Inversion of 2D cross-hole electrical resistivity tomography data using artificial neural network

Area of Science:

• Geophysics
• Artificial Intelligence
• Data Science

Background:

• Geophysical inversion is often ill-posed due to nonlinearity and limited measured data.
• Artificial Neural Networks (ANNs) offer a powerful approach for handling complex, nonlinear geophysical inversion problems.

Purpose of the Study:

• To apply a feedforward back-propagation neural network (FBNN) for inverting 2D cross-hole electrical resistivity tomography data.
• To evaluate the performance and accuracy of the proposed FBNN model against conventional inversion techniques.

Main Methods:

• Synthetic data generation using eighteen forward models with a dipole-dipole array configuration.
• Training and testing the FBNN model with specific hyperparameters (trainrp function, 4 hidden layers, 75 neurons/layer, 0.8 learning rate, 1 momentum coefficient) and 54,000 data points.
• Validation using laboratory testing results and error determination between actual and predicted areas.

Main Results:

• The optimized FBNN model achieved higher performance and accuracy.
• The FBNN model demonstrated a 15% to 18% lower error rate compared to conventional inversion models.
• Hyperparameter effects on FBNN model performance were systematically examined.

Conclusions:

• The developed FBNN approach is effective for 2D cross-hole electrical resistivity tomography inversion.
• The FBNN model offers a significant improvement in accuracy and error reduction over traditional methods.
• This study highlights the potential of ANNs in addressing challenges in geophysical data interpretation.