Advanced Microgravity Inversion Workflows for Uncharted Void Detection

When a gravity survey over a suspected karst terrain returns a -0.06 mGal residual anomaly at the center of a 10 m grid, the question isn't whether there's a void—it's how deep, how large, and how to interpret the inversion when the target is barely above the noise floor. This article is for geophysicists and survey managers who already understand Bouguer corrections and free-air gradients. We focus on the inversion decisions that separate a plausible cavity model from a statistical artifact.

Why Standard Inversion Falls Short for Uncharted Voids

Most commercial inversion packages assume a starting model with known density contrasts. For uncharted voids—caves, abandoned mines, sinkholes without surface expression—you lack that prior. The inversion becomes an ill-posed problem where multiple density distributions fit the same gravity data equally well. Without careful constraints, the result is often a smeared low-density region that tells you little about the void's geometry or location.

Practitioners frequently report that smoothness-constrained inversions, which minimize spatial gradients, produce blob-like anomalies that misrepresent sharp void boundaries. A void with a roof collapse, for example, generates a density contrast that changes abruptly at the cavity edge. Smoothness constraints smear that edge over several meters, leading to overestimates of void volume and underestimates of depth. One team I read about spent weeks drilling test holes based on a smooth inversion, only to find the actual cavity was 2 m shallower and 30% smaller than predicted.

The core problem is that gravity data alone cannot uniquely determine depth and density contrast simultaneously. A shallow, low-density anomaly can mimic a deeper, higher-contrast void. This non-uniqueness demands that you inject external constraints—geological boundaries, borehole logs, or joint inversion with other geophysical methods—early in the workflow.

When You Should Not Use Microgravity Alone

If the target void is deeper than twice its diameter, the gravity signal becomes too weak to invert reliably without additional constraints. Similarly, in areas with steep topography or heterogeneous overburden, the residual anomaly from the void may be indistinguishable from terrain-correction errors. In those cases, consider microgravity as a reconnaissance tool rather than a standalone inversion input.

Prerequisites: Data Quality First

Before running any inversion, you must settle three data-quality prerequisites: station spacing relative to target depth, temporal drift correction, and terrain density assumptions. Skipping these leads to inversion artifacts that no algorithm can fix.

Station spacing should be no larger than one-half the expected void depth to satisfy the Nyquist criterion for gravity anomalies. For a void at 15 m depth, stations at 7.5 m intervals are the minimum; 5 m is safer. Many surveys fail because the grid is too coarse for the target depth, and the inversion interpolates across a gap that never contained a measurement.

Temporal drift from instrument creep and tidal effects must be removed to better than 0.01 mGal. Loop closures with a base station reoccupation every 1–2 hours are standard, but for microgravity voids you need tighter control. Use a second instrument as a stationary reference if possible, or apply continuous GPS-tide corrections. Drift residuals of 0.02 mGal can mask a void anomaly that is only 0.05 mGal.

Terrain density is the most overlooked prerequisite. The Bouguer correction assumes a uniform density, typically 2.67 g/cm³ for sedimentary rock. But if your site has a 3 m layer of dry sand (1.6 g/cm³) over limestone, the terrain correction will be wrong by several tenths of a mGal. Measure density from outcrop samples or use a density logger in shallow boreholes. If that's not possible, run the inversion with a range of plausible densities and check whether the void model persists across all trials.

Elevation Accuracy Requirements

With a vertical gravity gradient of 0.3086 mGal/m, a 5 cm elevation error produces a 0.015 mGal error—enough to corrupt small void signals. Use real-time kinematic GPS with vertical accuracy better than 3 cm, or a total station for small grids. Correct for building or tree canopy effects that degrade GPS vertical precision.

Core Workflow: Sequential Inversion with Quality Gates

The workflow we recommend has four stages: data preparation, forward modeling of the regional field, inversion with constraints, and validation through synthetic testing. Each stage includes a quality gate that must pass before proceeding.

Stage 1: Data Preparation and Residual Separation

After standard corrections (drift, tide, latitude, free-air, Bouguer, terrain), compute the residual anomaly by subtracting a regional field. For uncharted voids, the regional is often a polynomial surface fitted to stations far from the suspected anomaly. But polynomial order matters: a second-order polynomial may absorb the void signal if the survey area is small. Cross-validate by fitting the polynomial to stations outside the anomaly zone, then predicting inside it. If the predicted regional shifts the anomaly center, try a different order or use a physically based regional from a known density model of the surrounding geology.

Quality gate: The residual anomaly should have a signal-to-noise ratio (peak anomaly divided by standard deviation of far-field residuals) of at least 3. If not, reoccupy stations to reduce noise or increase station density.

Stage 2: Forward Model of the Regional Background

Build a simple forward model of the background geology using any known interfaces (bedrock depth, fault contacts). This model is not the inversion target; it is used to remove predictable gravity effects from the residual. For example, if a limestone-shale contact crosses the survey area, forward-calculate its gravity effect and subtract it. This step prevents the inversion from assigning density contrasts to the void that actually belong to known geological boundaries.

Stage 3: Inversion with Constraints

Choose an inversion algorithm based on your prior knowledge. We compare three approaches in the next section. Regardless of algorithm, constrain the inversion with a density contrast range (e.g., -1.5 to -2.0 g/cm³ for air-filled voids) and a depth-weighting function that penalizes solutions far from the surface. Without depth weighting, the inversion tends to place anomalies at the surface because shallow sources produce the strongest signal.

Run the inversion with multiple starting models—a homogeneous half-space, a layered model, and a model with a small seed anomaly at the suspected location. If all converge to similar void geometry, confidence increases. If they diverge, the non-uniqueness is severe and you need additional data.

Quality gate: The modeled gravity data should fit the observed residuals within the noise level (typically 0.01–0.02 mGal). A chi-squared test per station can identify outliers that may indicate a bad measurement or an unmodeled shallow feature.

Stage 4: Synthetic Validation

Create a synthetic void model with known geometry and density, forward-compute its gravity, add realistic noise (0.01 mGal), then run the same inversion on the synthetic data. If the inversion recovers the void to within 20% of its true dimensions, you have confidence in the field inversion. If not, adjust constraints or station spacing and retest.

Tools, Setup, and Environment Realities

The inversion software landscape ranges from academic codes (e.g., UBC-GIF, ModelVision) to proprietary packages. Each has strengths and weaknesses for void detection. We focus on setup decisions that affect field outcomes.

Choice of Inversion Algorithm

Three algorithm families are common for microgravity void inversion. Smoothness-constrained (Occam-type) inversions are robust and fast but smear boundaries. Sharp-boundary inversions (e.g., with L1 norm regularization) preserve edges but require careful tuning of the regularization parameter. Probabilistic inversions (Markov chain Monte Carlo) explore the full solution space, giving uncertainty estimates, but are computationally expensive and require a prior density distribution.

For uncharted voids, we recommend a hybrid: start with a smooth inversion to locate the anomaly centroid, then run a sharp-boundary inversion with the centroid as a seed and with the regularization parameter chosen by L-curve analysis. Validate with a probabilistic run on a reduced grid if computational resources allow.

Field Setup Considerations

Instrument drift is worse in hot or windy conditions. Shade the gravimeter and allow 15 minutes for thermal equilibration after setup. On uneven terrain, use a bipod or tripod with a circular level to ensure consistent instrument height. Record the height above ground for each station; a 2 cm variation changes the free-air correction by 0.006 mGal.

Urban environments introduce additional noise: passing vehicles, vibrating machinery, and nearby buildings that produce gravity gradients. Schedule surveys during low-traffic hours and avoid stations within 5 m of large structures unless you model their gravitational effect.

Variations for Different Constraints

Not every project can afford a 5 m grid or a probabilistic inversion. Here are variations for common constraints: limited budget, rough terrain, and time pressure.

Limited Budget: Sparse Grid Plus Forward Modeling

If you can only afford 50 stations over a 100 m x 100 m area, do not attempt a full inversion. Instead, use a forward-modeling approach: hypothesize a void location from surface features (sinkholes, vegetation changes), then forward-calculate the expected anomaly for a range of depths and sizes. Compare with observed residuals using a chi-squared test. This approach cannot discover unknown voids but can test specific targets.

Rough Terrain: Adaptive Station Placement

In steep terrain, the terrain correction error dominates. Place stations on contour lines to simplify the correction, or use LiDAR-derived digital elevation models with 1 m resolution to compute terrain effects more accurately. In such conditions, the inversion should include the terrain as a known density boundary (air-rock contact) rather than subtracting it fully, because the terrain correction may remove part of the void signal if the void is near the surface.

Time Pressure: Fast Inversion with Fixed Depth Weighting

When you need results in 24 hours (e.g., before drilling decisions), run a smooth inversion with aggressive depth weighting (e.g., inverse of depth squared) to stabilize the solution. Accept that the void boundaries will be fuzzy. Use the centroid location and an estimated depth from the peak anomaly's half-width to plan a single borehole. Validate after drilling with a forward model of the actual void.

Pitfalls, Debugging, and What to Check When It Fails

Inversion failures usually fall into three categories: overfitting to noise, misinterpreting shallow versus deep sources, and ignoring non-uniqueness. Here is how to diagnose each.

Overfitting to Noise

If the inversion fits the data to 0.005 mGal but the void model looks like a spiky cluster of small anomalies, you have overfit the noise. Check the residual after inversion: it should be random, not structured. If you see a pattern (e.g., a bullseye at a station), that station may have a measurement error. Reoccupy it. If the pattern persists, increase the regularization parameter or use a coarser mesh.

Shallow Versus Deep Ambiguity

A classic symptom: the inversion places a shallow, wide anomaly when the true void is deeper and narrower. This happens because a deep void produces a broader, lower-amplitude anomaly that the inversion misinterprets as a shallow, low-density layer. To test, run the inversion with a fixed depth constraint—if the misfit increases significantly, the data cannot distinguish depth. In that case, add a seismic or GPR profile to constrain depth independently.

Ignoring Non-uniqueness

The worst pitfall is presenting a single inversion result as the truth. Always run at least three inversions with different starting models and compare them. If they differ by more than 50% in void volume, report the range. Use a probabilistic inversion to generate a posterior distribution of depth and size, and present the 90% confidence interval.

Frequently Asked Questions and Common Mistakes

Below we answer common questions that arise when applying this workflow, written in prose to provide context.

What is the minimum detectable void size for microgravity?

There is no fixed threshold because detectability depends on depth, density contrast, and noise level. As a rule of thumb, a spherical void at depth equal to its diameter produces a peak anomaly of about 0.04 mGal for a density contrast of -2.0 g/cm³. With 0.01 mGal noise, that void is detectable but near the limit. For shallower voids, smaller sizes are detectable. We recommend running a forward-model sensitivity test before the survey to determine the minimum size for your site conditions.

Can I invert gravity data without terrain correction?

No. Even gentle topography produces gravity variations comparable to void signals. A 2 m hill with density 2.0 g/cm³ creates a ~0.1 mGal anomaly. Without terrain correction, the inversion will fit the hill rather than the void. Always compute terrain effects using a digital elevation model with resolution at least the station spacing.

How do I integrate microgravity inversion with GPR?

GPR provides high-resolution depth information in conductive soils but may not penetrate to void depth if clay is present. Use GPR to identify shallow reflectors and constrain the top of the void, then use the microgravity inversion to estimate the void's lateral extent and lower boundary. Joint inversion algorithms that combine gravity and GPR traveltimes are still research-grade, but a pragmatic approach is to fix the void top from GPR and invert only for the bottom and density contrast.

Common Mistake: Using a Uniform Density for Bouguer Correction

Many practitioners use 2.67 g/cm³ without measuring site density. If the actual density is 2.0 g/cm³ (e.g., weathered sandstone), the Bouguer correction error is 0.67 g/cm³ times the thickness of the layer, potentially introducing a regional gradient that masks the void. Always measure density from samples or use a density profile from a nearby borehole.

What to Do Next: Specific Actions After Inversion

Completing an inversion is not the end. The following steps will validate your model and guide drilling or further surveying.

First, reoccupy three to five stations with the largest residuals to verify that the anomaly is repeatable. If the gravity value changes by more than 0.01 mGal between occupations, the station may have a setup error or near-surface disturbance. Exclude those stations and rerun the inversion.

Second, build a synthetic model that matches your inverted void geometry, forward-compute its gravity, and add random noise at the level observed in the field. Run the same inversion on the synthetic data. If the inversion recovers the synthetic void to within 20% in volume and depth, you have confidence in the field result. If not, adjust constraints or consider that the void may be more complex (e.g., partially filled).

Third, plan a validation borehole or geophysical line. For a void predicted at 12 m depth, drill to 15 m with continuous core logging. If the void is not found, re-examine the inversion: perhaps the anomaly came from a low-density clay lens rather than a cavity. Use electrical resistivity tomography along the same line to distinguish between a void (high resistivity) and a clay lens (low resistivity).

Fourth, document the entire workflow, including all starting models, regularization parameters, and quality gate results. This documentation is essential if the inversion is used for regulatory approval or litigation. A transparent workflow that acknowledges uncertainty is more credible than a single deterministic model.

Finally, consider publishing your results in an open repository (with location anonymized) to help the community build a database of microgravity void signatures. Empirical benchmarks from real voids are rare, and each well-documented case improves inversion practice for everyone.