Think of geosteering as a drill with brains.
Rather than drilling an unvaryingly vertical borehole into the Earth, in hopes of hitting oil, petroleum companies can now adjust the drill’s direction in real time based on geological logging measurements. The drill can even bore horizontally when seeking a “pay zone” – oil-speak for a deposit. The process relies on accurate data analysis.
“We can drill more accurately and take less time, and time is money in this business. We can log while drilling, and the people at Rice are helping us make that possible,” said Neilkunal Panchal, a lead research scientist with Shell.
Panchal is working with a team from Rice University, Purdue University and Shell to streamline geosteering with machine-learning algorithms. “In the past, the sensor readings we got were very noisy and very erratic. They weren’t reliable. We want to clean them up and make the information more dependable,” said Daniel Kowal, assistant professor of STAT.
“Oil companies routinely collect large sensors and drilling mechanical datasets from the rigs around the world. Many statistical explorations remain to be done, and some of them are crucial to make safe and smart decisions while drilling,” said Marina Vannucci
, Noah Harding Professor and STAT department chair at Rice.
Other members of the research group include Yinsen Miao, a fourth-year graduate student in STAT at Rice, whose internship at Shell brought the team together; Faming Liang, professor of STAT at Purdue University; and Mingqi Wu and Jeremy Vila with Shell.
“Human-based data processing and interpretation are easily affected by individual opinions and prone to various errors. Therefore, statistical machine learning algorithms are needed to avoid the individual variances and standardize the geosteering process,” Vannucci said.
At the start of their two-year industry/academia project, the research group devised a Bayesian machine learning approach based on state-space modeling. Next they will test the accuracy of the methods they have developed on real drilling data and investigate other non-linear state-space model solutions. They will test the models using a Python-based multi-thread GUI (graphical user interface) simulation software.
Shell will eventually test the result at their fields in West Texas, Argentina and Canada.
“Many projects don’t reach this state of effectiveness. The results we’ve seen are very promising. The algorithms we have developed at Rice are working. We’re happy,” Panchal said.