Details of the abstract
Title of paper | Convolutional Neural Networks Applied to 2D and 3D DC Resistivity Inversion |
List of authors | Weit, S., Börner, R.-U., Brändel, M., Gödickmeier, P., Gootjes, R., Kost, S., Rheinbach, O., Scheunert, M., Spitzer, K. |
Affiliation(s) | Institute of Geophysics and Geoinformatics TU Bergakademie Freiberg, Institute of Geophysics and Geoinformatics TU Bergakademie Freiberg, Institute of Numerical Mathematics and Optimization TU Bergakademie Freiberg, Institute of Geophysics and Geoinformatics TU Bergakademie Freiberg, Institute of Numerical Mathematics and Optimization TU Bergakademie Freiberg, Institute of Numerical Mathematics and Optimization TU Bergakademie Freiberg, Institute of Numerical Mathematics and Optimization TU Bergakademie Freiberg, Institute of Geophysics and Geoinformatics TU Bergakademie Freiberg, Institute of Geophysics and Geoinformatics TU Bergakademie Freiberg |
Summary | A neural network approach has been developed to invert 2D and 3D apparent resistivity data. The network utilizes convolutions as well as pooling and unpooling operations to transform pseudosections into an underground resistivity model. To train the network, synthetic data were produced using an in-house finite element routine. The subsurface models to produce the data consist of homogenous halfspaces with 0 to 5 conductive spherical anomalies per simulated measurement. The anomalies, if present, have a constant total cross-sectional area. The network was trained on 15300 simulated measurements in 2D and 7500 in 3D. Results show a fairly accurate match of anomaly resistivity and location between ground truth and prediction for larger anomalies, while smaller anomalies often blend together in the prediction. The background resistivity is often overestimated by the network. Due to the way training was performed, the applicability of the network is currently limited to a small number of subsurface scenarios. Despite its limitations, the speed of the prediction and the lack of required a-priori information are advantageous. Possible applications of neural networks in DC resistivity inversion lie in the generation of suitable starting models for other, more traditional inversion techniques. |
Session Keyword | 2.0 Theory, Modelling and Inversion |
File upload |
2.0_convolutional_neural_netw_spitzer_01.pdf
|