Offshore Subsurface Infrastructure Imaging From AI Deep Learning-Based Three-Dimensional Acoustic and Magnetic Data Inversion

Hurricane Ivan in 2004 was a catastrophic event that toppled an oil production platform in the Northern Gulf of Mexico, displacing it over 250 m from its original location. Oil well conductors attached to the platform and many other metallic structures were also displaced and scattered across the existing sea floor. In 2022, Couvillion engaged Kraken Robotics to deploy their proprietary imaging technology, Acoustic Coring (AC), to image these conductors, currently buried beneath a 30 to 45 m thick layer of delta-released sediments below the mudline. The three-dimensional imaging technology, developed by [1], was designed for the shallow sub-seabed to identify and characterize geohazards and stratigraphy. This technology requires specular and non-specular returns with spatial accuracies that exceed those of current conventional seismic surveys (i.e., towed streamer-based methods) [2]. This was achieved through sub-seabed interrogation using a stationary transmitter and receiver array, which holds a spatially centimetre-spaced network across a horizontal imaging dimension of 14 metres [1] [2]. A total of 63 AC scans, consisting of 21 rows along three scanning columns spanning a 205 m x 60 m survey area, were conducted to determine the extent, expanse, orientation, and characteristics of lost conductor pipes. Of significance were continuous linear conductors reposing onto a basement geology lying directly across from the containment dome towards the location of the original conductor bay, following the basement slope closely. The pipes' endings show an abrupt spread-out termination with an undulating humped, slightly down-looking character around survey Row 11. In the region around this termination, gas-laden soil inhibited the acoustic ability to get acoustic returns and collect further data. It was postulated that the conductors were severed at this location, no longer existed, or continued uninterrupted toward the original oil well location. To help clarify the possibilities, a high-resolution magnetometer survey comprised of four three-component flux gate magnetometers integrated into a Millenium Plus ROV was conducted. The aim was to survey a region defined as an area of 460m (1509ft) by 405m (1328ft) to determine whether magnetometer data collected beyond Row 11 and extending toward the conductor bay could indicate the presence of buried pipes. The high-quality magnetic data showed continuity and agreement with the pre-survey modelling results. The results concorded with the acoustic conductor features mapped initially, and the responding magnetic lobes were consistent with debris targets predicted by modelling. Out-of-plane influences, such as the rotated displaced oil rig platform and heavy accumulation of metal debris around the original well, complicated the signals. However, the recorded magnetic data by itself shows a termination of the high-intensity lobe 40 m NE of the known location of the well conductor bay. The nature and geometry of the conductor bundle, as it extended beyond the last imaged location in the AC and towards the well bay, was not easily deciphered by common human vision-driven interpretation of magnetic data. To obtain this information, extensive AI -based modelling was developed utilizing the precision and details of the Acoustic Corer data results to train the AI system on the magnetometer data [5]. The 3D Training Models were created using an 8 GPU computer cluster handling 110 Epochs (in computing, an epoch is a fixed date and time used as a reference from which a computer measures system time). High-performance cloud-based computation clusters from AWS were deployed to simulate forward models to generate input training data rapidly. 202,000 3D training models and their magnetic response at 2 m intervals (high resolution) were generated. A deep learning architecture was then deployed to learn how to predict the underlying 3D magnetic model from the simulated response. This was achieved on another cloud-based system hosted by Shadeform, utilizing a network of 8 GPU machines, each equipped with 80 GB RAM. In this paper, the AI imagery is presented. It shows that damage to the well site is evident, revealing debris around the original well and where the conductors appear after Row 11. The continuity trend of the conductor pipes exhibits variation in signal strength responses beyond Row 11, indicating that past Row 11 and nearing the location of the original well, the conductors have sustained damage and may contain separated and fragmented pipes. However, the AI enhancements delineate continuity, highlight the trajectory of the pipe infrastructure toward the well placement, and partially encircle it. Simultaneously, a secondary parallel conductor trajectory branches off and leads forty meters north of the original well. This secondary trajectory consists of fragmented conductors or pipe-related debris and terminates where the high survey-based magnetic main anomaly lobe also concludes. As will be shown in the paper, the resulting magnetometer imagery was influenced by a complicated geometry of scattered metallic debris. Aspects of the critical conductor system, as imaged in detail by the Acoustic Corer, correlated well. The conductor package continues beyond Row 11; the last imaged location of the conductors seen in the acoustic thins out in places but continues towards the original well bay, albeit 40 m to the northeast, as suggested by the high anomaly lobe in the magnetic data. However, the AI imaging brought out a hidden subtle secondary branch of the conductor package veering and connecting with the metallic debris near the well bay, establishing clear evidence of a potential flow path from the well bay back to the downed jacket and oil /gas collection dome. These results could be achieved only through the fusion of acoustic imaging, magnetic data, and the intensive AI Deep Learning analysis that could look deep into the magnetic responses to reveal the subtle, hidden, buried features at higher discriminating resolutions.