190-5 Evaluating Cross-Well Generalization for Subsurface Pore Pressure Prediction Using Explainable Machine Learning

Abstract:

Accurate pore pressure prediction is critical for safe drilling operations, well planning, and the prevention of hazardous events such as blowouts and formation damage. However, machine learning models trained on a specific set of wells often perform poorly when applied to well log data from different geological regions. While this limitation is widely recognized in the petroleum and geoscience industries, less is understood about why certain features and model architectures generalize more effectively across wells than others. Understanding this behavior is particularly important for building robust AI systems that can operate reliably in data-scarce subsurface environments.

To investigate this problem, the study evaluated five machine learning architectures, including Random Forest, XGBoost, Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and a deep feedforward neural network, using data from 21 wells in the Potwar Basin, Pakistan. The models were tested under a strict leave-one-well-out protocol in which each well was completely excluded during training and used only for blind evaluation. This setup was designed to simulate realistic cross-well deployment conditions rather than conventional random train-test splitting.

In their standard form, all models failed to generalize across wells and even performed worse than a naïve baseline that simply predicted the mean pore pressure. These findings demonstrate how conventional random cross-validation can conceal severe generalization failures and emphasize that evaluation design is as important as model selection itself. To overcome this limitation, the study incorporated targeted domain adaptation, which substantially improved performance. The ensemble framework achieved the best results, reaching an R2R^2R2 value of 0.938 on the blind test well.

Beyond predictive accuracy, the study also investigated why certain models performed better than others. SHAP-based explainability analysis revealed that the Eaton ratio was the most influential physics-informed feature in both tree-based models, contributing up to 29% of prediction influence in XGBoost. In contrast, deep resistivity contributed negligibly despite being included in training. The analysis further showed that shallow intervals were better handled by tree-based models, whereas deeper zones favored neural networks, reflecting changes in subsurface pressure mechanisms with depth. Overall, the findings demonstrate that physics-informed features contribute meaningfully to prediction performance and that explainability analysis provides a practical way to verify when models genuinely utilize physically relevant information.

Geological Society of America Abstracts with Programs. Vol. 58, No. 5, 2026

doi: 10.1130/abs/2026AM-15610

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.