Details of the Abstract
| Title of paper | A Neuro-Physical Inverter for Magnetotelluric Data of Geothermal Systems |
| List of authors | Kim, J., Ravela, S., Evans, R. |
| Affiliation(s) |
MIT-WHOI Joint Program; Earth Signals and Systems Group, Dept. of Earth, Atmospheric and Planetary Sciences, Massachusetts Institute of Technology; Dept. of Geology and Geophysics, Woods Hole Oceanographic Institution |
| Summary | The magnetotelluric (MT) method is often used to image and characterize geothermal systems as MT data are sensitive to the presence of hot fluids and may help delineate thermally enhanced zones. The interpretation of MT data typically relies on inversion methods, where a subsurface electrical resistivity model is inferred and progressively updated by minimizing the discrepancies between the measured data and the forward simulated data from the resistivity model. However, geophysical inverse problems are notoriously non-unique and the interpretation of inverted models requires substantial amounts of hypothesis testing to arrive at reasonable conclusions. This motivates the use of machine learning to improve the analysis and inversion methods of MT data. Here, we propose the use of physics-coupled machine learning to invert for resistivity from MT data and use it for direct interpretation, bypassing the need for a fully numerical inversion. We use physical forward models to generate synthetic data from subsurface resistivity models and then use deep learning to reconstruct the resistivity model from the MT data. We rely on a self-supervised framework based on an autoencoder architecture, where one half is fixed by physics (i.e., Maxwell’s equations) and the other half is neural. This allows us to train physically consistent “inverters”, which capture both the physics of EM induction and the complexities of real-world data. We call this type of neural network “Neuro-Physical Inverters” (NPI). We apply our NPI to MT data and models from Gabbs Valley, Nevada, USA and compare with our baseline from Gauss-Newton inversions and ensemble-approximated Gaussian Processes. |
| Session Keyword | 2.0 EM theory, modelling and Inversion |
| File upload |
2.0_a_neuro-physical_inverter_kim_03.pdf
|