Integrating Seismic Anisotropy, Attenuation, and Machine Learning for Advanced Subsurface Characterization

Seismic anisotropy and attenuation, often quantified by the inverse of the quality factor (), are powerful, but often underexploited, indicators of fracture architecture, fluid content, and small-scale heterogeneity in the subsurface. At the same time, machine-learning (ML) methods offer flexible, data-driven mappings between seismic attributes and subsurface properties yet are not often designed to exploit seismic anisotropy and attenuation. In this contribution, an integrated workflow that combines laboratory measurements, borehole and VSP data, and surface seismic attributes with ML modelling to achieve advanced subsurface characterization in fractured carbonate systems.

Seismic waveforms collected in Abu Dhabi in the United Arab Emirates (UAE), were recorded at wide frequency range from Hertz to MHz, in the field and laboratory. The lithology of Abu Dhabi subsurface is dominated by carbonates, which are known by their high heterogeneity and multiple fracturing systems. To address the complexity caused by lithology, new methods and processing workflows have been developed and applied on the data. This includes new methods for calculating seismic attenuation from surface seismic, vertical seismic profiling (VSP), and sonic data, allowing an estimate of attenuation magnitude and its anisotropy, in addition to separating between scattering and intrinsic attenuation.

The study includes a suite of field and laboratory studies that quantify azimuthal P-wave attenuation, separate intrinsic and scattering contributions, and relate these to fracture systems and tar-mat occurrence in Abu Dhabi carbonate subsurface. These include multi-offset azimuthal VSP analyses that recover fracture strike and discriminate between open and cemented fractures using attenuation anisotropy, detailed attenuation-mode separation from VSP and sonic data, AVAz-based fracture characterization from 3D surface seismic, and ultrasonic measurements that document the sensitivity of  to petrophysical properties and saturation in carbonate core plugs. Building on this physical understanding, we extend recent work on ML-based prediction of Thomsen’s parameters from synthetic and VSP data to explicitly incorporate multi-scale attenuation attributes. Training data is generated by finite-difference modeling in anisotropic, fractured carbonate media constrained by well logs, FMI, and core information from an offshore Abu Dhabi oilfield. Input features include azimuthally dependent amplitudes of direct and reflected waves, frequency- and traveltime–derived attributes. We benchmark several ML regressors (support vector regression, extreme gradient boosting, multilayer perceptrons, and 1D convolutional neural networks) and use explainable AI tools to rank the relative importance of attenuation- versus kinematics-based features.

This study demonstrates that jointly exploiting anisotropy, attenuation, and ML substantially improves the interpretability and resolution of fracture and fluid systems in complex carbonate media. The proposed workflow is generic and can be transferred to other fractured and heterogeneous settings, offering a practical route to physics-aware, data-driven seismic characterization for reservoir development and monitoring.