Training a Convolutional Neural Network to Locate and Quantify CH4 Leaks at Well Pads with High-Resolution Atmospheric Simulations for ARPA-E Field Tests

Fugitive methane (CH4) leaks from oil and gas production fields are a potential significant source of atmospheric methane. US DOE's ARPA-E is supporting research to locate small (< 1 scfh) fugitive methane leaks at 10 m size well pads to within 1 m so they can be isolated and fixed. An Aeris Technologies, LANL, Planetary Science Institute and Rice University team is developing an autonomous leak detection system (LDS) employing a compact rugged inexpensive laser absorption methane sensor, a sonic anemometer and multiport sampling. The LDS system is trained via a convolutional neural net (cNN) using ultra-high-resolution simulations of methane transport provided by LANL's coupled atmospheric transport model HIGRAD, for numerous controlled methane release scenarios and methane sampling configurations under variable atmospheric conditions. cNN learning is well suited to problems in which the training and observed data are noisy, or correspond to complex sensor data as is typical of meteorological and sensor data over a wellpad. Neural nets have been shown to achieve higher accuracy with more efficiency than other inverse modeling approaches in studies at larger scales, in urban environments over short time scales, and at small spatial scales for efficient source localization of indoor airborne contaminants. Our cNN is intended to characterize fugitive leaks rapidly, leading to a minimum time-to-detection and providing a first order improvement with respect to overall minimization of methane loss. Recent studies with our cNN on a variety of source and sensor location scenarios and different wind fields are discussed. Various filters that have been applied to the data (winds, measured methane concentrations) are compared for their relative success at minimizing error in locating and quantifying methane leaks. Length of multiport duty cycle sampling and flush time as well as number and placement of monitoring sensors can significantly impact ability to locate leaks. With a sufficient training base (about 30 cases), source location error is better than 10%, and usually much better. Quantifying source strength is even more successful, generally to within 1-2 %. Our objective is to transport our trained cNN code onto a Beaglebone microcomputer and integrate it with the LDS for field testing for ARPA-E.