⚡ electrical engineering

Efficient uncertainty quantification in FDEM inversion using an adjusted Bayesian information criterion

This study introduces a computationally efficient framework for frequency-domain electromagnetic (FDEM) inversion that utilizes an effective-degrees-of-freedom-adjusted Bayesian information criterion (EDF-BIC) to perform layer-count model averaging, thereby providing robust uncertainty quantification for resistivity structures without the prohibitive computational cost of traditional Bayesian methods.

Original authors: Zekai Deng, Hai Li, Ziteng Li

Published 2026-11-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zekai Deng, Hai Li, Ziteng Li

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Beneath the Earth's surface, the ground is rarely uniform. It is a complex stack of layers, some made of solid rock, others of water-saturated soil or mineral-rich veins. To understand what lies below, geophysicists use a technique called frequency-domain electromagnetic induction. They send electromagnetic signals into the ground and measure how the Earth responds at different frequencies. These signals travel differently depending on the electrical resistance of the materials they pass through, allowing scientists to build a picture of the subsurface. However, this process is not like taking a photograph where the image is clear and unique. Instead, it is more like trying to guess the shape of a hidden object by listening to the echo of a sound bouncing off it. Many different arrangements of rock layers could produce the exact same echo. This ambiguity is known as non-uniqueness. When scientists run a computer model to interpret these signals, they must make choices about how many layers to include in their guess. If they choose too few, they miss important details; if they choose too many, they might invent features that aren't really there. The challenge is to find the right balance and to understand how much confidence we can place in the final picture.

A team of researchers at the Institute of Geology and Geophysics in Beijing has developed a new way to handle this uncertainty. They created a framework that does not just pick a single "best" model of the underground layers but instead combines many different possibilities to create a more honest and complete picture. Their approach focuses on a specific problem: how to count the complexity of a model when the data has been smoothed out by mathematical rules. In standard methods, the complexity of a model is often counted by simply tallying the number of variables, such as the number of layers and their thicknesses. But when the data is smoothed to remove noise, these variables are no longer fully independent. The researchers realized that counting the raw number of variables over-penalizes complex models, leading scientists to prefer simpler, flatter models that might miss real geological features. To fix this, they introduced a method that measures the "effective" number of variables actually used to fit the data, rather than the total number of variables available.

The researchers tested this new method, which they call an adjusted information criterion, using computer-generated models of the Earth's subsurface. In one test, they simulated a five-layer underground structure. When they used the traditional counting method, the computer tended to favor a model with only four layers, missing the true complexity. When they applied their new adjusted method, the result shifted to favor a five-layer model, correctly identifying the true structure. This adjustment allowed the system to recognize that the extra layer was supported by the data, even after the smoothing rules were applied. The team then moved to more complex simulations where layers thinned out and disappeared, mimicking real geological pinch-outs. In these scenarios, the new method showed that the "best" number of layers changes from place to place along a survey line. In some spots, the data supported a simple two-layer model, while just a few meters away, a three-layer model was better.

Instead of forcing a single choice, the researchers combined the results from all the supported models. They assigned a weight to each model based on how well it fit the data, creating a weighted average that blended the different possibilities. This process, called layer-count model averaging, produced a final image that was smoother and more continuous than picking just one model. It also allowed them to separate the uncertainty into two types. The first type is the uncertainty about the exact thickness or resistance of a layer when the number of layers is fixed. The second type is the uncertainty about the number of layers itself. In their simulations, the uncertainty about the number of layers was highest exactly where the geological structures were changing, such as where a layer was pinching out or an interface was shifting. This helped them distinguish between features that were robust and those that depended heavily on the specific choice of model.

To ensure their method was reliable, the researchers ran fifty separate experiments with slightly different random noise added to the data. They found that their method consistently recovered the correct structures and that the range of uncertainty they calculated contained the true answer about 84 to 88 percent of the time. This level of accuracy suggests the method is stable and not overly sensitive to random errors in the data. Finally, they applied their technique to real-world data collected from the Yemaquan iron-lead-zinc deposit in western China. The survey lines covered hundreds of meters of terrain with varying geology. The new method successfully identified the main electrical structures, such as large resistive rock bodies and conductive zones, which matched previous independent studies. However, it also revealed that the exact positions of the boundaries between these zones were less certain. The method showed that while the large-scale picture was clear, the fine details of where a layer started or ended varied depending on the model chosen.

The computational cost of this new approach was surprisingly low. The researchers noted that calculating the effective complexity and averaging the models added less than one percent to the time required to run the standard inversions. This efficiency is crucial because it means the method can be used alongside existing workflows without requiring massive new computing power. By reusing the mathematical derivatives already calculated during the standard inversion, they avoided the need for millions of extra calculations that other probabilistic methods often require. The result is a practical tool that helps geophysicists see not just a single version of the underground, but a spectrum of possibilities. It highlights where the data is strong enough to define a structure and where the interpretation remains flexible. In the end, the work provides a clearer, more nuanced view of the Earth's hidden layers, turning a single guess into a well-informed range of possibilities.

Technical Summary: Efficient Uncertainty Quantification in FDEM Inversion Using an Adjusted Bayesian Information Criterion

Problem Statement
Frequency-domain electromagnetic (FDEM) inversion is inherently non-unique; a single set of observations can be explained by various combinations of resistivity, layer thickness, and model parameterization. While regularization stabilizes the inversion to yield a preferred solution, conventional deterministic approaches typically require the model dimension (specifically the number of layers, $K$ ) to be specified in advance. Selecting a single "best" layer count ignores the uncertainty arising from the choice of dimension itself, potentially overstating confidence in local interfaces and fine-scale structures.

Existing Bayesian approaches, such as trans-dimensional Markov chain Monte Carlo (MCMC), can characterize this uncertainty but often require millions of forward evaluations, making them computationally prohibitive for large-scale or routine applications. Conversely, standard information criteria like the Bayesian Information Criterion (BIC) are computationally efficient but often fail to account for the reduction in effective model complexity caused by regularization. Standard BIC uses the nominal parameter count as a complexity penalty, which can overpenalize regularized models with higher layer counts because regularization limits the independent contribution of added parameters to the data fit.

Methodology
The authors propose a framework for layer-count assessment and model-averaged uncertainty quantification that integrates with conventional regularized FDEM inversion. The workflow consists of three primary components:

Effective Degrees of Freedom (EDF) Adjustment:
The study replaces the nominal parameter count in the BIC formula with the effective degrees of freedom (EDF). The EDF is calculated as the trace of the data-space influence matrix ( $S_K = J_d H_K^{-1} J_d^T$ ), where $J_d$ is the Jacobian of the data residuals and $H_K$ is the Gauss-Newton Hessian. This metric quantifies the actual data-fitting flexibility retained after accounting for regularization, parameter correlation, and data sensitivity. The resulting criterion, EDF-BIC, is defined as:
$\text{EDF-BIC}_K = \chi_K^2 + \nu_K \ln N$
where $\chi_K^2$ is the weighted residual sum of squares, $\nu_K$ is the EDF, and $N$ is the number of data points.
Layer-Count Model Averaging (LCMA):
Instead of selecting a single preferred layer count, the framework evaluates a finite set of candidate models (inverted sequentially with increasing $K$ ). An adaptive stopping rule (2% relative improvement threshold) limits the expansion of candidate layer counts. The differences in EDF-BIC among candidates are converted into relative-support weights ( $w_{i,K}$ ) at each station. These weights are used to perform LCMA, where candidate models are mapped to a common depth grid and averaged in log-resistivity space.
Decomposition of Uncertainty:
Using the law of total variance, the total model-averaged uncertainty is decomposed into two components:
- Within-model uncertainty: Parameter uncertainty (resistivity, thickness) conditional on a fixed layer count, approximated via a Laplace approximation using the Gauss-Newton Hessian.
- Between-model uncertainty: Variance arising from differences in the mean resistivity structures of the supported layer counts.
  The framework combines these to produce P10–P90 credible intervals for the final resistivity structure.

Key Results

Synthetic Experiments (U05, G02, G05):
- In the U05 five-layer model, the EDF adjustment corrected the low-layer-count preference of nominal BIC. The weighted mean layer count shifted from 4.06 (nominal BIC) to 5.15 (EDF-BIC), significantly reducing the deviation from the true five-layer configuration.
- In laterally varying models (G02 and G05), EDF-BIC support varied spatially, correctly identifying regions where layers pinch out or disappear.
- LCMA recovered principal electrical structures with lower root-mean-square errors (RMSE) compared to fixed-dimension or maximum-support models.
- Across 50 noise realizations, the EDF-BIC LCMA approach achieved mean empirical coverage of 87.57% (G02) and 83.66% (G05) for P10–P90 intervals, demonstrating stability and robustness.
- Uncertainty decomposition revealed that while within-model uncertainty dominated globally, between-model uncertainty was concentrated in structural transition zones (e.g., pinch-outs), highlighting where layer-count choice most affects interpretation.
Field Data Application (Yemaquan L52 and L53):
- Applied to FDEM data from a polymetallic deposit, the framework recovered consistent principal resistive and conductive structures across candidate layer counts.
- While large-scale electrical zonation was robust, local interface positions and deep extensions showed higher sensitivity to layer-count choice, reflected in increased between-model uncertainty in transition zones.
- The computational cost of EDF calculation and uncertainty analysis was negligible, adding less than 1% to the total candidate-inversion time.

Significance and Claims
The paper claims to provide a computationally tractable approach to layer-count assessment and uncertainty propagation in regularized FDEM inversion. Its primary significance lies in:

Bridging the gap between efficiency and rigor: It offers a method to quantify model-dimension uncertainty without the prohibitive cost of full posterior sampling (MCMC).
Improved complexity penalization: By substituting nominal parameter counts with EDF, the method provides a more accurate measure of model complexity in regularized inversions, preventing the underestimation of necessary layer counts.
Explicit uncertainty propagation: The framework moves beyond selecting a single "best" model to propagating uncertainty from both parameter estimation and model dimension choice into the final resistivity structure.
Practical applicability: The method reuses Jacobians and Hessians from standard deterministic inversions, making it easily integrable into existing workflows with minimal additional computational overhead.

The authors maintain a modest stance, noting that the reported uncertainties are conditional on the chosen candidate set, error model, and regularization. They acknowledge that the Laplace approximation describes local uncertainty and that the method does not capture uncertainty in regularization parameters or broader model classes, nor does it fully address strong non-Gaussianity or multimodality in the posterior distribution. However, for the purpose of assessing layer-count uncertainty in regularized FDEM, the framework offers a robust and efficient alternative to full Bayesian sampling.

More like this