Efficient Uncertainty Quantification for Marine CSEM via Swi-DREI: A Physics-Driven Deep Reparameterization Ensemble Approach with Implicit Regularization
Abstract
Marine controlled-source electromagnetic (MCSEM) method serves as a pivotal geophysical tool for characterizing deep-sea hydrocarbon reservoirs or gas hydrates. However, restricted by the exponential decay of EM fields in conductive media and the spatial incompleteness of data acquisition systems, the inverse problem faces challenges of severe non-linearity and non-uniqueness. This study develops a Swish-Gated Linear Units (SwiGLU) based deep reparameterization ensemble Inversion framework, designated as Swi-DREI, for accelerated uncertainty quantification. By exploiting the implicit structural regularization in deep neural networks, the framework successfully resolves geological backgrounds and sharpens the boundaries of anomalies without requiring external training datasets. To efficiently map complex multi-modal posterior distributions, a parallelized deep ensemble strategy is executed, combined with a statistical selection scheme based on clustering and interface probabilities to eliminate the boundary blurring typical of conventional arithmetic averaging. Numerical experiments demonstrate that the proposed framework possesses significant advantages in delineating high-resistivity thin layers, maintaining robustness under challenging scenarios with 20% Gaussian noise and 45% missing data. Applications to field data further verify the practical feasibility of the proposed approach. The results indicate that the proposed method can successfully reconstruct high-resolution resistivity profiles matching actual geological sequences while providing quantitative risk assessments for target identification. This framework establishes a rigorous, training data-free paradigm for interpreting complex marine CSEM data and balances high-resolution physical imaging with high-confidence uncertainty evaluation.
Related
- Uncertainty analysis of a machine learning-aided ensemble forecast for coastal flood maps
- Uncertainty quantification for renewable energy generation from landfill methane
- Rigorous uncertainty quantification of probabilistic AI weather forecasts with conformal prediction
- Advancing Near-Term Water Quality Forecasting with Explainable Machine Learning and Probabilistic Uncertainty Quantification
- LFS-SEnDA: a global ocean forecast system with an efficient Sliding-Ensemble-based Data Assimilation scheme
BibTeX
@misc{liang2026efficient,
title = {{Efficient Uncertainty Quantification for Marine CSEM via Swi-DREI: A Physics-Driven Deep Reparameterization Ensemble Approach with Implicit Regularization}},
author = {Hao Liang and Changchun Yin and Zhanxiang He and Bo Zhang and Yunhe Liu and Xiuyan Ren and Yang Su and Thomas Hermans},
howpublished = {EarthArXiv},
year = {2026},
month = nov,
doi = {10.31223/x5vf92},
url = {https://eartharxiv.org/repository/view/15923/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Efficient Uncertainty Quantification for Marine CSEM via Swi-DREI: A Physics-Driven Deep Reparameterization Ensemble Approach with Implicit Regularization
AU - Hao Liang
AU - Changchun Yin
AU - Zhanxiang He
AU - Bo Zhang
AU - Yunhe Liu
AU - Xiuyan Ren
AU - Yang Su
AU - Thomas Hermans
PY - 2026
DA - 2026/11/03
PB - EarthArXiv
DO - 10.31223/x5vf92
UR - https://eartharxiv.org/repository/view/15923/
AB - Marine controlled-source electromagnetic (MCSEM) method serves as a pivotal geophysical tool for characterizing deep-sea hydrocarbon reservoirs or gas hydrates. However, restricted by the exponential decay of EM fields in conductive media and the spatial incompleteness of data acquisition systems, the inverse problem faces challenges of severe non-linearity and non-uniqueness. This study develops a Swish-Gated Linear Units (SwiGLU) based deep reparameterization ensemble Inversion framework, designated as Swi-DREI, for accelerated uncertainty quantification. By exploiting the implicit structural regularization in deep neural networks, the framework successfully resolves geological backgrounds and sharpens the boundaries of anomalies without requiring external training datasets. To efficiently map complex multi-modal posterior distributions, a parallelized deep ensemble strategy is executed, combined with a statistical selection scheme based on clustering and interface probabilities to eliminate the boundary blurring typical of conventional arithmetic averaging. Numerical experiments demonstrate that the proposed framework possesses significant advantages in delineating high-resistivity thin layers, maintaining robustness under challenging scenarios with 20% Gaussian noise and 45% missing data. Applications to field data further verify the practical feasibility of the proposed approach. The results indicate that the proposed method can successfully reconstruct high-resolution resistivity profiles matching actual geological sequences while providing quantitative risk assessments for target identification. This framework establishes a rigorous, training data-free paradigm for interpreting complex marine CSEM data and balances high-resolution physical imaging with high-confidence uncertainty evaluation.
ER -