During the 2018 expedition, acoustic surveys were conducted from the research vessel and the ROV. Ship-based echosounders were installed on an overboard pole that was oriented vertically during survey operations. Sensors included an HiPAP Model 350 Ultra-Short Baseline transducer (USBL) for tracking the ROV and beacons, a 300 kHz Teledyne RDI Workhorse Acoustic Doppler Current Profiler (ADCP), a 500 kHz Kongsberg Mesotech M3 MBES, and three frequencies of Simrad EK80 SBES transducers operating at 70, 120, and 200 kHz. A Teledyne DMS-25 was installed to account for ship motion. Acoustic mapping surveys were conducted daily to account for changing currents and consequent orientation of hydrocarbon plume in the water column (Supplementary Figure S1A). Initial passes were interpreted to establish the main axis of the plume, and subsequent passes were run parallel to that axis. Mapping of the plumes was conducted with the ROV with the use of Simard EK80 and Kongberg M3 sonars to localize and delineate separate components of the hydrocarbon plume and their sources on the seabed. Details of the calibration and data processing procedures can be found in the work of Taylor and Boswell (2019).

Water samples analyzed for methane concentration were collected in Niskin bottles and, to a lesser extent, using methods described in Collection and Imaging of Gas Bubbles. During the 2018 samplings, the bottles were deployed on a rosette and triggered at preset depths. The rosette casts were deployed over the site of the erosion crater while monitoring the acoustic signature of the hydrocarbon plume and attempting to maintain the rosette within the plume. Rosette casts were lowered to a maximum depth of 120 m to avoid entanglement with the platform jacket and midwater gear suspended above the jacket. Additional water samples were obtained from the bubblometer pressure chambers (described below). These samples were sealed at depth and were collected when oil bubbles were observed passing through the device.

During the 2021 samplings, the Niskin bottles were deployed from an ROV and triggered by its manipulator arm. The containment system, which was installed in May 2019, encloses the plume sources under a dome suspended approximately 3 m above the bottom. The contained hydrocarbon plumes are channeled into a patented separator system (Couch, 2010), through which oil is passively diverted into underwater storage tanks, while gas bubbles are continually released into the water at a depth of 123 m. These tanks are periodically emptied in a pump-off procedure that transfers the oil to storage tanks on the M/V Brandon Bordelon. Observations in 2021 and 2022 were cruises of opportunity accommodated by these operations. For the ROV sampling, the bubble plume was readily detected in the ROV scanning sonar, allowing the vehicle to maneuver to where bubbles were visible for water collections. Samples were generally collected in replicate pairs. Water samples were stored in gas-tight vials at 4 C. Methane concentrations were determined with the use of a GC coupled to a flame ionization detector, Shimadzu 8a packed carbosphere column, 140 C oven, and detector at 180 C. Standards were obtained from Restek with accuracy ±5% and precision ±1%.

A custom device called the “bubblometer” was fabricated to collect oil and gas samples in the water column and record digital image samples of the bubble plume for quantification (Figure 2). Its major component was an inverted funnel mounted atop a 30 × 30 cm-wide, 20 cm-high, three-sided visualization chamber, which was open at its bottom and at the side facing the ROV. The funnel was fed into a 300 ml acrylic collection tube that could accumulate water plus oil and gas. Four pressure cylinders, evacuated and sealed at the surface, were plumbed into the collection cylinder and could be filled individually by hydrostatic pressure when their respective valves are opened. Closing these valves then sealed the samples and prevented methane from degassing during ascent. A digital video camera (Deepsea Power and Light model HD Multi Seacam™) with two high-intensity lamps (Deepsea Power and Light Sealite™ 2,300 lumen) was mounted 30 cm from the rear opening of the visualization chamber. The entire device including camera and lamps was fixed to a frame with a hydraulic actuator. This arrangement allowed it to be extended from the front of the ROV so that the bubbles could flow through the bottom of the chamber for imaging or sample collection, or retracted to block the flow. Buoyant oil and gas would pass through the visualization chamber and be funneled into the collection tube. The camera and lamps on the bubblometer allowed observers to monitor bubbles passing through the chamber, while the separate video feed from the ROV allowed them to watch oil and gas displacing seawater at the top of the collection tube and determine when to open a valve and collect a sample.

Output from the bubblometer camera was monitored from the ship while the ROV navigated toward the MC20 hydrocarbon plumes. The ROV approached the bubble plumes with its sonar system until bubbles were observed in the camera, and then, thrusters were secured and drifted until a plume had been traversed and bubbles were no longer visible. A digital video was recorded when bubbles were observed passing through the chamber. Sample frames were subsequently captured at 5 s intervals from the recorded video. The camera was mounted at a fixed distance from the chamber, so images had a constant scale, but were cropped to a constant size of 1971 × 1,173 pixels, which showed only the interior of the chamber (Figure 3). In total, 665 individual image samples were collected from the video records during two lowerings of the ROV on 5 and 6 September 2018, respectively.

Post-containment estimates of bubble size for gas released from the separator were obtained from the video of bubbles passing in front of a panel with scale markings. This material was collected from the MV Brandon Bordelon ROV in January 2022. Frame captures were taken at intervals when visibility permitted, and bubbles were measured in comparison to the scale markings using Image-J.

Samples of gas captured in the water column by using the bubblometer and collected into pressure vessels at depth were transferred on the deck to evacuate foil gas bags with a valve containing a septum to allow subsampling. Triplicate, 10 μl aliquots of each gas bag sample were injected via a gas-tight syringe onto a HS-GC/FID (HP 5890) configured with a porous layer open tubular (PLOT) column to separate and quantitate the C₁–C₅ light hydrocarbon gases. Multiple injections of each gas bag sample were introduced into the GC/IRMS configured with a PLOT column to separate the C₁–C₅ hydrocarbon gases and determine their carbon isotopic ratios. A reference carbon dioxide standard (−37.5‰ versus PDB) was used to linearize the detector. An external standard containing all of the C₁–C₅ analytes of interest with known carbon isotopic ratios was used to verify the PDB accuracy +/1 1‰ of the GC/IRMS (Gaskins et al., 2019).

The front of the ROV acted as a baffle that only allowed bubbles to enter the bubblometer chamber, where the camera system recorded images at constant scale and illumination. However, objects in its images appeared larger or smaller depending on their distance from the camera within the chamber. Calibration in a laboratory setting showed that the camera resolved 8 pixels/mm at the front of the chamber closest to the camera, 4 pixels/mm in the center, and 2 pixels/mm at the rear of the chamber, with no discernable distortion due to vertical position at a given distance. The single camera could not reliably determine the distance between it and an object within the chamber. It was assumed that bubble distributions within the chamber were uniformly random and a constant scale of 4 pixels/mm was used to estimate the size of bubbles, which unavoidably meant that there was a two-fold uncertainty in any estimate of bubble diameter and an ∼eight-fold uncertainty in an estimate for the volume of a spherical bubble, while most bubbles were somewhat elliptical in shape, with dimensions that tended to vary as the bubbles moved within the chamber. For these reasons, bubble sizes were estimated for confirmation of general impressions gained by comparing bubbles to adjacent objects of known size and to provide parameters for calibration of the acoustic surveying (MacDonald et al., 2019), but are not used as quantification of bubble volumes in this paper.

A neural network process called Object Detection was chosen for the detection and classification of the bubbles in the image samples and implemented with use of a Faster Region Convolutional Neural Network (Faster R-CNN) variant. The MATLAB^© Computer Vision toolbox was used for performance of Faster R-CNN classification. Both the training set and test set are taken from the ground truth set. The training set uses 70% randomly assigned images, while the remaining 30% goes to the test set. The training process completed automatically by MATLAB generated a detector that can be used on any image, or a set of images, and it identifies, classifies, and measures their targets. Figure 4 shows examples of images processed by the algorithm. The detector was tested on the remaining 30% of the images. It shows an overall accuracy of 60%, which increases to 62% for oil bubbles (Table 1). Most of the error is attributable to false positives due to either overlapping detections or detections that were not considered in the training set.

Two classes were selected for the targets: gas bubbles with a minor fraction of oil and oil bubbles assumed to be predominantly oil. A total of 68 cropped and prepared images were chosen, and bubbles were manually classified and stored as the ground truth set. Bubbles resolved with less than 8 pixels (nominally 2 mm) of radius were excluded to reduce process noise. Half bubbles, partially occluded bubbles, or bubbles too close to the black boundary were also left out of the training set for similar reasons.

The atmospheric concentration of methane was continuously measured with the use of a cavity ring-down spectrometer (CRDS) Picarro^® G2203 Analyzer for CH₄/C₂H₂ that drew air samples (4 Hz) from an intake tube located on the starboard side of the ship, 3 m above the water surface and below the level of the exhaust stacks of the vessel. The instrument was calibrated to gas standard following a three-point procedure (Piccaro, 2011). The length of the intake line introduced a 60 s lag between intake and measurement. Data were recorded during the entire time the vessels R/V Brooks McCall or M/V Brandon Bordelon were on station. Ship tracks for surveys during the 2018 and 2021 campaigns are shown in Supplementary Figure S1. Sample readings were georeferenced in real time using a GPS antenna connected to the computer of the CRDS. Identical instrumentation and collection procedures were used on the 2018 and 2021 expeditions.

During the final day of operations in the 2018 expedition, a tracer release experiment was conducted in order to test whether methane concentrations in the air in the study area could be linked to the persistent hydrocarbon plume that reached the ocean surface near the toppled MC20 oil platform and also to establish the path and dispersion of methane gas in the area. Because the CRDS-Picarro 2,203 has the capability of detecting acetylene gas with a precision in the parts per million range, the tracer experiment also utilized this gas as a reference compound. The acetylene tracer release technique has been widely used in the quantification of fugitive methane in landfills (Mønster et al., 2014; Mønster et al., 2015). To adapt this technique for the open ocean, a floating raft was constructed which held a small acetylene tank. The acetylene tank was connected to a mass flow regulator to produce a constant release of tracer gas into the atmosphere (15 L/min). In addition, a meteorological station installed on the raft collected records of the ambient wind speed and direction, temperature, and humidity in the area during the controlled acetylene release event. The raft was moored, as close as was possible, to the area where the MC20 hydrocarbon plume reached the ocean surface. Once the raft was successfully deployed, the vessel drifted downwind to increase the chances of detecting both atmospheric methane concentrations and the known volume of tracer gas. In practice, the scope of the raft anchor line allowed it to drift away from the bubble surfacing area.

Inverse plume modeling combined with atmospheric methane concentrations is used to estimate methane emission rate from the hydrocarbon plume into the atmosphere. The employed approach originally presented in landfill applications (Kormi et al., 2017; Kormi et al., 2018; Ali et al., 2020) tackles the problem of determining a contaminant source emission rate for a given set of measurements. The Gaussian plume model is coupled with an optimization-based identification method, to estimate fugitive methane emissions. Methane concentration is used to infer emissions though dispersion modeling and optimization. This is achieved through tracing dispersed methane back to potential emission sources. In the subsequent sections, we briefly summarize this optimization-based approach (Silva et al., 2019) and refer the reader to the work of Kormi et al. (2017) and Kormi et al. (2018) for a more thoroughly detailed presentation of the method.

Input parameters of the methane emission estimation method include methane concentration measurements and locations along with meteorological conditions, the most important being wind speed and direction and temperature. An implemented code is used to generate multiple configurations of source positions and emission rates. Each sample configuration is evaluated through calculating the corresponding methane concentrations at each measurement point. This is carried out through the backward application of an atmospheric dispersion model. As such, source identification can be treated as an inverse optimization task where the objective is to obtain the configuration of sources (locations and emission rates) that best fits the measured concentrations. The performance of a source configuration is further evaluated through the difference between measured and predicted methane concentrations. To predict methane concentrations at locations where effective measurements are performed, an atmospheric dispersion model is needed. In the proposed method, modeling of methane dispersion is carried out using Gaussian dispersion Eq. 1. This equation models the dispersion of a non-reactive gaseous pollutant (here, methane) from an elevated point source. Eq. 1 predicts the steady-state concentration (C) in μg/m³ at a point (x, y, z) located downwind from the source. C ( x , y , z ) = Q 2 π u σ y σ x e − y 2 2 σ y 2 ( e ( z + H ) 2 2 σ z 2 + e ( z − H ) 2 2 σ z 2 ) (1)

In Eq. 1 “Q” is the emission rate (μg/s), “σy” and “σz” (m) are the horizontal and vertical spread parameters that are functions of wind distance “x,” respectively, and atmospheric stability is a measure of the resistance of the atmosphere to vertical air motion. Continuing, “u” is the average wind speed at stack height (m/s), “y” is the crosswind distance from the source to the receptor (m), “z” is the vertical distance above the ground (m), and “H” is the effective stack height (physical stack height plus plume rise) expressed in m.

The Gaussian dispersion equation uses a relatively simple calculation requiring only two dispersion parameters (σy and σz) to identify the variation of gas concentrations away from the diffusion source. These dispersion coefficients, σy and σz, are functions of wind speed, cloud cover, and surface heating by the Sun. Generally, the evaluation of the diffusion coefficients is based on atmospheric stability class. In the employed method, Pasquill–Gifford stability classes are used, and dispersion coefficients are calculated using the Briggs model (Briggs, 1965).

The optimization task is tackled using genetic algorithm routines in MATLAB. As a stochastic search method, including genetic algorithm optimization can efficiently explore complex and large solution space without getting trapped in low-quality minima. Although there is no guarantee of reaching a global optimum, near-optimal solutions are usually obtained.

A total of eight surface echosounder surveys were conducted between 1 and 7 September 2018. At least, one survey was conducted each day, except for 4 September, due to evacuation of the MC20 site during severe weather associated with Tropical Storm Gordon (composite displays of all survey results can be seen in Chapter 3 of the work of Mason et al. (2019)). Survey tracklines varied in number and orientation, depending on the orientation and extent of the hydrocarbon plume and the daily operating plan coordinated among research investigators (Supplementary Figure S1A). Along-track and cross-track observations of the plume in the SBES and MBES revealed two or more sub-plumes emanating from a seabed position within an erosional pit at the northwest corner of the platform jacket. Visually differentiating the components of the plume in the acoustic echograms depended on the trajectory of the plume in the water column and the orientation of the survey trackelines. The components of the plume showed differential rise rates consistent with faster rising gas bubbles separated from slower-rising oil bubbles. On some occasions, relatively low noise on the 200 kHz channel permitted detection of components of the plume with backscatter intensity relatively higher than backscatter intensity in the 120 kHz for portions of the plume, consistent with the expected backscatter intensity patterns of liquid-filled spheres of oil, i.e., oil bubbles. Similarly, the high-frequency 500 kHz M3 multibeam surveys provided further evidence of separate components of the plume consistent with separate oil and gas components in the plume.

Figure 5 shows the components of the echogram and multibeam survey results for operations on the evening of 5 September 2018, close in time to visual plume observations reported in Gas Bubbles in the Water Column. For this survey, the vessel tracked along current, starting north of the jacket and continuing along for approximately 1,000 m to the southwest (Figure 5A). A single ping from the multibeam swath shows the jacket and a cross section of the plume (Figure 5B). It is noticed how the plume appears truncated where the beam bisects it in midwater. The trackline composite shows the entire length of the plume originating near the base of the jacket and deflecting approximately 300 m SW before reaching the surface (Figure 5C).

Figure 6 shows the 3D interpreted components of the plume observed on 2 September 2018, when currents deflected the plume to the southwest. High-backscatter components appeared to diverge upwards from relatively lower-backscatter components, which were rising more slowly, consistent with lower buoyancy of oil bubbles compared with gas bubbles. A possible second divergence occurred within the high backscatter plume as it approaches the surface, suggesting the ongoing fractionation of gas and oil components occurring closer to the surface. Drone surveillance of the ship parked within the plumes as they surfaced showed the separation of gas-dominated and oil-dominated components of the plume along the length of the vessel. An additional, much fainter, plume target was observed on 2 September 2018 about 200 m NW of the erosion crater over the location of the platform’s original foundation and well template (Supplementary Figure S1C).

The ROV collected clear images of the plume components and as they vented within a ∼10 m-wide and ∼2 m-deep erosion pit at the base of the platform jacket (Figure 7). Closer examination with the M3 sonar revealed four or five subcomponents of the plume: a pair of smaller plumes to the southwest of the erosional pit and two separate larger plumes to the north. A fifth feature was less defined between the pair of smaller plumes and the southernmost of the two larger plumes. The characteristics of the backscatter intensity suggested the smaller pair was composed of oil, whereas the ones with higher backscatter intensity were composed of gas.

A total of 5,881 gas bubbles and 6,258 oil bubbles were counted and measured based on image samples collected by using the bubblometer. Combining the depth of the vehicle as image collections were made with its navigation track meant that each image sample, and the number of bubbles detected in that sample, could be mapped in three dimensions (Figure 8). Bubble counts and densities were used to fit a cross section that segments the high-bubble-density core of the plume, shown as the red, blue, and yellow polygons in Figures 8B–D. The volume of this “core” plume region is the height of each depth region multiplied by the core cross section. The results showed that bubble abundance was variable among image samples, while the apparent density of bubbles observed was different within different depth ranges; that is, image samples within the erosion crater (>135 m), between the seafloor and the upper extent of the platform (135–125 m) and in the water column above the jacket (<125 m), tended to show different bubble abundances (Table 2). The number of bubbles per sample could be extrapolated to estimate the abundance of bubbles per m³ according to the volume of the bubblometer’s imaging chamber (0.018 m³).

Bubble size and size frequency distributions were estimated for the observations prior to and after installation of the containment system. The pre-containment mean bubble diameter was 8.1 mm (n = 5,881: median 7.2 mm, stdev 3.16) for bubbles observed at water depths from >135 to 120 m; the post-containment mean bubble diameter was 7.4 mm (n = 585: median 6.7 mm, stdev 3.13) for bubbles observed at 120 m. Size frequency distributions were similar for the two datasets (Supplementary Figure S2). There were small differences. The 2018 distribution had several folds more of the 2.5 mm bubbles, and peaked at 7.5 mm, while the 2022 distribution peaked at 5 mm. This type of differences may not always be negligible in terms of bubble dissolution (see the work of McGinnis et al. (2006), for example).

The pre-containment collections were slightly skewed toward bubbles of large diameter (>20 mm) compared with post-containment observations. Bubble measurement procedures for the post-containment observations were based on ROV video under marginal water clarity. Both sets of measurements were approximate due to scaling uncertainty.

Analysis of the gas samples collected in midwater with the use of the pressure chamber showed that the gas bubbles comprised predominantly methane with a mixture of higher hydrocarbons consistent with thermogenic gases typical of the Gulf of Mexico (Table 3). Additional details regarding hydrocarbon analysis from MC20 can be found in the work of Gaskins et al. (2019).

The concentration of methane in all water samples, including the 2018 and 2021 Niskin samples and the 2018 bubblometer water samples, spans in five orders of magnitude (Figure 9), from an expected background level of 0.003 µM CH₄, for seawater equilibrated with ambient air, to extreme values of >60 µM CH₄ measured in water samples from the bubblometer pressure cylinders and associated most closely with the copious flux of oil. Generally, the 2018 Niskin samples, the collection of which was targeted using the acoustic signal of the plume rather than visual observation of bubbles in the water column, showed methane levels that were slightly elevated versus expected background, consistent with the influence of the plume. The 2021 Niskin samples were collected using the ROV video feed to verify the presence of bubbles. The reader should recall that these bubbles had passed through a passive separator that removed oil and allowed gas to pass through unimpeded. Reference samples, collected >30 m from the bubble release point, with no bubbles visible, showed background concentrations of methane. The highest methane concentrations in the 2021 collections (∼1 µM CH₄) were observed in samples collected among copious visible bubbles at depths of 123 m directly above the release point and 82 m amid copious visible bubbles. Overall, Niskin samples from the bubble plume taken at >30 m depth in 2021 (n = 20: mean = 0.550 µM CH_4, stdev = 0.680) were significantly greater (p < 0.005) than the comparable samples from 2018 (n = 34: mean = 0.014 µM CH_4, stdev = 0.0157). Extreme methane concentrations in the bubblometer water collections from 2018 reflect collection into a pressure cylinder that includes a head volume of gas and liquid oil.

The cavity ring-down spectrometer (CRDS) was in near-continuous operation recording at 4 Hz the atmospheric concentrations of methane (ppmv) at 3 m above the ocean surface in the vicinity of the Taylor platform site in MC20 for 1–7 September 2018, except for about 36 h when the ship had to vacate during the passage of Tropical Storm Gordon. The CRDS also operated continuously during 13–17 November 2021, except for a 1 h restart of the instrument during 15 November, also at 3 m above the ocean surface. These measurements are summarized in Table 4. Note that the CRDS recorded a slightly elevated average methane concentration in 2021 compared to 2018. The 2021 summary data are presented with and without the interval when oil was being pumped into transfer tanks on the ship deck. The major difference between the two CRDS surveys was the broad distribution of peak methane concentrations in 2018. Figure 10 shows the comparative plots of mean atmospheric methane concentrations observed over the MC20 site; Figure 10A shows the results for 2018 and Figure 10B, for 2021—excluding the period when oil was being pumped up to the ship. Note that, during 2018, acoustic surveys and other operations meant that the ship’s track covered a much broader area at the MC20 site, while in 2021, broader area surveys were more curtailed and the ship was mostly positioned directly over the platform wreckage (Supplementary Figure S1).

Results were strongly dependent on the circumstances of hydrocarbon release at the seafloor and operations of the vessels. During 2018, gas and oil rose unimpeded to the surface, with water column currents largely slack or moderate (<0.3 m/s). In April 2019, authorities installed the containment system that diverted oil from the hydrocarbon plume into storage tanks. Oil recovery rates from these storage tanks subsequently showed that oil being released into the water column and rising to the surface would have exceeded 4.5 m³/d during 2018 (O'Reilly, 2020). Oil was continuously observed over the site during 2018 and produced the large surface oil slicks typical of this spill (Daneshgar Asl et al., 2015). During 2021, in contrast, there was little visible oil that rose to the surface. However, large amounts of oil were pumped up from the underwater storage tanks to the ship and into transport tanks that vented to the air. The pump-off period lasted 12 h and was followed by an exercise to moor a CTD so that it was suspended over the erosion crater and under the containment. Figure 11 shows the annotated timelines of CRDS observations over the MC20 sites during 2018 and 2021.

During the acetylene tracer experiment conducted in 2018 and the routine CRDS measurements, background acetylene concentrations fluctuated between 0 and 0.9 parts per billion (ppb). During the tracer experiment, when acetylene was being released from the raft tethered over the wreck site at a controlled rate of 15 L/min (0.29 g/s at STP), three major spikes on the CRDS for acetylene were detected with values that exceeded 200 ppb of acetylene in the air. The highest concentration of the acetylene tracer gas was detected only a few meters away from the raft deployment in the “bubbling zone” (1,024.2 ppb), and after that, two more spikes on the tracer gas were detected downwind (11.16 knots, ESE) at approximately 310 and 430 m away from the deployment site, respectively. Because the vessel R/V Brooks McCall was drifting downwind during the tracer study raft deployment, there were at least three opportunities where the tracer spikes almost perfectly matched the relatively high measurements of methane in the air, further confirming that the source of additional methane in the air was sourced from the “bubbling zone” where the hydrocarbon plume was actively reaching the ocean surface (Supplementary Figure S3).

Tracer gas (acetylene) air concentration measurements were also employed to calibrate/validate the inverse plume measurements method that is proposed to estimate methane emission rate estimates. Acetylene measurements are used as an input for the method in order to test its ability to predict the actual emission rate of the tracer gas (15 L/min). Wind direction and speed, along with a set 4,175 data points (acetylene concentrations and measurement locations), were used as inputs for the identification method. The average acetylene emission rate predicted by the method was equal to 0.26 g/s which approximately corresponded to the actual emission rate of 15 L/min (0.29 g/s at STP). The results obtained with the tracer (acetylene) gas release, in a controlled manner, show that the inverse plume modeling method can be used to estimate methane emission rates under marine conditions.

The inverse plume modeling combined with atmospheric methane concentrations was used to estimate transfer from the hydrocarbon plume into the atmosphere. These estimates provide an average methane emission rate equal to 9 g/s with a standard deviation of 1.1. This corresponds to a discharge equivalent to 0.8 t CH₄/d. However, it is important to note that this estimation method is prone to variability in wind direction and speed.