The Silicon Valley team training AI to prevent another Deepwater Horizon oil spill
Tehmina Amjad and a team of graduate computer science students in Silicon Valley have set out to address one of the critical root causes of accidents like the Deepwater Horizon spill.
OAKLAND – Devastating images from the Deepwater Horizon oil spill in the Gulf of Mexico serve as a powerful warning 16 years after the record-setting environmental disaster.
Photos of the BP-operated offshore drilling rig emitting a plume of smoke, wildlife stained with slick, onyx-black oil and the miles-long streak of escaped oil discoloring the ocean continue to be haunting reminders of the risks associated with drilling. A “blowout” explosion killed 11 people and the rig eventually collapsed, allowing 210 million gallons of oil to leak across the seabed over nearly three months in what became the largest marine oil spill in history.
A 2011 report by the U.S. Coast Guard and Bureau of Ocean Energy Management, Regulation and Enforcement determined that a central cause of the catastrophe was a failure of a cement barrier in the offshore well, which resulted in an upflow of crude oil to the rig and, ultimately, an explosion. A White House commission at the time stated that without major reforms to the drilling industry, a similar disaster could occur.
Despite the years since the April 20, 2010, spill and the federal report, the risk remains high, according to experts in the field.
“This is actually a very serious problem, and the consequences can be very dangerous,” Tehmina Amjad, a professor in the Khoury College of Computer Sciences on Northeastern University’s Silicon Valley campus, said in a recent interview.
As part of the college’s Research Runway program, Amjad and a team of graduate computer science students have set out to address one of the critical root causes of accidents like the Deepwater Horizon spill.
Using artificial intelligence and data from oil wells around the world, her team is attempting to optimize the ability for AI models to predict the pressure of fluid trapped thousands of feet underground. The metric is considered crucial for the safe and efficient extraction of oil and gas, she said.
By more accurately predicting the subsurface geopressure conditions of new well sites based on real and synthetically-generated data, the algorithms could help drilling companies save time and resources as they plumb the earth for resources. And in the most dire cases, they could act as a safeguard against devastating spills, Amjad said.
Drilling companies perform a delicate balancing act when managing the pressure of fluid trapped in the pore spaces deep beneath the surface. Engineers must pump a precise amount of heavy mud down holes during the extraction process, creating enough pressure to prevent natural oil and gas from entering the hole too soon, but not so much pressure that it fractures the surrounding rock, Amjad said.
“Higher fluid pressures within the rocks necessitates a proportional increase in drilling pressure inside the well to maintain a pressure balance for better well control,” she said.
If engineers miscalculate pore pressure, they could lose valuable fluid, destabilize the well or, in the worst case, cause a buildup inside the hole that leads to a dangerous blowout.
According to existing research cited by Amjad and her team, oil and gas fields around the world are overpressured in varying magnitudes, and unanticipated pressure anomalies during drilling can lead to significant operational delays, increased non-productive time and additional drilling costs.
AI, or machine learning, models fed with the proper geological data can already calculate pore pressure in specific areas, Amjad said. They cannot yet be trusted to give accurate recommendations in new areas – a particular issue for companies that drill in multiple locations or countries.
“These algorithms need data to learn, but real drilling data is expensive to acquire,” she said. “The more data we can provide to these algorithms, the more they learn about a particular area’s geology.”
Amjad and the team of four graduate students took it upon themselves to think up a two-part solution: using machine learning to make informed generalizations about pore pressures based on real data and, when necessary, creating synthetic well data via a homemade Python tool called WellBench.
Editor’s Picks
Their data-generalization project uses data of hydrocarbon wells in Pakistan, acquired with help from Amjad’s brother, Muhammed Raiees Amjad, associate professor at Barhia University in Islamabad, Pakistan. The logs contain subsurface information on the physical properties of rocks, including their radioactivity, electrical resistivities, density and porosities.
Amjad and her team believe a predictive algorithm can draw from this information to make recommendations for engineers to design safe, secure drilling plans for new wells in the same region as existing ones.
“If the wells are three or four miles apart, and they’re going to dig a new well in between them, our model knows the data from the surrounding wells and should be very good in predicting the pressure in that area,” she said.
Tiago Monteiro, who is pursuing his master’s degree in data science and AI at Northeastern’s Silicon Valley campus, is tackling the team’s second project. The Portugal native created WellBench, a reusable tool in Python which can be used to generate synthetic well data to supplement the lack of available data in some drilling areas, he said.
“It doesn’t matter where we are,” Monteiro said. “If we are in China or India, we can generate synthetic data related to that region so that we can quickly create machine learning models that can predict the safest path to drill.”
WellBench is able to come up with accurate predictions by using interpolation methods, Monteiro said. That involves a machine learning process in which an algorithm estimates unknown or missing data points that lie within the range of a known dataset.
With additional contributions from researchers Pranav Patel, Ritik Mahyavanshi and Ramyashree, Amjad hopes the team’s machine learning applications can someday fill the gap in available well data and ensure safer and more accurate drilling.
“Geological conditions are not the same everywhere. We need to be very careful when we are working in any area,” she said. “If we can achieve our goal, it will be a very good thing.”
