A MACHINE LEARNING APPROACH FOR WELL INTEGRITY PREDICTION USING CEMENT BOND LOGS

Oil and gas wells with compromised cement can permit unwanted migration of hydrocarbons and brine in the subsurface or into the atmosphere, with associated potential environmental and/or human health risk. Such wells also represent a potential pathway for fluid migration from other engineered subsurface operations such as geologic carbon storage. Acoustic logging (i.e., cement bond logging) is a widely used technique that indirectly measures the quality of the cement bond along the wellbore by emitting an acoustic vibration that then travels through the drilling mud, casing, cement, and surrounding formation, before reflecting back to a receiver on the same assembly. The acoustic wave is dampened by thick, well-formed cement, and undampened by free space in the annulus, allowing a user to interpret the quality of cement from these acoustic readings. Operators in Colorado are required to submit cement bond logs (CBLs) with well permitting information; the Colorado Oil and Gas Conservation Commission (COGCC) makes these CBLs publicly available. In this study, we use a machine learning approach to make region-scale forecasts of oil and gas well integrity using only cement bond log data. We gathered a large set of CBLs from the wells in the Wattenberg Field of Colorado that have exhibited gas leak in their outermost annulus (i.e., sustained casing pressure, or SCP) indicative of well integrity loss. We characterized acoustic amplitude measurements from each well over the length of the well, and over target oil and gas bearing formations in the region, to produce statistical representations of the wells’ cement quality. Using these statistical parameters, an ensemble decision tree model was trained to forecast SCP occurrence. This modeling approach shows potential to understand the relationship between CBL measurements and SCP occurrence.