Integration of the Biot-Gassmann Fluid Substitution Method and Machine Learning-Based Velocity-Stress Relationship...

Area of Science:

• Geophysics and Rock Mechanics
• Machine Learning Applications in Earth Sciences

Background:

• Accurate in situ stress estimation is crucial for subsurface engineering, including geothermal energy extraction.
• Traditional methods often rely on saturated rock data, but low-frequency interpretations may benefit from dry rock properties.
• The impact of frequency dispersion on velocity-stress relationships for in situ stress prediction remains an open question.

Purpose of the Study:

• To enhance machine/deep learning (ML/DL) workflows for in situ stress prediction.
• To investigate the efficacy of using Biot-Gassmann-derived equivalent saturated velocities from dry rock data for ML/DL model training.
• To assess the impact of low-frequency acoustic velocities on the accuracy of in situ stress prediction.

Main Methods:

• Acquisition of true triaxial ultrasonic velocity (TUV) data from dry core samples under various stress configurations.
• Application of Biot-Gassmann fluid substitution to derive equivalent saturated velocities from dry rock ultrasonic velocities.
• Training and validation of ML/DL models using the derived low-frequency equivalent saturated velocities and TUV data from Utah FORGE site cores.

Main Results:

• ML/DL models trained with equivalent saturated velocities achieved high predictive performance for in situ stress.
• Validation/testing yielded R-squared values of 0.86, 0.971, and 0.975 for vertical, minimum horizontal, and maximum horizontal stress, respectively.
• Shapley additive explanations (SHAP) analysis confirmed model reliability and improved scientific understanding of the velocity-stress relationship.

Conclusions:

• The enhanced ML/DL workflow using low-frequency acoustic velocities from dry rock is a viable and accurate method for in situ stress prediction.
• This approach offers a reliable alternative for interpreting lower frequency measurements in subsurface geological settings.
• The study validates the use of Biot-Gassmann theory in conjunction with ML/DL for robust geophysical stress analysis.