Maps of known mineral deposits are a lot like broken clay tablets. Both are records with holes in them, and in both cases the holes are not random. A tablet breaks along its edges and corners. The mineral record goes silent wherever rock is buried under sediment, because nobody saw the outcrop and nobody drilled. Recent work trains neural networks to "fill in" missing deposits on a continental map by hiding parts of the map and asking the model to reconstruct them. It hides random tiles, so it never tests the case that matters, which is predicting deposits that lie under cover.
We propose to treat burial as the damage process. Training and evaluation masks will follow where geological cover actually lies, not random squares. To stop the model from inventing deposits, we add a constraint drawn from fractal geometry. In well-exposed terrain, known deposits cluster in a way that looks statistically similar across scales. That multi-scale clustering can be measured by counting occupied grid boxes at several box sizes. Predictions under cover should reproduce the same counting curve. The claim is simple: an infilling model that is trained with damage-shaped masks and held to the clustering law of exposed terrain will recover concealed deposits better, and hallucinate fewer, than one trained on random masks or one handed fractal statistics as extra inputs.
Core Claim and Where It Sits
The closest existing work is Masked Mineral Modeling, or M3. It frames prospecting across the conterminous United States as masked image infilling, in the spirit of masked autoencoders. A ResNet U-Net reconstructs ten metal layers from the USGS Mineral Resource Data System (MRDS), guided by geophysical, lithological and soil-spectral layers. It reports a test Dice of 0.31 ± 0.01 and recall of 0.22 ± 0.02. M3 motivates this proposal in two ways. It describes MRDS as "an incomplete collection of true positives" biased toward roads and valleys, and it reports that its infills hallucinate. M3 also constrains our design. Kriging failed to generalize to unseen tiles, and Vision Transformers did clearly worse than the convolutional network. Neither result came with a number. We therefore keep M3's U-Net backbone and change only what is masked and what the loss rewards.
Fractal analysis supplies the constraint. Cui et al.'s Luanchuan study showed with 2D box-counting that ore-controlling NW faults reach a capacity dimension near 1.65, while barren NE faults reach 1.21. The Southern Jiangxi tungsten study overlaid fault fractal dimensions on Weights-of-Evidence maps; its top prospectivity tier contained 58.47% of known occurrences. Both studies use fractal dimension as a hand-set threshold on a map. Neither uses it as a training signal, and neither deals with cover. Liu et al. add a caution that we adopt directly. Box-counting is corrupted by artifacts such as roads mistaken for faults. Self-similarity also holds only within a limited range of scales, so any constraint must be restricted to a measured scaling window.
Historical-record restoration supplies the evaluation discipline and a warning. Ancient-text infilling, such as the bidirectional LSTM for Late Babylonian texts (85% Hit@1), succeeds partly because grammar tightly constrains what can fill a gap. Spatial infilling has no grammar. Critical Confabulation names ungrounded filling of archival silences as the central failure mode. Our position is that a measured scaling law is the closest thing geology offers to a grammar, and that the gaps must be the shape of real damage before any success means anything.
The input direction on deposit depth enters as the application regime, and we are honest about its reach. The deep three-dimensional datasets in this literature are site-specific and largely confidential; Zhaoxian's drill logs are explicitly withheld (Li et al., 2022). So "depth" here means burial beneath mapped cover, not a predicted z-coordinate. Prediction of a true depth coordinate is out of scope for this note.
Research Program
We build three things on the M3 data stack, which consists of MRDS metal layers, USGS elevation, faults, slip rates, rock type and age, and RaCA soil spectra, on the same 50-mile tiles at one-square-mile resolution.
Cover masks. From the lithology and age layers already in the stack, we label pixels mapped as young sedimentary or surficial units as "covered." We then generate training masks in two ways. The first reuses real cover footprints by transplanting them onto exposed tiles. The second applies synthetic burial to exposed pixels: we hide the deposit label and also degrade the evidence a buried target would lose. Concretely, fault presence is removed, and soil spectra are replaced with spectra drawn from covered pixels elsewhere. This mirrors realistic damage simulation in text restoration, where the shape of the gap matters as much as its size.
Clustering reference. On tiles with little cover, we compute the box-occupancy curve of known deposits per metal. This is $N_{\text{ref}}(\epsilon)$, the number of boxes of side $\epsilon$ that contain at least one deposit. We fit it only inside the range of $\epsilon$ where $\log N$ is linear in $\log(1/\epsilon)$, and we report that window explicitly. Fault traces are screened against road and rail layers before any fault fractal statistic is computed.
Fractal-matching loss. We add a scale-matching penalty to M3's reconstruction loss. Let $p$ be the model's per-pixel deposit probability on a tile. Let $\tilde N_p(\epsilon)$ be a soft box count, computed as the sum over boxes of side $\epsilon$ of the maximum of $p$ within each box, which max-pooling makes differentiable. Let $\mathcal{E}$ be the set of box sizes inside the measured scaling window. The added term is
$$ \mathcal{L}_{\text{frac}} = \sum_{\epsilon \in \mathcal{E}} \big(\log \tilde N_p(\epsilon) - \log N_{\text{ref}}(\epsilon \mid \text{area})\big)^2 , $$
where $N_{\text{ref}}$ is rescaled to the masked area. The penalty punishes predictions that are too diffuse, such as smearing probability across the cover. It also punishes predictions that are too sparse, such as predicting nothing under cover. It says nothing about where deposits go; that remains the job of the evidence layers.
Baselines
The naive reference point has two parts. The first is a prior that assigns each covered pixel the regional occurrence rate of its lithology class. The second is "predict no deposits under cover," which is effectively what an unbiased reading of MRDS does. Any method must beat both on recall at matched area.
The strongest existing method is M3 itself, a ResNet U-Net trained with random tile masks, whose reported result is Dice 0.31 ± 0.01 and recall 0.22 ± 0.02 on random test tiles (Nair et al., 2026). It is the strongest competitor because it is the only infilling model in this literature on this data, and it already uses the same evidence layers. We will retrain it from scratch on our splits with identical architecture, inputs, epochs and hyperparameter search budget as our variants. We will report it on its own random split, to reproduce the 0.31, and on our cover split.
We add a Random Forest on per-pixel evidence features. This is the method that beat logistic regression in Meng et al.'s 3D skarn study, where it reached AUC 0.989 and captured ore in the top 4% of the volume, compared with 20% for logistic regression. It is the classical competitor a geologist would actually use.
Two ablations isolate the key ingredients. The first is cover masks without the fractal loss, which isolates the masking. The second is M3 plus local fault box-counting dimension maps as extra input channels, in the Cui and Jiangxi style, with no loss term. This isolates "fractal geometry as a constraint" from "fractal geometry as a feature." A control trains the same loss against a reference curve taken from a mismatched metal. If that control helps equally, the benefit is generic smoothing and not the clustering law.
Beating M3 means higher recall on covered held-out pixels at the same fraction of area flagged, with a confidence interval over five seeds that excludes M3's.
Why not just…?
Why not just feed fractal dimension in as an input feature? This is the most plausible competitor, and it is cheap. The Luanchuan and Jiangxi results show fault fractal dimension separates fertile from barren structure. But a feature describes the evidence around a pixel; it does not stop the model from distributing too many or too few deposits under cover, which is the hallucination M3 reports. The feature-input ablation tests this directly. If it matches the loss on covered recall and on occupancy-curve error, we will conclude that the constraint adds nothing and report the cheaper method.
Why not just treat this as positive-unlabeled learning and reweight the unlabeled pixels? Reweighting addresses the fact that unlabeled pixels are not true negatives, which M3 and the Zaozigou study both handle crudely. It is a real threat. However, standard reweighting assumes labels go missing at a uniform rate, which is exactly what cover violates. We will include a reweighted variant of M3 with label frequency estimated per lithology class. If it closes the gap, the contribution narrows to the evaluation protocol alone.
Why not just krige or use Weights-of-Evidence? M3 reports that kriging fit training areas but failed on unseen tiles, though no exact number was given. Weights-of-Evidence assumes conditional independence that fault-derived layers violate. We include Weights-of-Evidence with Jiangxi-style fractal overlays for completeness, but we do not expect it to compete.
Why not just go to true 3D depth prediction? That is the scientifically sharper target. The 3D benchmarks here, however, are single deposits with confidential or unreleased data: Zaozigou with 72 drill holes, Dayingezhuang with more than 16,000 assays (Deng et al., 2021). A public, reproducible test of the burial-as-damage idea has to come first.
Proposed Experiments
The first experiment is the decisive check on the evaluation protocol. Retrained M3 is scored on random held-out tiles and on held-out covered pixels from geographically separate tiles. If covered-pixel recall falls well below the random-tile recall of 0.22, random masking overstates usefulness for concealed targets, and that alone is a reportable finding.
The second experiment compares all baselines and ablations on the cover split. Metrics are Dice, recall at fixed flagged area, area under the success-rate curve, and occupancy-curve error. Occupancy-curve error is computed against held-out exposed tiles, which the model never trained on.
The third experiment is synthetic burial. On exposed tiles we progressively degrade evidence layers, from no degradation to all layers degraded, and plot recall against degradation level. The question is whether the fractal loss flattens this curve. A flatter curve would be the map-scale analogue of the steep loss of accuracy with depth that the Zaozigou study reports.
The fourth experiment is a falsification check through the mismatched-reference control and a per-metal analysis. If benefits appear only for metals whose exposed-terrain occupancy curves show a clean linear window, the mechanism is the scaling law. If they appear uniformly, it is regularization in disguise.
No reported result was found on Lacuna for kriging, for Vision Transformers on the MRDS benchmark, or for any model evaluated on covered terrain specifically. Those numbers will be established here.
References
- Masked Mineral Modeling: Continent-Scale Mineral Prospecting via Geospatial Infilling
- Fractal analysis of structural characteristics and prospecting of the Luanchuan polymetallic mining district, China
- Joint Application of Fractal Analysis and Weights-of-Evidence Method for Revealing the Geological Controls on Regional-Scale Tungsten Mineralization in Southern Jiangxi Province, China
- Fractal and Multifractal Characteristics of Lineaments in the Qianhe Graben and Its Tectonic Significance Using Remote Sensing Images
- Restoration of fragmentary Babylonian texts using recurrent neural networks
- Critical Confabulation: Can LLMs Hallucinate for Social Good?
- Combining 3D Geological Modeling and 3D Spectral Modeling for Deep Mineral Exploration in the Zhaoxian Gold Deposit, Shandong Province, China
- Three-Dimensional Mineral Prospectivity Modeling for Delineation of Deep-Seated Skarn-Type Mineralization in Xuancheng–Magushan Area, China
- 3D Mineral Prospectivity Mapping of Zaozigou Gold Deposit, West Qinling, China: Deep Learning-Based Mineral Prediction
- Learning 3D Mineral Prospectivity from 3D Geological Models with Convolutional Neural Networks: Application to a Structure-controlled Hydrothermal Gold Deposit