Using rock physics analysis driven feature engineering in ML-based shear slowness prediction using logs of wells from different geological setup

• Published in: Acta geophysica 2024, Vol.72 (5), p.3237-3254
• Main Authors: Chakraborty, Shantanu, Datta Gupta, Saurabh, Devi, Varsha, Yalamanchi, Pydiraju
• Format: Article
• Language: English
• Subjects: Algorithms; Cost analysis; Data analysis; Datasets; Deep water; Deepwater drilling; Earth and Environmental Science; Earth Sciences; Geology; Geophysics/Geodesy; Geotechnical Engineering & Applied Earth Sciences; Machine learning; Modelling; Physics; Porosity; Predictions; Research Article - Applied Geophysics; Reservoirs; Rocks; Shear; Structural Geology; Water wells
• ISSN: 1895-7455; 1895-6572; 1895-7455
• DOI: 10.1007/s11600-023-01266-3

Shear slowness data are crucial data in rock physics analysis and seismic reservoir characterization. In petrophysical formation evaluation, the use of sonic data is limited, and hence, sonic data, especially shear sonic, are not considered as critical. In many deep-water wells to save the cost of operations, shear sonic data are not recorded. In these scenarios for rock physics analysis, it becomes necessary to predict shear sonic data from other available datasets. Conventional techniques for shear slowness predictions rely on empirical relations and rock physics modeling. However, these approaches require extensive information as input and additionally carry assumptions and multiple prerequisites. Presently with the advancement of computing power Machine learning (ML) emerges as a robust and optimized technique for predicting precise DTS in quick time and with fewer input datasets. In this study, wells located in the deep-waters of the East Coast of India and penetrated siliciclastic reservoirs of both compacted sand and soft high porosity sands were used as input to train the ML algorithm. Random Forest machine learning algorithm is best used for both classification and regression tasks, and this algorithm is used here for the data prediction. As a comparison, the convolutional LSTM method is also used for data prediction. To comply with the geological variability in the prediction and to enhance the prediction accuracy, rock physics understandings were used as a guide in feature engineering. The RF prediction shows a good match of ~ 93%, and the LSTM model prediction shows ~ 94% correlation at validation well. Both the model predicted data show good agreement with the rock physics modeling interpretations at the target well.