Exploration and development of fractured carbonate reservoirs containing according to various estimates, between 35% and 48% of world's oil reserves and between 23% and 28% of natural gas reserves is one of the most challenging problem in petroleum geophysics.
Seismic data processing enables more accurate localization of faults, fractures, and caverns containing hydrocarbons and also generating diffraction/scattering of seismic waves bearing crucial information about properties of the target objects, since we know that fractures often determine the orientations of fluid flows, while caverns point to a petroleum accumulation zone. Scattered waves are used for constructing seismic diffraction image attributes and their geological interpretation.
Researchers from IPGG SB RAS developed an algorithm for extracting scattered waves from the data domain using machine learning and a synthetic training set that closely approximates real data.
The algorithm was successfully tested in the course of the experiments, both using synthetic data obtained for a realistic 3D model with fractures for one of the field in Eastern Siberia, and real data.
Project Leader: M.I. Protasov, DSc (phys.-math.)
The study has demonstrated a clear correlation between the algorithm's performance and the size and variety of the synthetic training set.
M.I. Protasov, DSc (phys.-math.), head of the Laboratory of Methods for the Interpretative Processing of Seismic Data at IPGG SB RAS noted that it is critical for a model to include different types of noise affecting the training of neural networks to enhance their performance and robustness. Precisely this model adapted to simulated real noise conditions has shown the best results in seismic survey data processing and interpretation.
Applications of the tested algorithm is a promising strategy for solving real problems of exploration geophysics.
For reference
For more information, see the article by:
Protasov M.I., Kenzhin R.M. Diffraction separation from seismic wave-fields in data domain using machine learning // Geophysics – No. 2 – pp. 9-16 – 2026
This work was funded by the Russian Science Foundation, RSF grant no. 21-71-20002-P. The results were obtained using advanced computing resources of the Supercomputer Center of Peter the Great St. Petersburg Polytechnic University (scc.spbstu.ru).
Published by IPGG Press Service