{"id":{"repo_id":"woods-hole","oai_identifier":"oai:darchive.mblwhoilibrary.org:1912/73001"},"canonical_url":"https://search.dev.ndltd.org/etd/woods-hole/oai:darchive.mblwhoilibrary.org:1912/73001","repository":{"repo_id":"woods-hole","name":"Woods Hole Oceanographic Institute","base_url":"https://darchive.mblwhoilibrary.org/server/oai/request"},"display":{"title":"Physics-based and data-driven inversion of magnetotelluric data for subsurface imaging","abstract":"Electromagnetic (EM) geophysical inversion is fundamentally limited by the ill-posed and non-unique nature of the inverse problem, where diffusive physics, limited depth sensitivity, and observational noise lead to broad sensitivity kernels and multiple admissible subsurface models. Conventional deterministic approaches stabilize the inversion through regularization but depend strongly on prior assumptions and do not explicitly characterize uncertainty, while fully probabilistic methods are often computationally prohibitive for practical applications. These limitations are evident in field studies, where even high-quality inversions yield ambiguous interpretations of subsurface resistivity structure. This thesis introduces a hybrid inversion framework, termed the Neuro-Physical Inverter (NPI), that integrates ensemble-based conditioning with physics-guided residual learning. The approach combines an ensemble-approximated conditional Gaussian process (EnsCGP), which produces physically consistent resistivity estimates and ensemble-based uncertainty within the span of a prior model space, with a residual neural network that learns structured corrections to the ensemble-conditioned solution while preserving forward-physics consistency. Synthetic experiments demonstrate that this formulation yields stable reconstructions under realistic noise conditions and systematically reduces depth-dependent bias while preserving meaningful ensemble spread, avoiding overconfident collapse of the model distribution. The framework is applied to magnetotelluric data from the Gabbs Valley geothermal system, where it is adapted through physics-constrained fine-tuning. The resulting resistivity models reproduce observed responses and recover spatially coherent conductivity structures consistent with independent three-dimensional inversions, while ensemble variability highlights regions of limited constraint. These results demonstrate that EM geophysical inversion can be formulated as a composition of ensemble-based conditioning and constrained residual learning. More broadly, this work establishes a scalable and uncertainty-aware framework for integrating forward physics, ensemble-based conditioning, and machine learning in geophysical inverse problems.","abstract_html":"Electromagnetic (EM) geophysical inversion is fundamentally limited by the ill-posed and non-unique nature of the inverse problem, where diffusive physics, limited depth sensitivity, and observational noise lead to broad sensitivity kernels and multiple admissible subsurface models. Conventional deterministic approaches stabilize the inversion through regularization but depend strongly on prior assumptions and do not explicitly characterize uncertainty, while fully probabilistic methods are often computationally prohibitive for practical applications. These limitations are evident in field studies, where even high-quality inversions yield ambiguous interpretations of subsurface resistivity structure. This thesis introduces a hybrid inversion framework, termed the Neuro-Physical Inverter (NPI), that integrates ensemble-based conditioning with physics-guided residual learning. The approach combines an ensemble-approximated conditional Gaussian process (EnsCGP), which produces physically consistent resistivity estimates and ensemble-based uncertainty within the span of a prior model space, with a residual neural network that learns structured corrections to the ensemble-conditioned solution while preserving forward-physics consistency. Synthetic experiments demonstrate that this formulation yields stable reconstructions under realistic noise conditions and systematically reduces depth-dependent bias while preserving meaningful ensemble spread, avoiding overconfident collapse of the model distribution. The framework is applied to magnetotelluric data from the Gabbs Valley geothermal system, where it is adapted through physics-constrained fine-tuning. The resulting resistivity models reproduce observed responses and recover spatially coherent conductivity structures consistent with independent three-dimensional inversions, while ensemble variability highlights regions of limited constraint. These results demonstrate that EM geophysical inversion can be formulated as a composition of ensemble-based conditioning and constrained residual learning. More broadly, this work establishes a scalable and uncertainty-aware framework for integrating forward physics, ensemble-based conditioning, and machine learning in geophysical inverse problems.","abstract_has_math":false,"creators":["Kim, Jae Deok"],"institution":"Massachusetts Institute of Technology and Woods Hole Oceanographic Institution","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["Evans, Rob L.","Ravela, Sai"],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026-05","date_published":"2026-05","updated_at":"2026-07-27T22:05:19Z","subjects":["Magnetotellurics","Geophysical inversion","Machine learning in geophysics"],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier.doi","label":"DOI","values":["10.1575/1912/73001"],"render_values":[{"text":"10.1575/1912/73001","href":"https://doi.org/10.1575/1912/73001","code":true}]}]},"links":{"outbound_url":"https://hdl.handle.net/1912/73001","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Evans, Rob L.","Ravela, Sai"]},{"key":"dc:creator","label":"Author","values":["Kim, Jae Deok"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2026-05-29T20:05:10Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2026-05-29T20:05:10Z"]},{"key":"dc:date.issued","label":"Date","values":["2026-05"]},{"key":"dc:publisher","label":"Institution","values":["Massachusetts Institute of Technology and Woods Hole Oceanographic Institution"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Magnetotellurics","Geophysical inversion","Machine learning in geophysics"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.doi","label":"DOI","values":["10.1575/1912/73001"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/1912/73001"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy at the Massachusetts Institute of Technology and the Woods Hole Oceanographic Institution May 2026."]},{"key":"dc:description.abstract","label":"Abstract","values":["Electromagnetic (EM) geophysical inversion is fundamentally limited by the ill-posed and non-unique nature of the inverse problem, where diffusive physics, limited depth sensitivity, and observational noise lead to broad sensitivity kernels and multiple admissible subsurface models. Conventional deterministic approaches stabilize the inversion through regularization but depend strongly on prior assumptions and do not explicitly characterize uncertainty, while fully probabilistic methods are often computationally prohibitive for practical applications. These limitations are evident in field studies, where even high-quality inversions yield ambiguous interpretations of subsurface resistivity structure. This thesis introduces a hybrid inversion framework, termed the Neuro-Physical Inverter (NPI), that integrates ensemble-based conditioning with physics-guided residual learning. The approach combines an ensemble-approximated conditional Gaussian process (EnsCGP), which produces physically consistent resistivity estimates and ensemble-based uncertainty within the span of a prior model space, with a residual neural network that learns structured corrections to the ensemble-conditioned solution while preserving forward-physics consistency. Synthetic experiments demonstrate that this formulation yields stable reconstructions under realistic noise conditions and systematically reduces depth-dependent bias while preserving meaningful ensemble spread, avoiding overconfident collapse of the model distribution. The framework is applied to magnetotelluric data from the Gabbs Valley geothermal system, where it is adapted through physics-constrained fine-tuning. The resulting resistivity models reproduce observed responses and recover spatially coherent conductivity structures consistent with independent three-dimensional inversions, while ensemble variability highlights regions of limited constraint. These results demonstrate that EM geophysical inversion can be formulated as a composition of ensemble-based conditioning and constrained residual learning. More broadly, this work establishes a scalable and uncertainty-aware framework for integrating forward physics, ensemble-based conditioning, and machine learning in geophysical inverse problems."]},{"key":"dc:title","label":"Title","values":["Physics-based and data-driven inversion of magnetotelluric data for subsurface imaging"]}]}],"canonical_facts":{"dc:contributor.advisor":["Evans, Rob L.","Ravela, Sai"],"dc:creator":["Kim, Jae Deok"],"dc:date.accessioned":["2026-05-29T20:05:10Z"],"dc:date.available":["2026-05-29T20:05:10Z"],"dc:date.issued":["2026-05"],"dc:description":["Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy at the Massachusetts Institute of Technology and the Woods Hole Oceanographic Institution May 2026."],"dc:description.abstract":["Electromagnetic (EM) geophysical inversion is fundamentally limited by the ill-posed and non-unique nature of the inverse problem, where diffusive physics, limited depth sensitivity, and observational noise lead to broad sensitivity kernels and multiple admissible subsurface models. Conventional deterministic approaches stabilize the inversion through regularization but depend strongly on prior assumptions and do not explicitly characterize uncertainty, while fully probabilistic methods are often computationally prohibitive for practical applications. These limitations are evident in field studies, where even high-quality inversions yield ambiguous interpretations of subsurface resistivity structure. This thesis introduces a hybrid inversion framework, termed the Neuro-Physical Inverter (NPI), that integrates ensemble-based conditioning with physics-guided residual learning. The approach combines an ensemble-approximated conditional Gaussian process (EnsCGP), which produces physically consistent resistivity estimates and ensemble-based uncertainty within the span of a prior model space, with a residual neural network that learns structured corrections to the ensemble-conditioned solution while preserving forward-physics consistency. Synthetic experiments demonstrate that this formulation yields stable reconstructions under realistic noise conditions and systematically reduces depth-dependent bias while preserving meaningful ensemble spread, avoiding overconfident collapse of the model distribution. The framework is applied to magnetotelluric data from the Gabbs Valley geothermal system, where it is adapted through physics-constrained fine-tuning. The resulting resistivity models reproduce observed responses and recover spatially coherent conductivity structures consistent with independent three-dimensional inversions, while ensemble variability highlights regions of limited constraint. These results demonstrate that EM geophysical inversion can be formulated as a composition of ensemble-based conditioning and constrained residual learning. More broadly, this work establishes a scalable and uncertainty-aware framework for integrating forward physics, ensemble-based conditioning, and machine learning in geophysical inverse problems."],"dc:identifier.doi":["10.1575/1912/73001"],"dc:identifier.uri":["https://hdl.handle.net/1912/73001"],"dc:publisher":["Massachusetts Institute of Technology and Woods Hole Oceanographic Institution"],"dc:subject":["Magnetotellurics","Geophysical inversion","Machine learning in geophysics"],"dc:title":["Physics-based and data-driven inversion of magnetotelluric data for subsurface imaging"],"dc:type":["Thesis"]},"updated_at":"2026-07-27T22:05:19Z"}