A First Step for Fracturing: Successful Real Time Brain Calibration Lays the Foundation for Predicting Fracturing Outcomes

Abstract

Abstract Fracturing utilizes significant heuristics and biases in their execution. A pressure history match, changing simulator inputs to match actual field test, is limited by rig time and capacity of a human brain. A virtual agent or brain can quickly iterate across small step sizes for a wide range of inputs to find optimal real time solutions. The ability to match pressures in real time will allow time and cost savings on minifrac tests and reverse outs. More importantly, the virtual brain can learn from past mistakes, get better over time (reinforcement learning), never forget past tests, and compare pressure signatures from vast historical data. This would be the foundational step for predictive ability on frac treatments that would eventually automate operational decision making. This also allows for consistent execution and serves as a design aid in service of building enduring capability for completion engineers. This agile project is currently the only reinforcement learning/machine learning project that interfaces with the reservoir within Wells at bp. A reward function that uses root mean square error between the brain's prediction and actual field data is used as the basis to reward or punish the brain i.e., reinforcement learning. Blind training of the brain was utilized on several deployments in the field. This led to a better understanding of the data engineering requirements to feed real time data to the virtual machine. As a result of the deployment, two brains in series are being run, one for pad fluid and one for frac fluid with proppant. By obtaining a real time calibration with pad, the cost of diagnostic tests is eliminated. More importantly, the brain can significantly shorten the learning curve for new engineers and help avoid typical execution mistakes. Finally, the vast memory capacity of an agent and its ability to compare various pressure traces in short time frames can eventually be used for predictive ability as the technology matures. This project is acknowledged as a groundbreaking project within bp and industry. The concept can be replicated at scale to various other petroleum engineering operations given the vast amount of historical data available in various disciplines.