[Editor's Note: This article is part two of a three-part series. Part one, Before the First Well: AI-Driven Field Development Planning, is available here.]
Well planning is inherently multidisciplinary and an iterative process. Once a field-development strategy has defined the recovery targets and well sequencing, the focus shifts from deciding where and when to develop to determining how each well will be drilled. A target on a map must become a viable wellbore: a well path that reaches the reservoir at the required location and orientation with constraints such as dogleg severity and kickoff points, a casing program that isolates every hazard with all the load cases, a drilling-fluid system that keeps the hole stable, and a bottomhole assembly capable of drilling it reliably. As for a concept of selection on field level, the well-planning phase carries a massive leverage. The decisions made here decide a major aspect of the cost, most of the risk, and most of the drilling time long before a rig is mobilized.
Trajectory design, casing and tubing design, drilling-fluid design, hydraulics, torque and drag, wellbore stability, and bottomhole-assembly dynamics each depend on the others, and a change in one propagates through the rest. A revised casing setting depth can alter the available mud-weight window while a trajectory modification intended to improve reservoir contact can raise torque, drag, or collision risk beyond the string limits. Historically, this iterative interdependency has forced long cycles of manual iteration between specialists. Recent advances in the digital drilling engineering environments have reduced the time required for these iterations, while AI is increasingly helping improve both the quality of engineering inputs and the speed at which feasible design alternatives can be evaluated.
Faster, Safer Well Design
The most immediate gains come from automation, autonomy, and integrated engineering workflows. Modern well-planning platforms embed the engineering calculations directly in the design environment, so that each time an engineer adjusts a trajectory or a casing seat, the system re-evaluates hydraulics, surge and swab, torque and drag, collapse and burst, and cement placement, and flags any constraint that has been violated. Errors that once surfaced days later are now identified as the design is being built and iterated.
Machine learning (ML) adds a layer of informed recommendations on top of these checks. Models trained on offset-well data can suggest mud windows by hole section, eligible casing grades, and fluid systems that performed well in comparable formations, giving engineers a defensible starting point rather than a blank template. Where log coverage is sparse, physics-informed models estimate pore pressure and fracture gradient by combining data-driven learning with the underlying geo-mechanical relationships. This keeps predictions credible in the conditions where purely statistical models tend to fail. For the bottomhole assembly, models trained on downhole vibration and dysfunction data can show which parameter combinations are likely to induce stick/slip or whirl, allowing engineers to select strong operating envelopes instead of single-point settings.
Digital twin models and generative assistants extend this further. A digital twin maintains a shared, continuously updated model of the planned well, aligning subsurface, engineering, and operational views. Generative tools can draft the basis of design, populate templates from offset databases, and assemble initial checklists and hazard registers. The engineer remains accountable for the decision in each case. What changes is that routine assembly and validation no longer consume the hours that should be spent on judgment. The benefit within a field development plan context is not simply faster engineering. It is the ability to mature larger portfolios of development wells more quickly, evaluate alternative development scenarios, and reduce the uncertainty associated with delivering the development plan.
From Designing Individual Wells to Optimizing Development Outcomes
A technically valid well design is not necessarily the best design for the field. Every trajectory represents a trade-off between reservoir exposure, drilling complexity, completion effectiveness, operational risk, and development cost. Automated checks and ML recommendations improve the quality of a given design, but they do not, by themselves, determine the best design. Selecting a well trajectory is a true multi-objective optimization problem. The path must reach the geological target at an acceptable inclination and azimuth while honoring anti-collision separation from offset wells, limiting dogleg severity, minimizing measured depth and torque, and respecting casing and completion requirements, and staying within the safety envelope for all these aspects.
Conventional optimization techniques can identify efficient trajectories for a specific target and a fixed set of constraints. However, development planning rarely operates under such certainty. Reservoir interpretations can evolve, appraisal results can refine target locations, and development priorities can change as new information becomes available. As a result, engineers are often required to revisit and re-evaluate previously optimized designs.
This challenge closely mirrors the uncertainty-management problem encountered in field development planning. Rather than optimizing a single deterministic case, the goal becomes identifying solutions that remain strong across a range of plausible subsurface outcomes and development scenarios.
What an AI Well-Planning Agent Actually Does
A fundamental challenge in well planning is that trajectory design is inherently a sequential decision-making process. Engineers must balance geological objectives, anti-collision requirements, drilling limitations, completion needs, and operational risks, while accounting for uncertainty in the subsurface interpretation. As assumptions change, designs often require multiple rounds of evaluation and refinement.
In this context, just as a drilling engineer becomes better at planning wells by understanding the outcomes of previous designs, an Al well planning agent improves by evaluating many possible trajectories and learning which decisions lead to safer, more efficient, and drillable outcomes. Instead of optimizing only one trajectory, the agent learns a decision-making strategy that can be applied to new wells and changing subsurface conditions.
In practical terms, AI well-planning agents treat well-trajectory design as a step-by-step planning problem. At each step, the agent considers the current target interpretation, nearby wells, engineering limits, and drilling objectives before selecting the next move in the path. By training across many well and pad configurations, and across uncertainty in the target position, the agent learns the trade-offs between reach, curvature, collision risk, and drilling difficulty. This helps it propose efficient paths for new wells without starting every design from a blank sheet.
A well-planning agent is valuable only if its recommendations remain consistent with established engineering practices and operational limitations. The governing engineering limits must therefore be treated as firm constraints rather than preferences. The agent may try to improve objectives such as measured depth, torque, and drilling time, but it must not violate limits such as maximum dogleg severity, minimum anti-collision separation, casing-setting-depth windows, or the build and turn rates that the selected steering assembly can realistically deliver. In a physics-constrained AI–learning agent formulation, these limits define the region the agent is allowed to explore. The result is an agent that can search aggressively for efficient paths while remaining inside the engineering envelope a drilling engineer would review and approve.
Historical field data can provide valuable context during model development. These agents also learn from offset wells. As-drilled surveys, planned-vs.-actual trajectories, and drilling outcomes recorded across a field show which doglegs were achievable in each formation, how steering tools behaved, and where earlier wells encountered difficulty. These records can be used as demonstrations, allowing an agent to be pre-trained on sound, previously drilled paths before an AI well-planning agent refines the decision strategy further. This grounds the agent in regional drillability rather than idealized geometry, improves learning efficiency because the agents do not have to learn from scratch, and helps generate trajectories that local drilling teams can recognize as realistic. As additional wells are drilled, their validated trajectories and outcomes can be incorporated into later model-training cycles, subject to engineering review and data-quality checks.
One potential advantage of these AI-based approaches is their ability to evaluate designs across multiple geological interpretations rather than a single deterministic target. When appraisal data shifts the expected position of a boundary, the agent adapts its recommended path without requiring the planning process to restart. The result is a well plan that can be generated and revised quickly enough to keep pace with evolving subsurface understanding, while still respecting the engineering constraints that make a well safe to drill.
These safeguards do not remove the need for engineering oversight. A planning agent optimizes only the objectives and explores only the configurations and constraints it has been guided to follow. If a limit such as anti-collision separation or wellbore stability is misspecified or left out of the formulation, the agent can still produce a trajectory that is efficient on paper yet unacceptable in practice, which is why the constraint set itself must be defined and reviewed by engineers. For this reason, AI planning agents should be viewed as decision-support capabilities rather than autonomous planning systems. Their value lies in accelerating the generation and evaluation of alternatives, while engineers remain responsible for defining objectives, validating assumptions, assessing risks, and approving final designs.
The Engineer's Role
This is why well planning, like field-development planning, is moving toward collaborative intelligence rather than automation. As checking, recommendation, and trajectory optimization become increasingly automated, the engineer's contribution shifts toward framing the problem correctly: defining objectives and constraints, encoding the operational realities the model does not see, and judging whether a proposed design is physically and operationally sound. AI does not replace the well-planning engineer. It raises the value of that engineer's judgment by removing the routine work that used to obscure it.
The next article in this series follows the well from the planning environment to the rig floor, where the design meets the formation in real time, and the central question changes from designing the path to executing and steering it.
Ashish Fatnani, SPE, is an energy technology leader with over 15 years of international experience spanning petroleum and drilling engineering, digital transformation, AI, commercial strategy, and business development. He is an active contributor to SPE, serving in leadership and volunteer roles while mentoring young professionals. He has authored several technical publications and holds patents related to predictive analytics, drilling-data automation, and digital workflows. He holds an MS in petroleum engineering from the University of Alaska Fairbanks, a BE in petroleum engineering from the University of Pune, and an MBA from the Indian Institute of Management Bangalore.
Shashwat Verma, SPE, is a data scientist at Halliburton based in London where he develops and deploys advanced data science and AI solutions for the global energy industry. His expertise spans machine learning, quantitative analytics, optimization, and predictive modeling, with a focus on solving complex challenges across the oil and gas value chain. By combining domain knowledge with modern AI techniques, he helps improve operational efficiency, optimize decision-making, and unlock greater value from engineering data. He is an advocate for data-driven innovation and enjoys sharing insights on the evolving role of AI in transforming the future of energy through smarter, more efficient operations.
Yusuf Ajibola Falola, SPE, is a senior technologist at Halliburton based in Houston where he develops advanced digital solutions that leverage AI, machine learning, and data analytics to address complex challenges in the energy industry. His expertise spans production optimization, predictive analytics, digital twins, intelligent field operations, and industrial AI, helping operators improve efficiency, enhance decision-making, and maximize asset performance. He is passionate about applying emerging technologies to solve real-world engineering problems and bridging the gap between data science and production engineering. He is also an active advocate for innovation and enjoys sharing his knowledge on the growing role of AI and advanced analytics in shaping the future of energy.
Vivek Kesireddy, SPE, is a senior data scientist at Halliburton Landmark, where he is part of the Computational Sciences and Engineering for Energy team. His work focuses on developing advanced AI and machine learning solutions for the energy industry, with particular expertise in drilling automation, reinforcement learning, deep learning, and intelligent decision support systems. His interests include autonomous drilling systems, optimization algorithms, and the application of next-generation AI technologies to enhance operational efficiency, safety, and sustainability across the energy sector. He holds a PhD in petroleum engineering from Texas A&M University and an MS from Missouri University of Science and Technology.
Aman Srivastava, SPE, is a product owner at Halliburton Landmark with more than 17 years of experience in the energy industry, specializing in well construction, drilling engineering, digital technologies, and AI. He is an inventor, published author, and recipient of two SPE Regional Awards. He is deeply involved with SPE, serving as chair of the webinars and virtual meetings program for the SPE Research and Development Technical Section, an editorial board member of the Journal of Petroleum Technology, and a frequent speaker on digital transformation, AI, and the future of energy engineering. He holds a master's degree in petroleum engineering from the University of Oklahoma and a bachelor’s degree in mechanical engineering from the National Institute of Technology, Surat.
Geetha Gopakumar Nair, SPE, is a computational physicist and data science leader with more than 2 decades of international research. She is the global lead for the Data Science Center of Excellence at Halliburton, she leads the development and application of AI and machine learning solutions that address complex challenges in the energy industry while driving collaboration with global operators and technology teams. She previously held research faculty positions at the University of Tokyo and Tokyo Metropolitan University. She has authored numerous peer-reviewed publications and is passionate about advancing data-driven innovation through research, education, and mentorship. A frequent speaker at technical conferences and workshops, Nair is recognized for her ability to bridge research with practical industry applications, inspiring the next generation of scientists, engineers, and AI practitioners.