Shear wave velocity prediction using Long Short-Term Memory Network with generative adversarial mechanism

Area of Science:

• Geophysics
• Petrophysics
• Machine Learning

Background:

• Shear wave velocity (Vs) is vital for subsurface characterization but difficult to acquire.
• Conventional long short-term memory (LSTM) networks predict Vs but struggle with dataset discrepancies.
• Existing methods often yield suboptimal performance due to domain shift issues.

Purpose of the Study:

• To propose an enhanced LSTM architecture with a generative adversarial mechanism (LSTM-GAM) for improved Vs prediction.
• To address the limitations of conventional LSTMs in handling training and prediction dataset differences.
• To enhance the cross-domain adaptability and generalization capability of Vs prediction models.

Main Methods:

• Developed an LSTM-GAM framework combining an LSTM backbone for sequential data and a generative adversarial module.
• The generator minimizes reconstruction errors, while the discriminator identifies common essential features across datasets.
• Employed adversarial feature alignment to improve cross-domain adaptability.

Main Results:

• The LSTM-GAM model achieved a mean absolute error (MAE) of 59.4 m/s and R² of 0.9064 on South China Sea well data.
• Demonstrated superior prediction accuracy compared to conventional LSTM networks.
• Ablation studies confirmed consistent performance improvements and enhanced generalization capability.

Conclusions:

• The proposed LSTM-GAM offers an effective data-driven solution for shear wave velocity estimation.
• The model shows significant potential for applications in complex geological environments.
• This advancement improves the reliability of petrophysical parameter estimation using logging data.