| Abstract Submission No. | ABS-2.4-0035 |
| Title of Abstract | Estimation of Gas Hydrate Saturation Using Machine Learning |
| Authors | Anchal Jaiswal* |
| Organisation | IIT (ISM) Dhanbad |
| Address | IIT(ISM) Dhanbad Jharkhand Dhanbad, Jharkhand, India Pincode: 826004 E-mail: 23mc0009@iitism.ac.in |
| Country | India |
| Presentation | Poster |
| Abstract | Sediments that contain gas hydrates can pose a geohazard, affecting slope stability and climate feedbacks. They also represent a promising energy source. Evaluating gas hydrate saturation (Sh) is crucial for understanding geomechanical risks and resource potential. Conventional estimation methods, such as resistivity, acoustic velocity, and nuclear magnetic resonance (NMR) logging, have notable limitations. For instance, resistivity methods are affected by pore-water salinity and require specific Archie parameters. Acoustic methods depend on accurate elastic models and can be influenced by variations in rock types. NMR measurements are direct but can be expensive and often not available. The accuracy and usefulness of traditional workflows are constrained by these challenges, particularly in data-limited situations. To address the challenges of parameter sensitivity and financial constraints in traditional Methods, this study develops a machine learning (ML) framework to predict gas hydrate saturation more accurately and flexibly. The process begins with detailed well-log interpretation and quality control of NMR, acoustic, and resistivity data from multiple hydrate-bearing sites. Despite their uncertainties, standard resistivity, acoustic, and NMR techniques were initially applied to estimate gas hydrate saturation to create reference labels. After validation on separate datasets lacking one or more complex measurements, a supervised machine learning model was trained using these multi-parameter petrophysical estimates. This model aimed to capture complex, nonlinear relationships between routine logging measurements and gas hydrate saturation (Sh). With about an 80% accuracy rate in replicating benchmark saturation estimates, the ML model surpassed individual physical methods and showed strong performance across various lithologies and depositional environments. The suggested ML approach offers a transferable and affordable tool for accurate gas hydrate saturation prediction in settings where resistivity, acoustic, or NMR data are lacking or economically limited by incorporating multi-log information and discovering hidden relationships beyond the presumptions of physical models. |
| Keywords | as hydrates; Machine learning; Well log interpretation; Saturation estimation; Petrophysics |