In the rugged and geologically complex terrain of Sichuan and Chongqing, where the earth’s crust tells a story of ancient upheavals and hidden fractures, the oil and gas industry has long battled a formidable, invisible foe: lost circulation. This phenomenon, where vital drilling fluid vanishes into the depths of the earth, is not merely an operational hiccup; it is a multi-million dollar problem that halts progress, risks catastrophic downhole accidents, and turns ambitious exploration projects into exercises in costly frustration. For years, the response has been reactive, guided by the intuition of seasoned engineers and the hard-won lessons of past failures. But intuition, no matter how sharp, is no match for the sheer randomness and complexity of subsurface fractures. The game needed to change, and it has. A new era of intelligent, data-driven well control has dawned, spearheaded by a team of researchers who have turned petabytes of chaotic drilling data into a powerful predictive weapon.
The story begins not with a eureka moment in a laboratory, but with a persistent, nagging problem on the drilling floor. In the Sichuan-Chongqing region, a cornerstone of China’s “13th Five-Year Plan” for energy production, the geological deck is stacked against drillers. The formations are notoriously friable, riddled with natural and induced fractures that act like subterranean sponges, greedily soaking up drilling mud. The consequences are dire: projects hemorrhage time and money, with some wells requiring multiple, expensive attempts to plug a single leak. The success rate for the first plugging attempt languished at a dismal 39.1%, a statistic that spoke volumes about the industry’s reliance on experience-based, trial-and-error methods. It was clear that a paradigm shift was needed—from reactive firefighting to proactive, intelligent prediction.
Enter Deng Zhengqiang, Lan Taihua, Lin Yangsheng, He Tao, Huang Ping, Luo Yufeng, Wang Jun, and Xie Xiantao from the Drilling Fluid Technology Service Co., Ltd., CNPC Chuanqing Drilling Engineering Company Limited. Their mission was audacious: to harness the power of big data and artificial intelligence to not only understand why wells leak but to predict where and when they will leak before the drill bit even gets there. The result of their labor is nothing short of revolutionary: an “Intelligent Assistant Decision Making Platform for Lost Circulation Prevention and Control.” This is not a simple database or a glorified spreadsheet; it is a sophisticated, self-learning brain for drilling operations, designed to turn historical failure into future success.
The platform’s foundation is an ocean of data—2.4 gigabytes of it, to be precise. The team didn’t just collect data; they went on a digital archaeological dig, painstakingly extracting 2.1 million data points from 2,796 disparate Excel sheets and other electronic documents spanning 240 wells. This raw data was a chaotic mess—noisy, inconsistent, and riddled with gaps. Before any AI could make sense of it, it had to be cleaned, structured, and normalized. Using Python, they built custom tools to migrate this mountain of information into a structured MySQL database, creating a unified, 1.4 million-record repository. They then applied mathematical normalization, scaling 9 critical input parameters to a standard range, ensuring that no single variable could skew the AI’s learning process. This meticulous data preparation, often the unsung hero of any AI project, was the crucial first step in transforming anecdotal experience into actionable intelligence.
The next challenge was teaching the machine to think. With 23 potential input parameters—from rock type (YX) and bit size (ZTCC) to pump pressure (BY1) and mud density (ZJJMD)—the team needed to identify which factors truly mattered. Feeding all 23 into a model would create unnecessary noise and computational bloat. So, they employed statistical correlation analysis using IBM’s SPSS software. The results were illuminating, stripping away irrelevant variables and pinpointing the 20 most critical predictors of a lost circulation event. This wasn’t guesswork; it was data-driven triage, ensuring the AI’s brain was focused only on the signals that mattered.
The heart of the platform is its predictive engine, built using advanced machine learning algorithms. The team tested four powerful models: Neural Networks, Decision Trees, Random Forest, and Support Vector Machines (SVM). Their task was singular: predict the exact depth of a future thief zone. The results were put to the test, with each model’s predictions plotted against actual thief zone depths. While the SVM model showed a near-perfect linear fit, a deeper dive into the error metrics revealed a critical insight. The Neural Network, despite a slightly less perfect fit, demonstrated superior stability and lower prediction error (as measured by RMSE). In the high-stakes world of drilling, where a small error can mean millions in losses, stability trumps theoretical perfection. The team made the bold, data-backed decision to go with the Neural Network, prioritizing robust, reliable predictions over aesthetically pleasing graphs.
But prediction is only half the battle. To make the platform truly intelligent, it needed to understand context and similarity. This is where the concept of “clustering” came into play. Imagine being able to group thousands of wellbore sections based on their pre-leak behavior—similar weight on bit, similar lithology, similar torque readings. By applying the K-means clustering algorithm, the platform can identify these “sibling intervals.” If interval A in Well 1 leaked under specific conditions, and interval B in Well 2 is exhibiting those exact same conditions, the platform can sound an alarm: “High risk of leakage here.” It’s like having a digital forecaster who has seen every storm that’s ever happened and can now predict the next one with uncanny accuracy.
The final piece of the puzzle was turning these predictions and clusters into actionable advice. This required a different kind of AI magic: association rule mining. The team used a modified version of the Apriori algorithm, a technique often used in retail to say, “Customers who bought diapers also bought beer.” In this case, the rules were: “In intervals with lithology X and weight on bit Y, the most effective plugging method was Z, using formula A, B, and C.” The modified Apriori algorithm was crucial because the standard version generates too many redundant rules and is computationally slow. The team’s improved version was leaner, faster, and more precise, capable of sifting through the clustered data to find the golden nuggets of plugging wisdom hidden within.
The platform that emerged is a marvel of modern engineering. It’s not a black box; it’s a transparent, user-friendly console that integrates seamlessly with existing drilling data systems. Engineers can log in, pull up real-time data from an actively drilling well, and within moments, receive a comprehensive risk assessment. The system doesn’t just say, “You might have a leak.” It says, “At 5,354 meters, in the Emeishan Basalt formation, you have an 80% probability of a loss event. The predicted thief zone width is X millimeters. The recommended plugging strategy is bridge plug with a 20 cubic meter slug of 2.05 g/cm³ mud, composed of 3-6% Sui Du, 6-10% rigid particles, and so on.” It’s a level of prescriptive, intelligent guidance that was simply unimaginable a decade ago.
The true test of any technology is in the field, and the results are compelling. The team rigorously tested the platform on five completed wells and three actively drilling wells. For the completed wells, the plugging plan recommended by the AI matched the plan actually used by engineers 60% of the time. For the actively drilling wells, where conditions are more dynamic and unpredictable, the match rate was still an impressive 50%. These aren’t just numbers; they are proof of concept. A 50-60% alignment means the AI’s recommendations are not wild guesses but grounded, valuable insights that engineers can trust and build upon. One case study, the MX023-H1 well, is particularly telling. The AI accurately predicted an 80% loss risk in the treacherous Emeishan Basalt. When the leak occurred, it recommended a specific bridge plug formula. The field engineers, using their judgment, tweaked the formula slightly, resulting in a 60% match. The leak was successfully controlled. This is the perfect synergy of human expertise and machine intelligence.
The most significant metric, however, is the bottom line: success rate. Before the platform, the region’s first-time plugging success rate was a paltry 39.1%. After deploying the AI assistant across 17 well operations in the shale gas and Gaomo blocks, that number jumped to 52.9%—a staggering 13.8 percentage point increase. This translates directly into saved days of rig time, millions of dollars in avoided costs, and a dramatic reduction in the risk of downhole accidents. It’s a testament to the platform’s ability to turn data into dollars and safety.
This innovation is not an endpoint; it’s a beginning. The team openly acknowledges that the 50% match rate for actively drilling wells indicates room for growth. The AI, like any learner, needs more data. As more wells are drilled and more loss events are recorded, the platform’s neural networks will continue to learn, its clusters will become more refined, and its association rules will grow ever more precise. The dream is a system that doesn’t just assist but anticipates and prevents, a true guardian angel for drilling operations.
The implications of this work ripple far beyond the Sichuan-Chongqing basin. It represents a blueprint for how the entire global oil and gas industry can leverage AI to tackle its most persistent, costly problems. It proves that with the right data, the right algorithms, and the right team, even the most chaotic, unpredictable challenges can be brought under control. It’s a victory not just for efficiency, but for safety and sustainability, reducing the environmental footprint of drilling by minimizing fluid loss and non-productive time.
In a world increasingly driven by data, the story of this intelligent platform is a powerful reminder that the future of heavy industry lies not in replacing human ingenuity, but in augmenting it. The engineers on the rig floor are still the heroes, making the final calls and wielding the tools. But now, they have a powerful new ally in their corner—an AI that has learned from every past mistake and is dedicated to ensuring those mistakes are not repeated. It’s a partnership forged in data, tested in the field, and destined to reshape the future of drilling.
This groundbreaking research, “Research and application of intelligent assistant decision making platform of lost circulation prevention and control in Sichuan-Chongqing area,” was conducted by Deng Zhengqiang, Lan Taihua, Lin Yangsheng, He Tao, Huang Ping, Luo Yufeng, Wang Jun, and Xie Xiantao from the Drilling Fluid Technology Service Co., Ltd., CNPC Chuanqing Drilling Engineering Company Limited, Chengdu, Sichuan 610056, China. It was published in the journal Oil Drilling & Production Technology, Volume 43, Issue 4, July 2021, pages 461-466. The article can be identified by its DOI: 10.13639/j.odpt.2021.04.008.