📄 earth science

ALTYN Inversion: A MATLAB-Based 2D/2.5D Finite-Strike Gravity Inversion Framework with Box-Constrained IRLS Regularization

This paper introduces ALTYN, a transparent MATLAB-based 2D/2.5D gravity inversion framework utilizing box-constrained IRLS regularization and a projected Newton solver to generate geophysically admissible density models, demonstrating through synthetic and field tests that while the method effectively fits data and highlights shared structural features, it cannot uniquely reconstruct complex subsurface geometries due to inherent non-uniqueness.

Original authors: Ahmad Ali Muckharom

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmad Ali Muckharom

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

The Earth is not a uniform sphere of rock; it is a complex, layered cake of materials with different weights. Some parts are dense and heavy, like solid granite or metal-rich ore, while others are light and porous, like fractured volcanic ash or water-saturated soil. Because gravity is a force that pulls on mass, these hidden differences in density create tiny, measurable variations in the pull of gravity at the surface. Geophysicists use this fact to look beneath their feet without digging. They measure the strength of gravity at many points along a line and then try to work backward to figure out what is down there. This process, known as gravity inversion, is like trying to guess the shape of a hidden object inside a box by only feeling how heavy the box feels from the outside. It is a powerful tool for finding oil, minerals, and geothermal energy, but it is also notoriously tricky. The same pattern of gravity can often be created by many different shapes and depths of rock, meaning there is no single, obvious answer to what lies below.

To tackle this challenge, a researcher named Ahmad Ali Muckharom has developed a new digital framework called ALTYN, designed to map these hidden density variations with greater clarity and honesty. The tool is built to handle a specific type of geological problem: when a rock formation stretches out sideways but not forever. In many real-world scenarios, a layer of rock or a volcanic body does not extend infinitely in both directions as older, simpler models assumed. Instead, it has a finite length, like a long, thick slab of stone that ends after a certain distance. ALTYN accounts for this by treating the underground as a series of rectangular blocks that have a defined width, allowing the computer to calculate the gravity pull more accurately than if it assumed the rock went on forever. The system also respects the actual shape of the ground above, adjusting its calculations to match hills and valleys rather than pretending the surface is flat.

The researchers tested this new framework using three carefully constructed computer simulations where they knew the exact answer beforehand. In the simplest test, they created a single, compact blob of heavy rock underground. The ALTYN system successfully found where the blob was and correctly identified that it was heavy, but it could not perfectly recreate its sharp edges. Instead, the recovered image showed the heavy rock as a slightly blurry, stretched-out shape. This is a common limitation in gravity studies: the system can tell you where a heavy object is, but it struggles to say exactly how deep it goes or how sharp its boundaries are. In a second test with two separate blobs, one heavy and one light, the system correctly found both and distinguished their opposite natures. However, the light blob appeared deeper and more spread out than it really was. The most difficult test involved a complex arrangement of light and heavy rocks stacked on top of each other beneath a hilly surface. Here, the system managed to match the gravity measurements perfectly, yet it failed to reconstruct the deeper heavy rock in its true location. Instead, it distributed the weight in a different way that produced the same surface reading. This result highlights a crucial truth: a perfect match to the gravity data does not guarantee that the picture of the underground is correct.

To see how different mathematical choices affect the final picture, the team applied the tool to a real-world gravity survey from the Dieng geothermal field in Indonesia. This area is a volcanic landscape with a long profile of measurement points stretching over 9 kilometers. The researchers ran the data through the system twice, using two different mathematical approaches to smooth out the results. The first approach, known as L2 regularization, produced a model where the density changes were gradual and spread out, like a soft watercolor painting. The second approach, called IRLS, allowed for sharper, more distinct boundaries, creating a model that looked more like a blocky, segmented structure. Both methods fit the actual gravity measurements equally well, satisfying the statistical requirements for a good match. However, the two resulting maps of the underground looked quite different in their details. The areas where both maps agreed—specifically the location of a large, heavy zone beneath a mountain and a broad, light zone nearby—were considered the most reliable findings. The differences between the two maps, such as exactly how deep the heavy zone extended or how sharp its edges were, were identified as less certain.

The study concludes that while the ALTYN framework is a transparent and effective way to build geologically plausible models, it cannot magically reveal a unique, perfect picture of the subsurface. The system is best viewed as a tool for constructing and comparing different possible versions of the underground, rather than as a machine that spits out a single, definitive truth. The most trustworthy information comes from the broad features that remain consistent across different methods, such as the general location and weight of major rock bodies. The finer details, like the exact depth of a fault or the precise shape of a volcanic intrusion, remain open to interpretation and require other types of data to confirm. In the end, the work provides a clearer, more honest way to look at the Earth's hidden density, reminding us that even with advanced mathematics, the deep Earth often keeps some of its secrets.

Technical Summary: ALTYN Inversion Framework

Problem Statement
Gravity inversion is fundamentally ill-posed and non-unique; distinct subsurface density distributions can generate identical or nearly identical gravity fields at the surface. This ambiguity is exacerbated by the trade-off between data fit and model complexity, the decay of sensitivity with depth, and the difficulty of representing geological structures that have finite strike lengths rather than infinite extent. Conventional 2D modeling often assumes infinite strike, while full 3D inversion incurs high computational costs. Furthermore, existing tools like the GROWTH family of software utilize a "model-growth" approach where predefined density contrasts are assigned to selected cells to build geometry. There is a need for a transparent, constrained framework that estimates continuous density contrasts on a fixed mesh while explicitly accounting for finite-strike physics, topography, and various regularization strategies without claiming unique geological reconstruction.

Methodology
The paper introduces ALTYN Inversion, a MATLAB-based 2D/2.5D framework designed for profile gravity inversion. The methodology integrates several key components:

Forward Modeling (2.5D Finite-Strike): The subsurface is discretized into a 2D X–Z mesh where each cell is treated as a 3D rectangular prism with a prescribed finite strike length ( $L_y$ ). The forward operator uses analytical expressions for the vertical gravitational attraction of a rectangular prism (Nagy et al., 2000), accounting for the finite extent perpendicular to the profile. This avoids the infinite-strike assumption of standard 2D models without requiring a full 3D density parameterization.
Topography and Discretization: The model uses a topography-consistent X–Z mesh. Cells above the surface are excluded, and the uppermost active cells are truncated to match local surface elevations. This geometry is passed directly to the forward kernel, ensuring consistency between the topography, the sensitivity matrix, and the regularization terms.
Data Fitting and Uncertainty: The framework employs uncertainty-weighted data fitting. If specific standard deviations are unavailable, it constructs a combined absolute-relative uncertainty model. The objective function minimizes a weighted data misfit term ( $\phi_d$ ) alongside a model regularization term ( $\phi_m$ ).
Regularization Strategies: Two primary regularization schemes are implemented:
1. L2 Regularization: A quadratic minimum-structure approach ( $P=2$ ) favoring smooth, distributed density variations.
2. IRLS (Iteratively Reweighted Least Squares): A non-quadratic approach ( $P < 2$ ) to promote focused or piecewise-constant models. The implementation allows for directional control; for instance, the field application uses $P_z = 0.5$ for vertical gradients while maintaining $P_x = P_s = 2$ , encouraging sharp vertical transitions while preserving horizontal smoothness.
Constraints and Solver: The inversion enforces explicit box constraints (density bounds) on every active cell during the optimization process, rather than clipping results post-inversion. It utilizes a projected Newton/active-set preconditioned conjugate-gradient (PCG) solver. A regional nuisance component (constant or linear trend) is estimated simultaneously with the density model to prevent long-wavelength trends from being misinterpreted as subsurface structure.
Convergence: The algorithm uses a cooling schedule for the regularization trade-off parameter ( $\beta$ ) and terminates based on a normalized discrepancy criterion (target WRMS $\approx$ 1) or relative model change.

Key Contributions
The primary contribution of ALTYN is not the introduction of a novel regularizer, but the transparent integration of established concepts into a single, reproducible profile-inversion framework. Key features include:

The combination of finite-strike prism physics with topography-consistent active-cell discretization.
The simultaneous estimation of density contrasts and regional nuisance parameters within a box-constrained optimization loop.
A comparative analysis of L2 and IRLS regularization on the same dataset to demonstrate how different priors select different geophysically admissible models.
A rigorous distinction between data-space fit (residuals) and model-space recovery (geometric fidelity), emphasizing that a good data fit does not guarantee a correct geological model.

Results
The framework was validated through three synthetic benchmarks of increasing complexity and applied to a 171-station gravity profile (9.23 km) in the Dieng geothermal area, Indonesia.

Synthetic Tests:
- Benchmark 1 (Simple Anomaly): Successfully recovered the lateral position and polarity of a compact source but exhibited significant vertical smearing, a characteristic of minimum-structure inversions.
- Benchmark 2 (Dual Polarity): Correctly identified separated positive and negative bodies and the regional trend, though amplitudes were attenuated and vertical extents broadened.
- Benchmark 3 (Complex Structure): Demonstrated the limits of non-uniqueness. While the inversion achieved a WRMS consistent with the noise level, it failed to distinctly reconstruct the deeper positive intrusion at its true location, instead distributing density over a larger volume. This confirmed that a statistically acceptable data fit can coexist with substantial model-space error.
Field Application (Dieng):
- Both L2 and IRLS solutions satisfied the discrepancy criterion (WRMS $\approx$ 0.95–0.97) and identified the same first-order features: a strong positive-density zone beneath 2–3 km (associated with Mount Prau) and a secondary positive zone at 6–7 km, separated by broad negative domains.
- L2 Solution: Produced smoother, more distributed density contrasts with gradual transitions.
- IRLS Solution: Produced sharper vertical transitions and more columnar density structures due to the non-quadratic vertical-gradient term.
- Residuals: Both models showed localized misfit (standardized residuals > 4 $\sigma$ ) near the steepest gravity gradient (2.6–3.0 km), suggesting that the 2.5D parameterization or regional model may be insufficient for this specific high-amplitude anomaly.

Significance and Claims
The paper modestly positions ALTYN not as a tool for unique geological reconstruction, but as a transparent framework for constructing and comparing geophysically admissible density models.

Non-Uniqueness: The study reinforces that gravity inversion is intrinsically non-unique. The synthetic results explicitly show that satisfying the data discrepancy criterion does not validate the recovered geometry.
Stability vs. Detail: The author claims that first-order features (anomaly polarity and approximate lateral position) are more defensible and stable across different regularization schemes than detailed boundaries, density extrema, or deep continuations.
Prior Dependence: The results highlight that the recovered model is heavily dependent on prior assumptions, including regularization type, density bounds, strike length, and mesh resolution. The paper does not claim quantitative robustness to perturbations in these parameters, as dedicated perturbation experiments were not performed.
Geological Interpretation: For the Dieng application, the author concludes that the recovered models represent relative density contrasts consistent with the data and assumptions. While these contrasts are geologically plausible (e.g., positive anomalies potentially indicating intrusive bodies, negative anomalies indicating alteration or pyroclastics), the paper explicitly states that these interpretations remain hypotheses requiring independent constraints (e.g., borehole data, MT, seismic) for validation. The framework serves to make the assumptions underlying the inversion explicit rather than to produce a single "correct" geological cross-section.

More like this