289 Estimating Natural Resources Using Geoscience Data
Methods to estimate the availability, location, and quantity of natural resources (e.g., geothermal, petroleum, critical minerals) vary by resource type, scale, and data availability. This session addresses modeling that targets any of these aspects.
Wednesday, 22 October 2025
1:30 PM - 5:30 PM
HBGCC, 302C
Beyond the Surface: Deep Neural Networks for Subsurface Pore Pressure Estimation
Paper number: 289-7
Program: Technical Sessions
Time: 3:30 PM – 3:50 PM
Authors
-
Muhammad AmjadBahria School of Engineering and Applied Sciences Bahria University, Islamabad, Pakistan, Islamabad
-
Ilmi YoonNortheastern University, San Jose, California
Abstract
Accurate prediction of subsurface pore pressure is essential for safe and cost-effective drilling in the field of geosciences and hydrocarbon exploration. The traditional physics-based methods for prediction have limitations in capturing the nonlinear, depth dependent variations in pore pressure, particularly in complex geological basins. The machine learning based methods are often highly domain specific. They perform well on the data similar to the one they are trained on but may struggle to generalize to unseen or out of distribution data. In this study, we evaluate and enhance the performance of Deep Learning models including Convolutional Neural Networks (CNN) and Deep Feed Forward Neural Networks (DFNN) for pore pressure prediction from conventional well log data to improve their generalization capability with fine tuning and limited data injection. The key predictors used in experimentation include, Gamma Ray, Sonic, Normal Compaction Trend, Density, and Resistivity collected from four wells of Potwar Basin, which is a major oil and gas producing basin of Pakistan. Three wells are used for training and testing while a fourth blind well is used for validation. Models performed well on test data with an R2 of 0.7885 for CNN and 0.7988 for DFNN but are unable to generalize for blind well. To improve the performance on blind well, we conducted a series of experiments and fed the model with 10%, 15% 20% and 25% of the labeled blind well data into training and validated the model predictions on remaining unseen data from blind well. This strategy significantly improved the model’s ability to generalize. The CNN models showed gradual improvement in value of R2 with 0.6888 at 10% injection to 0.7390 for 25% injection. The DFNN model on the other hand shows remarkable performance with R2 value of 0.7062 with only 10% injection and achieved maximum R2 of 0.7349 with 25% injection. The findings also show fine tuning with limited blind well data can enhance model generalization without any need of additional features or architectural revamps. In future, we will explore the outcomes with adaptive sampling strategies and hybrid ensemble frameworks for further robustness. These findings underscore the potential of advanced machine learning models to enhance pore pressure prediction, offering a scalable solution to challenges in subsurface exploration and drilling risk assessment.
Geological Society of America Abstracts with Programs. Vol. 57, No. 6
doi: 10.1130/abs/2025AM-412529
© Copyright 2025 The Geological Society of America (GSA), all rights reserved. Permission is hereby granted to the author(s) of this abstract to reproduce and distribute it freely, for noncommercial purposes. Permission is hereby granted to any individual scientist to download a single copy of this electronic file and reproduce up to 20 paper copies for noncommercial purposes advancing science and education, including classroom use, providing all reproductions include the complete content shown here, including the author information. All other forms of reproduction and/or transmittal are prohibited without written permission from GSA Copyright Permissions.