Document Type : Original Research Paper
Authors
- Mahdi Eskandari
- Alireza Arab Amiri
- Abolghasem Kamkar-Rouhani
- Amir Joolaei
- Meysam Davoodabadi Farahani
School of Mining, Petroleum and Geophysics, Shahrood University of Technology, Shahrood, Iran
Abstract
Rapid and accurate prediction of effective resistivity in azimuthally anisotropic media remains a computational challenge in geoelectrical studies, as conventional 3D finite element method (FEM) simulations are numerically expensive for iterative inversion. This study presents a hybrid ANN-PSO framework for predicting the effective resistivity of azimuthally anisotropic media, leveraging three-dimensional finite element modeling (FEM) as a high-fidelity data engine. A comprehensive synthetic dataset (n=1000) was generated by simulating geoelectrical responses over two-layer anisotropic structures with varying resistivity tensors and strike angles. To overcome the limitations of manual hyperparameter tuning, Particle Swarm Optimization (PSO) was integrated to systematically optimize the Artificial Neural Network (ANN) architecture, including neuron counts, dropout rates, learning rate, and batch size. The PSO-optimized ANN achieved a 54.2% reduction in Mean Squared Error (MSE) and a 36.4% reduction in Mean Absolute Error (MAE) compared to the baseline model, reaching a coefficient of determination (R2) of 0.9937. Furthermore, robustness analysis under 10% Gaussian noise confirmed the superior resilience of the hybrid framework (R2=0.8985) over standalone ANN architectures. The results demonstrate that while the model functions as a highly efficient surrogate for computationally intensive FEM simulations, its performance is optimized for the defined parameter space of HTI media. This hybrid approach offers a robust and rapid predictive tool for accelerating geoelectrical data interpretation and inversion in mineral exploration and subsurface conductivity characterization.
Keywords
- Electrical resistivity
- Azimuthal anisotropy
- Finite element method
- Artificial neural network
- Particle swarm optimization
Main Subjects
)