📄 earth science

The Role of Geological Prior Information in Diffusion-Based Probabilistic Gravity Inversion

This study demonstrates that explicitly incorporating geological prior information into diffusion-based probabilistic gravity inversion significantly improves reconstruction accuracy and structural fidelity, though the optimal prior choice depends on whether the goal is minimizing error or preserving geological structure.

Original authors: Omid Mahmoodi

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Omid Mahmoodi

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 not a uniform block of rock but a complex tapestry of hidden layers, faults, and dense bodies that shift and stretch in ways we cannot see. To understand what lies below, geophysicists measure the subtle pull of gravity at the surface. Because gravity is a force that travels through everything, a heavy chunk of rock deep underground pulls slightly harder on a sensor than a lighter one nearby. By measuring these tiny variations, scientists attempt to work backward, calculating the shape and density of the hidden structures. However, this process is like trying to guess the shape of a hidden object solely by the shadow it casts; many different shapes could produce the exact same shadow. To solve this puzzle, scientists must bring in outside knowledge, or "priors," about what the geology is likely to look like. For decades, they have relied on mathematical rules that favor smooth, simple shapes, but these rules often miss the sharp edges of real-world features like faults or volcanic dikes. Recently, a new type of computer model, known as a diffusion model, has emerged. These models learn from thousands of examples to understand the complex, messy patterns of real geology, offering a way to generate realistic underground maps that respect the sharp boundaries nature actually creates.

In a recent study, researcher Omid Mahmoodi set out to test exactly how much this geological knowledge matters when using these advanced computer models. The central question was not just whether the models work, but how the specific amount of information provided to the computer changes the final answer. Imagine a geologist standing at a drilling site with three different levels of knowledge: they might know absolutely nothing about the local rock types, they might know exactly which family of rocks is present, or they might be unsure and have to consider several equally likely possibilities. Mahmoodi created a controlled digital experiment to simulate these three scenarios. He generated thousands of synthetic underground worlds, organized into six distinct families of shapes, ranging from smooth, horizontal layers to jagged, faulted blocks and narrow, vertical dikes. He then tasked the computer with reconstructing these hidden worlds using only the simulated gravity data, testing it under the three different information conditions: total ignorance, perfect knowledge of the rock family, and a balanced mix of all possibilities.

The results revealed that the amount of geological information fed into the system fundamentally alters the quality of the map produced. When the computer was given no specific information about the rock types, it still performed better than traditional methods, producing maps that were closer to the truth than the old, smooth-solution approaches. However, when the computer was told the exact family of rocks it was looking for, the maps became significantly sharper. The model successfully preserved the crisp edges of faults and the distinct shapes of intrusions, capturing the structural details that other methods blurred away. This suggests that if a geologist has strong evidence pointing to a specific geological setting, feeding that specific knowledge into the model yields the most accurate picture of the underground structure.

Yet, the study also uncovered a surprising trade-off when dealing with uncertainty. When the computer was asked to consider all six rock families at once, treating them as equally likely, the resulting map had the lowest overall numerical error. In other words, the average of all these different possibilities was mathematically closer to the true density values than any single guess. However, this statistical advantage came at a cost: the map looked smoother and less defined. By averaging out the different possibilities, the sharp contacts between rock types were softened, making the final image look less like a real geological formation and more like a hazy blend. The researchers found that while this "ensemble" approach reduced the average error, it failed to capture the specific, sharp structures that a geologist would need to identify a fault line or a mineral deposit.

The study also examined how well the computer could express its own uncertainty. In the scenarios where the computer was unsure, it produced a wider range of possible maps, which is a good sign that it knew it was guessing. However, even in these cases, the range of possibilities was not wide enough to fully cover the true answer, meaning the computer was still too confident in its guesses. This is a common limitation in current simulations, where the computer tends to underestimate the full scope of the unknown. The research concludes that there is no single "best" way to invert gravity data. If the goal is to find the most precise numerical values for density, averaging multiple scenarios works well. But if the goal is to see the actual shape of the rocks, knowing the specific geological family is essential. The choice of what information to provide to the model is not just a technical detail; it is a decision that determines whether the final map reveals the sharp, broken reality of the Earth or a smoothed-over approximation.

Technical Summary: The Role of Geological Prior Information in Diffusion-Based Probabilistic Gravity Inversion

Problem Statement
Gravity inversion is an inherently non-unique problem where the recovered subsurface density distribution depends heavily on prior assumptions. While conventional deterministic approaches (e.g., SimPEG) utilize smoothness or minimum-structure regularization to stabilize the inverse problem, these constraints fail to encode complex geological features such as contacts, faults, dykes, or compact intrusive geometries. Although diffusion models have emerged as powerful tools for encoding complex geological structures and performing probabilistic inversion, the specific impact of varying levels of geological information available prior to inversion on posterior behavior remains insufficiently quantified. This study addresses the gap in understanding how the "amount" of geological information—ranging from unknown to known to uncertain—affects reconstruction error, structural fidelity, and uncertainty quantification.

Methodology
The author implemented a controlled measurement-guided diffusion workflow to isolate the effects of geological prior information. The study utilized a synthetic benchmark comprising six procedurally generated geological families: blobs, horizontal layers, dipping layers, faulted layers, dykes, and intrusions. These families were selected to span distinct morphological axes (compact vs. distributed, subhorizontal vs. inclined, continuous vs. disrupted).

The experimental design compared three specific information states under an identical inversion framework:

Unknown Geology (Unconditional Prior): Inference uses a diffusion model trained on all six families without labels ( $p_\theta(m)$ ).
Known Geology (Conditional Prior): Inference uses a diffusion model conditioned on the correct synthetic family label ( $p_\phi(m | g)$ ). This serves as an "oracle" case.
Uncertain Geology (Scenario Ensemble): Inference runs the conditional sampler across all six families with equal weights, pooling the samples to form an ensemble ( $q_{ens}(m | d)$ ).

The inversion mechanism employed Diffusion Posterior Sampling (DPS) with a stabilized, delayed guidance update to ensure numerical stability. The benchmark included 100 fixed in-distribution (ID) and 100 fixed within-family out-of-distribution (OOD) test cases. Performance was evaluated against a deterministic SimPEG L2 baseline and across multiple metrics: Root-Mean-Square Error (RMSE), correlation, Structural Similarity Index (SSIM), Intersection over Union (IoU) for compact anomalies, and empirical interval coverage for uncertainty.

Key Contributions
The paper makes four primary contributions:

Controlled Hierarchy: A systematic implementation of geological-information states using fixed unconditional and conditional diffusion priors.
Common Benchmark: A standardized gravity-inversion benchmark utilizing both ID and OOD test cases to evaluate generalization within known conceptual families.
Scenario Ensemble Baseline: A simple equal-weight ensemble approach for unresolved geological interpretations, serving as a sensitivity analysis rather than a Bayesian model average.
Metric Disagreement Analysis: An evaluation demonstrating that pixelwise RMSE and structure-aware metrics (SSIM, IoU) can rank posterior estimates differently, highlighting trade-offs between error minimization and structural preservation.

Results

Reconstruction Error: All learned-prior settings reduced mean RMSE relative to the smooth SimPEG L2 baseline on both ID and OOD splits. The equal-weight scenario ensemble achieved the lowest ID mean RMSE (0.333), followed by the known prior (0.343) and the unknown prior (0.350).
Structural Recovery: Correct family conditioning (Known Prior) provided the strongest structural recovery, yielding the highest mean SSIM and anomaly overlap (IoU). While the known prior lowered RMSE in 53% of cases compared to the unknown prior, the mean difference was not statistically significant. However, the improvement in structural fidelity (SSIM) was clear.
Trade-offs: The scenario ensemble achieved lower aggregate RMSE by averaging across incompatible family-conditioned reconstructions. However, this averaging resulted in smoother estimates that attenuated contacts and compact bodies, leading to lower SSIM and IoU scores compared to the known prior.
Uncertainty Quantification: The scenario ensemble produced the broadest empirical intervals and highest coverage, though all methods remained under-dispersed relative to nominal levels (especially in OOD cases). Ensemble spread showed a positive correlation with absolute reconstruction error, suggesting it serves as a useful relative indicator of reconstruction difficulty rather than a calibrated Bayesian uncertainty.
Morphological Heterogeneity: The value of family labels varied by morphology; for instance, known information helped recover compact targets (dykes, intrusions) but did not uniformly improve performance across all families (e.g., OOD horizontal layers were sometimes better recovered by the unknown prior).

Significance and Claims
The paper claims that geological prior information fundamentally alters the character of the recovered solution, rather than producing a single universally superior method.

Explicit Specification: The study argues that geological prior specification must be treated explicitly when evaluating probabilistic gravity inversion, as the inversion favored by RMSE can differ from that favored by structure-aware metrics.
Interpretive Consequence: When independent evidence strongly supports one geological family, family-conditioned inference preserves expected morphology. When multiple interpretations are plausible, the scenario ensemble provides a transparent sensitivity baseline, though the pooled mean may not resemble any geologically plausible realization.
Limitations: The author maintains a modest scope, noting that the benchmark is synthetic, two-dimensional, and restricted to discrete single-family geology. Consequently, conclusions regarding mixed geology, unseen concepts, or 3D field data are not directly extrapolated. The OOD results specifically test parameter extrapolation within known families, not generalization to unseen geological concepts.

In conclusion, the research demonstrates that while learned priors generally outperform smooth deterministic baselines, the choice of prior information (unknown vs. known vs. uncertain) dictates a trade-off between minimizing cellwise error and preserving structural sharpness.

Technical Summary: The Role of Geological Prior Information in Diffusion-Based Probabilistic Gravity Inversion

More like this