Abstract

This study presents a machine learning-driven approach for hydrocarbon probability mapping using Low-Frequency Passive Seismic (LFPS) data from the Akholjuni Field in the Indian Cambay Basin. LFPS leverages naturally occurring low-frequency seismic waves (0.1–10 Hz) and their interactions with hydrocarbonbearing formations to detect spectral anomalies, particularly in the 2–6 Hz frequency range. Virtual stations were generated through seismic interferometry to enhance spatial coverage, and a suite of spectral and statistical attributes were extracted from vertical and horizontal components. Supervised machine learning models including CNN, FNN, Random Forest, and XGBoost were trained on labelled data derived from proximity to known wells. The best-performing models demonstrated high accuracy and generalization, with a weighted ensemble model aligning with known reservoir structures and blind well probability outcomes. The resulting probability map offers a valuable tool for optimizing well placement, delineating reservoir extent, and supporting reserve estimation, highlighting LFPS as a promising Direct Hydrocarbon Indicator (DHI) technique.