activity and time it could have taken if parameters were optimized.
Well design, planning
BCG and client teams developed tools based on advanced machine learning models to support engineers in making fast and accurate advanced analyses to identify risks for a new well trajectory. Machine learning models use data from previously drilled wells in two categories:
- Well events. Circumstances that deviate from plans which cause NPT such as stuck pipes, fluid influxes, and wellbore fluid losses. These events are typically captured in reports highlighting lessons learned from the incidents. They are frequently not stored in a structured manner and their availability for new well plans is often scattered.
- Sensor data. Measurements captured by sensors on surface and inside downhole tools are not often stored in structured databases because of the variety of tools and third-party service providers. Manually extracting key information is time consuming. For long drilling times, records can be in the hundreds, and datapoints can add up to millions per well and billions per field.
Fig. 2 shows how data are extracted with the use of natural language processing (NLP) algorithms to create integrated databases storing heterogeneous data, unstructured data, and structured data in support of engineering teams. Fig. 3 shows that events mined can be four times the number of events previously identified (NPT codified).
NLP starts with a collection of unstructured files stored in many formats such as drilling reports (Acrobat), lessons-learned presentations (PowerPoint), and pore pressure-fracture gradient plots (Excel).
NLP algorithms extract events and metadata into a defined data hierarchy. These are stored in a systematic manner on the newly created platform. Data for well-design studies increase by an order of magnitude, and time required to prepare a study is reduced from weeks to a day.
Advanced ML spatial regression and classification models support engineers by coupling large amounts of data from previously drilled wells to well design parameters (e.g. well trajectory and mud program). These models predict potential risks for a new well trajectory. They provide risk estimates every 100 m with 80-90% of actual events predicted with 75-85% precision (one false alarm out of five).
NPT prevention, execution
Due to subsurface uncertainties, the operator must execute crucial decisions while reacting to dynamic circumstances. Advanced analytics and real time (RT) sensor data reduce uncertainties by warning of potential high risk situations hours before they materialize. The tool inputs predictions from well-planning ML models (trained on data from previously drilled wells) and fine tunes them with RT data coming from surface sensors, downhole measurements, and M/LWD tools. An end-to-end data stream within the driller’s ecosystem collects and processes data which subsequently are fed into ML models (Fig. 4). The models run continuously during drilling operations and communicate RT with field personnel through a monitor inside the doghouse. The tool gives warnings hours before a risky situation develops (Figs. 5-6). and suggests potential mitigation actions combining advanced analytics with human expertise. To retrain ML models and improve predictive performance, the driller can add feedback in the user interface.
New modules can be added to predict additional events such as stuck pipes, tight spots, wellbore fluid losses, fluid influxes, and well kicks if enough history is available for model training. RT monitoring of torque and drag can be added to give additional insights.
A recent pilot case deployed the tool in an offshore pilot and other wells (including onshore) in several geographies. In excess of 20 wells were drilled with 80-90% of actual kick events correctly predicted an hour ahead of actual risks at a precision of 60-70%.
ILT reduction
ILT reduction comes from continuous guidance in the field, providing target parameter recommendations for each routine activity during well construction. The focus is on cumulative time reduction instead of a single NPT event.
Drillers face three options when deciding on uncertain drilling parameters (such as rpm, tripping speed out of hole, or mud flow rate): follow the planned drilling program, spend downtime hours performing calculations to adjust parameters to the actual situation, or learn through trial and error.
The first option may fail under operations with high variance like geological uncertainty. The second option may not be possible during rapidly changing conditions. The third option increases risk of NPT and often results in suboptimal performance.
Rate of penetration (ROP) is primarily dependent on rpm, flow rate, and weight on bit (WOB). Historical reference values for these parameters in highly uncertain geological environments and continual changes to bottom hole assembly (BHA) components will not improve performance alone. Early analytical approaches could not adapt quickly enough to changes in drilling conditions in unexpected geological transitions when considering optimal points using offset well data. They could not understand or accurately predict behavior using deviations in drilling mechanics with new components.
AI models analyze historical data and look at data across different time horizons in current and offset wells. They quickly understand drilling dynamics and detect changes in well condition and retrain accordingly. The first step in the digital workstream for this tool captures all RT data from rig sensors, mud logging, and downhole M/LWD. It compares these data in context with offset-well historical data and any technical information required for physical estimations, such as BHA components, well geometry, friction factor curves, pressure drop estimations, pressure drilling windows, and formation prognosis.
Using contextualized information, the system identifies drilling activity then activates and trains the proper set of models that predict ROP. Once accuracy is confirmed, the model will offer recommendations for rpm, WOB, and flow rate for ROP optimization (Fig. 7).
Recommendations go through a final filter before reaching the end user. Rig mechanical component operational limits are considered and extended to other physical models applicable to specific fields or company best practices (e.g., stick and slip avoidance, optimal depth of cut, bit aggressiveness, mechanical specific energy, bit wear, or hole -cleaning considerations).
Retraining and recommendations update every 3-5 sec with current well conditions but are only displayed to the driller when significant changes are required. This discrimination avoids time loss from constant parameter adjustments. The drilling team supervises and manages recommendations by exception; they can be overwritten when the team identifies other criteria or phenomena outside the prescriptive tool-data flow. In Fig. 8 this same digital workflow is followed but applied to continuous sequential activities, in this case drilling, connection, and tripping.
Field case
An onshore hard-rock well used the ROP module in a field characterized by geological unpredictability. This environment, along with high abrasiveness of the formations, made it impossible for the operator to establish a drilling parameter roadmap with historical data. Premature bit wear and excessive bit trips were common.
The prescriptive tool was implemented in the 17 ½-in. section of the well. Well design, S-shaped trajectory, and comparable BHA components were kept the same as offset wells for direct comparison between the AI guided recommendations and conventional drilling plans. Recommendations for rpm, WOB, and flow rate were sent to the driller every 15 min or when a considerable change occurred in well behavior. With adoption of the RT guidance tool, ROP improved more than 30% compared to historic wells (Fig. 9). This result came from the model’s ability to detect and quickly adapt to changes within formations.
Average WOB recommendations were 30% higher than historical WOB, but discrete increases only occurred in select formations with elevated uniaxial compressible strengths. Energy was used only when necessary. The number of parameter changes was at least three times the number of adjustments a driller typically would make (Fig. 10), illustrating the tool’s ability to rapidly adapt to geological changes without human comfort-zone biases.
The system also minimized bit damage and downhole tool wear by recommending parameters that could be sustainable throughout the section, leveraging RT mechanical specific energy calculations and monitoring vibrations. This sustainable drilling approach allowed the operator to drill the full section with only fewer bit-trips, half the initially planned amount.
Authors
Ivan Kozak ([email protected]) is a Partner and Associate Director at Boston Consulting Group, Houston, and is the global subsurface topic leader for BCG. He has also previously served as a vice president with Schlumberger and a geoscientist with ExxonMobil. He holds a BS in Geology (1997) from Sewanee, The University of the South, an MS (2000) in Geology from the University of California, Santa Barbara, and an MBA (2006) from the Darden Graduate School of Business at the University of Virginia. He is a member of the American Association of Petroleum Geologists (AAPG), the Society of Petroleum Engineers (SPE), and is a member of the US National Committee for the World Petroleum Council.
Pedro Alcala ([email protected]) is a Principal at Boston Consulting Group, Houston, and is a core member of BCG’s Energy practice. He has also served in different operational and engineering positions in D&C for companies such as Schlumberger, Halliburton, Imperial Oil and Remedy Energy. He holds a double degree in Chemical and Petroleum engineering (2009) from Universidad de Oriente, an MS from University of Alberta, and an MBA (2011) from IE Business School.
Ferrante Benvenuti ([email protected]) is a principal at BCG office in Milan, Italy, and is a core member of BCG’s Energy practice. He has previously worked for Tecpetrol (E&P company based in Argentina) as Head of Planning & Continuous Improvement for the Neuquen region assets. He holds a MA (2012) in International Business and Energy Policy from the Fletcher School, Tufts University, and a BA (2008) in Ancient History from the University College, London.
Lorenzo Angelo Veronelli ([email protected]) is an expert consultant at BCG, Milan, Italy, and is a core member of BCG’s Energy practice. He has also served in operational and resource planning roles in Schlumberger. He holds an MS (2012) in Energy Engineering from Politecnico di Milano.