Numerous geochemical approaches have been proposed to ascertain if methane concentrations in groundwater, [CH4], are anomalous, i.e., migrated from hydrocarbon production wells, rather than derived from natural sources. We propose a machine-learning model to consider alkalinity, Ca, Mg, Na, Ba, Fe, Mn, Cl, sulfate, TDS, specific conductance, pH, temperature, and turbidity holistically together. The model, an ensemble of sub-models targeting one parameter pair per sub-model, was trained with groundwater chemistry from Pennsylvania (n=19,086) and a set of 16 analyses from putatively contaminated groundwater. For cases where [CH4] ≥ 10 mg/L, salinity- and redox-related parameters sometimes show that CH4 may have moved into the aquifer recently and separately from natural brine migration, i.e., anomalous CH4. We applied the model to validation and hold-out data for Pennsylvania (n=4,786) and groundwater data from three other gas-producing states: New York (n=203), Texas (n=688), and Colorado (n=10,258). The applications show that 1.4%, 1.3%, 0%, and 0.9% of tested samples in these four states, respectively, have high [CH4] and are ≥50% likely to have been impacted by gas migrated from exploited reservoirs. If our approach is indeed successful in flagging anomalous CH4, we conclude that: i) the frequency of anomalous CH4 (# flagged water samples / total samples tested) in the Appalachian Basin is similar in areas where gas wells target unconventional as compared to conventional reservoirs, and ii) the frequency of anomalous CH4 in Pennsylvania is higher than in Texas + Colorado. We cannot, however, exclude the possibility that differences among regions might be affected by differences in data volumes. Machine learning models will become increasingly useful in informing decision-making for shale gas development.
All Science Journal Classification (ASJC) codes
- Environmental Engineering
- Civil and Structural Engineering
- Ecological Modeling
- Water Science and Technology
- Waste Management and Disposal