Sensor Location-Invariant Multi-Regime Methane Leak Localization and Quantification with Knowledge-Guided Deep Learning

Quick, reliable, and precise source methane emission localization and quantification can help reduce upstream oil and gas operational costs and enable a reduction of greenhouse gas emissions associated with oil and gas production activities. Our primary goal is to create a community based, open-source software package of machine learning (ML)-based inversion algorithms ingesting local scale meteorology (anemometer) and CH4 continuous monitoring instrument observations to accurately localize and quantify emissions. The inversion algorithms are intended to generalize across arbitrary site infrastructure configurations, sensor types and sensor counts, time-series lengths, as well as varying climatologies. Two ML-based algorithmic approaches are explored including a relatively simple ML backtracking multilayer neural network approach and a more sophisticated pure ML approach that uses a masked encoder-decoder transformer architecture with multi-head attention. Training of the ML models leverages thousands of turbulence-resolving large-eddy simulation (LES) scenarios carried out using NCAR’s GPU-accelerated FastEddy® spanning a wide variety of meteorological conditions, run at a spatial resolution of 50 centimeters and with results of virtual instrument platforms extracted at a frequency of 1Hz. This presentation provides a summary of favorable characteristics towards robust generalization along with training, verification and skill assessment using both the purely virtual LES simulated emissions scenario cases and controlled release field observations.