# Probabilistic 3D lithology classification from elastic property volumes an advanced inversion workflow at Desouq Gas Field, West Nile Delta, Egypt

## TL;DR

This paper presents a new workflow combining seismic inversion and machine learning to improve 3D lithology classification in gas fields, enhancing reservoir understanding.

## Contribution

A novel workflow integrating seismic inversion, rock physics, and machine learning for probabilistic 3D lithology classification.

## Key findings

- Probabilistic lithology classification revealed compartmentalized gas-sand channels and eight new gas-charged zones.
- Integration with rock physics reduced misclassification risks caused by thin anhydrite layers.
- The workflow improved reservoir outlines and subsurface heterogeneity insights in the Desouq Gas Field.

## Abstract

Advanced 3D lithology prediction is vital for reducing uncertainty in reservoir characterization and exploration planning. Traditional post-stack inversion yielded only acoustic impedance volumes, limiting facies discrimination. However, pre-stack simultaneous inversion enables direct estimation of elastic property volumes, particularly P-impedance, shear impedance and Vp/Vs, linking seismic inversion to rock physics evaluation. In this study, the applied workflow integrates seismic inversion products and borehole information within a structured, multi-stage lithology classification framework. Initially, 3D seismic inversion volumes and available well logs were subjected to comprehensive quality control and jointly interpreted to establish a consistent geological framework. This geological interpretation guided subsequent petrophysical analysis of the well logs, from which key reservoir properties were derived. Based on these petrophysical results, rock physics crossplots were conducted to define and discriminate the lithology classes generating a litho-facies log for each well and characterize their elastic responses. The classified well-based data were then, used to generate probability density functions (PDFs) for each lithology class, forming the statistical foundation of the lithology classification model. The dataset used to train an ML algorithm (trained model) was subsequently applied to the 3D seismic inversion volumes to predict lithological distributions away from the wells. Finally, the resulting lithology classification volumes were visualized, interpreted, and quality-controlled to delineate reservoir outlines and assess their spatial continuity and geological credibility. This workflow applied to the Abu Madi Formation in the west onshore Nile Delta, with a focus on the Desouq Gas Field. The probabilistic classification revealed compartmentalized gas-sand channels, refined hydrocarbon facies outlines in the northwest sector, and identified eight previously unrecognized gas-charged zones in the southwest sector. Validation using classification metrics and confusion-matrix analysis confirmed the robustness of the workflow, while integration with elastic property crossplots clarified ambiguities caused by thin anhydrite layers that commonly generate misleading amplitude responses which reduced misclassification risks. The resulting 3D lithology volumes (gas sand, wet sand, shale, and tight anhydrite formation) provide enhanced insights into subsurface heterogeneity and hydrocarbon potential, demonstrating the added value of integrating seismic inversion, machine learning, and rock physics analysis.