The use of machine learning toward an accurate initial model for seismic inversion

Abstract

Seismic inversion is used routinely in hydrocarbon exploration and development as an effective geophysical tool. Inversion methodologies vary from relative to absolute poststack, prestack, and stochastic seismic inversion. Most of the inversion algorithms suffer from the “nonuniqueness” problem, which means there is more than one possible geologic model that fits the seismic data. Because of this, other information must be provided in the form of a low-frequency initial model. Construction of an accurate initial model requires honoring the subsurface geology of the field, adequate well control, and accurately picked horizons. The lack of all or some of these items makes it hard to build a reliable initial model, and the subsequent inversion will be subject to cumulative errors and accordingly questionable results. To mitigate these limitations, we implement the multiple linear regression algorithm to blend selected seismic attributes with seismic velocities to build a robust 3D initial model. First, we analyze the elastic properties of a given well against the seismic internal attributes and the available processing velocities. Next, we build a set of multilinear equations to convert the chosen attributes into 3D elastic properties. Finally, we smooth the results and produce the elastic initial models. No horizons are needed in this process. To prove its validity, the proposed approach was applied in the offshore Nile Delta, Egypt. The new initial models for the case studies show better consistency with the geology of the area and contain very fine details compared with the conventional initial models. Hence, the proposed approach suits exploration of deep targets as it provides a reliable link between seismic internal attributes, velocities, and elastic logs for the construction of an accurate 3D initial model even though the data may be insufficient.