Climate change is one of the greatest challenges facing humanity, and we, as machine learning (ML) experts, may wonder how we can help. Here we describe how ML can be a powerful tool in reducing greenhouse gas emissions and helping society adapt to a changing climate. From smart grids to disaster management, we identify high impact problems where existing gaps can be filled by ML, in collaboration with other fields. Our recommendations encompass exciting research questions as well as promising business opportunities. We call on the ML community to join the global effort against climate change.

Methane (CH4) emissions from oil and natural gas (O&NG) systems are an important contributor to greenhouse gas emissions. In the United States, recent synthesis studies of field measurements of CH4 emissions at different spatial scales are ~1.5–2× greater compared to official greenhouse gas inventory (GHGI) estimates, with the production-segment as the dominant contributor to this divergence. Based on an updated synthesis of measurements from component-level field studies, we develop a new inventory-based model for CH4 emissions, for the production-segment only, that agrees within error with recent syntheses of site-level field studies and allows for isolation of equipment-level contributions. We find that unintentional emissions from liquid storage tanks and other equipment leaks are the largest contributors to divergence with the GHGI. If our proposed method were adopted in the United States and other jurisdictions, inventory estimates could better guide CH4 mitigation policy priorities.

Climate change is one of the greatest challenges facing humanity, and we, as machine learning experts, may wonder how we can help. Here we describe how machine learning can be a powerful tool in reducing greenhouse gas emissions and helping society adapt to a changing climate. From smart grids to disaster management, we identify high impact problems where existing gaps can be filled by machine learning, in collaboration with other fields. Our recommendations encompass exciting research questions as well as promising business opportunities. We call on the machine learning community to join the global effort against climate change.

Limiting emissions of climate-warming methane from oil and gas (O&G) is a major opportunity for short-term climate benefits. We deploy a basin-wide airborne survey of O&G extraction and transportation activities in the New Mexico Permian Basin, spanning 35 923 km2, 26 292 active wells, and over 15 000 km of natural gas pipelines using an independently validated hyperspectral methane point source detection and quantification system. The airborne survey repeatedly visited over 90% of the active wells in the survey region throughout October 2018 to January 2020, totaling approximately 98 000 well site visits. We estimate total O&G methane emissions in this area at 194 (+72/–68, 95% CI) metric tonnes per hour (t/h), or 9.4% (+3.5%/–3.3%) of gross gas production. 50% of observed emissions come from large emission sources with persistence-averaged emission rates over 308 kg/h. The fact that a large sample size is required to characterize the heavy tail of the distribution emphasizes the importance of capturing low-probability, high-consequence events through basin-wide surveys when estimating regional O&G methane emissions.

Methane leakage from point sources in the oil and gas industry is a major contributor to global greenhouse gas emissions. The majority of such emissions come from a small fraction of “super-emitting” sources. We evaluate the emission detection and quantification capabilities of Kairos Aerospace’s airplane-based hyperspectral imaging methane emission detection system for methane fluxes of 18–1,025 kg per hour of methane (kgh(CH4)). In blinded controlled releases of methane conducted over 4 days in San Joaquin County, CA, Kairos detected 182 of 200 valid nonzero releases, including all 173 over 15 kgh(CH4) per meter per second (mps) of wind and none of the 12 nonzero releases below 8.3 kgh(CH4)/mps. Nine of the 26 releases in the partial detection range of 5–15 kgh(CH4)/mps were detected. There were no false positives: Kairos did not detect methane during any of the 21 negative controls. Plume quantification accuracy depends on the wind measurement technique, with a parity slope of 1.15 (σ = 0.037, R2 = 0.84, N = 185) using a cup-based wind meter and 1.45 (σ = 0.059, R2 = 0.80, N = 157) using an ultrasonic anemometer. Performance is comparable even with only modeled wind data. For emissions above 15 kgh/mps, quantification error scales as roughly 30%–40% of emission size, even when using wind reanalysis data instead of ground-based measurements. This reflects both uncertainty in wind measurements and in Kairos’ estimates. These findings suggest that at 2 mps winds under favorable environmental conditions in the United States, Kairos could detect and quantify over 50% of total emissions by identifying super-emitting sources.

Methane detection limits, emission rate quantification accuracy, and potential cross-species interference are assessed for Bridger Photonics’ Gas Mapping LiDAR (GML) system utilizing data collected during laboratory testing and single-blind controlled release testing. Laboratory testing identified no significant interference in the path-integrated methane measurement from the gas species tested (ethylene, ethane, propane, n-butane, i-butane, and carbon dioxide). The controlled release study, comprised of 650 individual measurement passes, represents the largest dataset collected to date to characterize GML with respect to point-source emissions. Binomial regression is utilized to create detection curves illustrating the likelihood of detecting an emission of a given size under different wind conditions and for different flight altitudes. Wind-normalized methane detection limits (90% detection rate) of 0.25 (kg/h)/(m/s) and 0.41 (kg/h)/(m/s) are observed at a flight altitude of 500 feet and 675 feet above ground level, respectively. Quantification accuracy is also assessed for emissions ranging from 0.15 to 1,400 kg/h. When emission rate estimates were generated using wind from high-resolution rapid refresh (HRRR) model (the primary wind source that Bridger uses for their commercial operations), linear regression indicates bias of 8.1% (R2 = 0.89). For 95% of controlled releases above Bridger’s stated production-sector detection sensitivity (3 kg/h with 90% probability of detection), the accuracy of individual emission rate estimates produced using HRRR wind ranged from −64.1% to +87.0%. Across all controlled releases, 38.1% of estimates had error within ±20%, and 87.3% of measurements were within a factor of two (−50% to +100% error). At low wind speed (less than 2 m/s) and low emission rates (less than 3 kg/h), emission estimates are biased high, however when removed do not impact the regression significantly. The aggregate quantification error including all detected emission events was +8.2% using the HRRR wind source. The resulting detection curves and quantification accuracy illustrate important implications that must be considered when using measurements from GML or other remote emission measurement techniques to inform or validate inventory models or to audit reported emission levels from oil and gas systems.

Methane (CH4) from oil and gas (O&G) activities is a known contributor to global anthropogenic methane emissions and recent research has demonstrated that a small fraction of large emitters contribute to the majority of total emissions. In this study, we perform a single-blind evaluation of the quantification capabilities of three airplane-based technologies (Bridger Photonics’ Gas Mapping LiDAR, Carbon Mapper’s Global Airborne Observatory, and GHGSat-AV) with a focus on large emitters (10-2,000+ kg h-1 CH4). In two 2021 campaigns, metered natural gas was released concurrently with overpasses by the tested technologies. Results were submitted by operators in a three-stage unblinding process. All teams detected 100% of releases above 50 kg h-1 CH4. The teams report parity slopes of 0.35 to 1.06, with R2 values of 0.35 to 0.78. After 10-meter anemometer wind measurements were unblinded, two out of three teams significantly reduced variance in the parity slope, highlighting the importance of accurate wind data. After half of metered release volumes were subsequently unblinded, improvement was mixed. These results suggest that multiple commercially available technologies can reliably detect larger point-source methane emissions, with varying quantification performance and trade-offs between survey area coverage and instrument sensitivity.

Satellites are increasingly seen as a tool for identifying large greenhouse gas point sources for mitigation, but independent verification of satellite performance is needed for acceptance and use by policy makers and stakeholders. We conduct to our knowledge the first single-blind controlled methane release testing of satellite-based methane emissions detection and quantification, with five independent teams analyzing data from one to five satellites each for this desert-based test. Teams correctly identified 71% of all emissions, ranging from 0.20 [0.19, 0.21] metric tons per hour (t/h) to 7.2 [6.8, 7.6] t/h. Three-quarters (75%) of quantified estimates fell within ± 50% of the metered value, comparable to airplane-based remote sensing technologies. The relatively wide-area Sentinel-2 and Landsat 8 satellites detected emissions as low as 1.4 [1.3, 1.5, 95% confidence interval] t/h, while GHGSat’s targeted system quantified a 0.20 [0.19, 0.21] t/h emission to within 13%. While the fraction of global methane emissions detectable by satellite remains unknown, we estimate that satellite networks could see 19–89% of total oil and natural gas system emissions detected in a recent survey of a high-emitting region.