Learning Electrical Structure Representations from Ore-Bearing Cores ERT Data Using a Dual Branch CNN Transformer Architecture

Electrical Resistivity Tomography (ERT) provides an effective means for probing the internal electrical structure of rock cores and plays an important role in understanding the electrical properties of ore-related geological bodies. Recovering informative structural representations from limited and highly coupled measurement data, however, remains challenging, particularly for drill cores, where complex resistivity distributions are commonly observed. Restricted electrode configurations and scale effects further hinder the ability of conventional inversion schemes and existing convolutional neural network (CNN)–based approaches to preserve structural continuity and spatial correlations in core-scale ERT imaging.

In this study, we investigate a dual-branch CNN–Transformer architecture designed for learning electrical structure representations from core-scale ERT data. The proposed approach adopts an end-to-end image-to-image learning paradigm to explore how complementary data organizations can be leveraged for representation learning. Two dedicated Transformer branches are incorporated: the first branch exploits potential difference data acquired from multiple sets of sequentially excited adjacent electrode pairs with consistent relative spatial configurations, while the second branch utilizes potential difference measurements collected at multiple spatial locations under a single electrode excitation.

By integrating the local feature extraction capability of CNNs with the global dependency modeling strength of Transformers, the proposed architecture aims to construct more expressive representations of complex electrical structures, thereby supporting improved structural coherence and spatial resolution in ERT imaging. Preliminary results, evaluated using quantitative imaging metrics including correlation coefficient and structural similarity index, suggest that the learned representations capture coherent electrical features under varying anomaly geometries, resistivity contrasts, and spatial distributions. These early findings demonstrate the feasibility of combining CNNs and Transformers for electrical structure representation learning in core-scale ERT and provide a methodological foundation for subsequent development of effective deep learning–based inversion strategies oriented toward deep mineral exploration applications.

This work is supported by National Science and Technology Major Project for Deep Earth Probe and Mineral Resources Exploration under Grant 2025ZD1008500.