Deep {Learning} based {Electrical} {Resistivity} {Tomography} {Inversion}

Basile Gandon and Paul Cupillard. ( 2025 )
in: 2025 {RING} meeting, pages 81--94, ASGA

Abstract

Electrical Resistivity Tomography (ERT) inversion is challenged by subsurface heterogeneity, survey variability, and the intrinsic non-linearity of the inverse problem. While deep learning approaches are increasingly explored as alternatives for ERT inversion, most existing models assume fixed survey configurations—such as array length, number of electrodes, and electrode geometry—thereby limiting their applicability to real-world scenarios. This study presents a deep learning framework designed for rapid ERT inversion capable of handling diverse geological scenarios and acquisition parameters. A large synthetic dataset was created using 3D geological models (Noddyverse) to generate varied 2D resistivity sections, for which apparent resistivity data were simulated using different array types and electrode counts using the SimPEG package. A U-Net architecture was trained to map normalized pseudosections and survey metadata (electrode count, array type, geometry) to subsurface resistivity models. Training utilized gradient accumulation to manage variable input sizes and incorporated a sensitivityweighted loss function, prioritizing data-constrained model regions. Evaluation on synthetic data shows the network accurately resolves shallow features, with decreasing resolution at depth correlating strongly with lower data sensitivity. Inference is computationally negligible ({\textless}1s). When applied to experimental data from the PEGGHy test site, the network produced results comparable in main structural features to conventional inversion outputs. This deep learning approach offers a computationally efficient alternative for ERT inversion, particularly suited for rapid initial appraisals, though considerations regarding generalization beyond the training distribution and model resolution remain important.

Download / Links

BibTeX Reference

@inproceedings{Gandon2025RM,
 abstract = {Electrical Resistivity Tomography (ERT) inversion is challenged by subsurface heterogeneity, survey variability, and the intrinsic non-linearity of the inverse problem. While deep learning approaches are increasingly explored as alternatives for ERT inversion, most existing models assume fixed survey configurations—such as array length, number of electrodes, and electrode geometry—thereby limiting their applicability to real-world scenarios. This study presents a deep learning framework designed for rapid ERT inversion capable of handling diverse geological scenarios and acquisition parameters. A large synthetic dataset was created using 3D geological models (Noddyverse) to generate varied 2D resistivity sections, for which apparent resistivity data were simulated using different array types and electrode counts using the SimPEG package. A U-Net architecture was trained to map normalized pseudosections and survey metadata (electrode count, array type, geometry) to subsurface resistivity models. Training utilized gradient accumulation to manage variable input sizes and incorporated a sensitivityweighted loss function, prioritizing data-constrained model regions. Evaluation on synthetic data shows the network accurately resolves shallow features, with decreasing resolution at depth correlating strongly with lower data sensitivity. Inference is computationally negligible ({\textless}1s). When applied to experimental data from the PEGGHy test site, the network produced results comparable in main structural features to conventional inversion outputs. This deep learning approach offers a computationally efficient alternative for ERT inversion, particularly suited for rapid initial appraisals, though considerations regarding generalization beyond the training distribution and model resolution remain important.},
 author = {Gandon, Basile and Cupillard, Paul},
 booktitle = {2025 {RING} meeting},
 language = {en},
 pages = {81--94},
 publisher = {ASGA},
 title = {Deep {Learning} based {Electrical} {Resistivity} {Tomography} {Inversion}},
 year = {2025}
}