PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning Volumetric Segmentation for Prediction of Cavity Geometry in Underground Jet Boring Mining

Giovanni Zingaro, Yuan Wang, Deepak Meena, Peter Cheng, Russell Buchanan, John Quilty, Robert Gracie

arXiv:2610.36303v1cs.CE

TL;DR

Cavity prediction in underground mining has largely relied on aggregate extraction outcomes rather than spatial geometry. This paper maps JBS data onto a voxel grid for 3D U-Net segmentation; the ensemble outperforms the cylindrical baseline on held-out test metrics, though predictions are weakest for underrepresented small virgin-ground cavities.

  • Problem

    Existing extraction studies commonly estimate aggregate outcomes rather than spatial geometry, in part because high-resolution geospatial training data are limited.

    Source: §1 · p. 2

  • Method

    The workflow projects cavity scans, geology, drilling, and jetting data onto a common voxel grid and trains a 3D U-Net with a backfill residual branch on historical cavities.

    Source: §1 · pp. 3–4

  • Results

    Test IoU 0.601±0.088 and F1-Score 0.747 ± 0.077; versus the baseline, IoU improves +22.4% and F1-Score +14.7%.

    Source: §5 · pp. 31–32

  • Takeaways & Limitations

    The workflow is intended to improve mine planning and support JBS operations.

    Source: §1 · p. 3

    Performance is weakest on small cavities excavated in virgin ground, which are underrepresented in the training data.

    Source: §4.2 · p. 27

Abstract

from arXiv · show

The Jet Boring System (JBS) is a proprietary technology developed by Cameco Corporation to extract high-grade uranium ore at the Cigar Lake Mine in Northern Saskatchewan, Canada. The JBS remotely excavates ore using a high-pressure water jet that cuts cavities within an artificially frozen ore body. This study outlines a Deep Learning (DL)-based workflow that formulates the cavity prediction problem as a volumetric binary semantic segmentation task. A preprocessing pipeline assembles drilling and jetting records, cavity scans, and a mine-wide geology block model onto a common voxel grid, where each voxel in the prediction is labelled as consumed or non-consumed during jetting. A three-dimensional U-Net with a backfill residual branch is trained on historical field cavities. The model is benchmarked against a baseline consisting of a cylindrical cavity centered along the pilot hole, with a diameter defined by domain experts. The proposed workflow achieves an IoU of 0.601 $\pm$ 0.088 and an F$_1$-Score of 0.747 $\pm$ 0.077 on a held-out test set, improving on the cylindrical baseline currently used in cavity planning by 22.4% and 14.7%, respectively. An eXplainable Artificial Intelligence (XAI) feature attribution approach is applied to identify the most influential input driving the predicted cavity geometry. The accuracy of the proposed model demonstrates its potential to support mine planning.

1. Introduction

Existing extraction models often predict aggregate outcomes rather than excavated geometry, while underground mining has not applied volumetric architectures to geometry prediction. This work formulates JBS cavity prediction as voxel-wise segmentation using geological and operational inputs.

1. Introduction

  • Mining data combine geological and operational measurements, but remote underground settings constrain collection through restricted access, poor visibility, and harsh conditions.These constraints are amplified in remote uranium mining, where inhospitable conditions limit data availability and accuracy.Source: §1 · p. 2
  • Existing extraction studies mostly predict aggregate measures, while volumetric architectures had not been applied to predict excavated geometry in underground mining.Spatial prediction is constrained by limited high-resolution geospatial data and the small number of cavities available at a mine.Source: §1 · pp. 2–3
  • The workflow assembles drilling and jetting records, cavity scans, and geology onto a common voxel grid, then classifies voxels as consumed or non-consumed.Its inputs include per-voxel geological and geotechnical properties, voxel-wise jetting and drilling parameters, distance to the pilot hole, and a backfill mask.Source: §1 · pp. 3–4
  • A three-dimensional U-Net with a backfill residual branch is trained on historical field cavities, and Integrated Gradients quantifies input-channel influence on predicted geometry.Final predictions are obtained from an ensemble of independently initialized models.Source: §1 · p. 4

2. The Jet Boring System and Available Data Streams

At Cigar Lake, JBS excavates frozen ore by jetting along a pilot hole, and cavity, geology, drilling, and jetting records provide the data streams for each cavity.

2.1. Excavation Process at Cigar Lake Mine

  • JBS accesses the ore body through a production tunnel after ground freezing stabilizes the water-saturated deposit.The jet extends from the tunnel through the unconformity and ore body to the cavity; retrieved slurry is collected for processing.Source: §2.1 · p. 4
  • Jetting proceeds downward from the top of the ore body as a rotating nozzle erodes a conical region along the pilot hole.The pilot hole is divided into constant-height slices between top and bottom jetting intervals; jetting time, rotation speed, and pump power form each slice’s recipe.Source: §2.1 · pp. 4–5

2.2. Data Streams

  • Each cavity has four data streams: a surveyed cavity point cloud, neighbouring-cavity surveys, mine-wide geological and geotechnical block data, and drilling and jetting records.Because the streams use independent coordinate systems, projecting them onto one voxel grid introduces spatial alignment uncertainty.Source: §2.2 · p. 5

3.1. Problem Formulation

The problem is to map voxelized geological, jetting, and drilling inputs to a binary voxel mask of the excavated cavity, learning the mapping from historical surveyed cavities.

3.1. Problem Formulation

  • The cavity is represented as a binary voxel mask, with each voxel indicating whether it was consumed during excavation.The forward prediction depends on voxelized geological properties, the jetting recipe, and the drilling configuration.Source: §3.1 · p. 6
  • The neural network’s forward function maps the finite-dimensional input channels to the predicted cavity mask.Inputs include per-voxel geological and geotechnical properties, voxel-wise jetting parameters, and drilling channels such as pilot-hole distance and the backfill mask.Source: §3.1 · p. 6
  • The model is trained on historical cavities whose input channels are paired with post-excavation surveyed masks, optimizing network weights through gradient-based backpropagation.The loss minimizes the difference between network predictions and observed voxelized cavities.Source: §3.1 · p. 6

3.2. Data Preprocessing

The preprocessing workflow projects Cameco’s data streams onto a common voxel grid for each cavity, producing model inputs and a binary cavity-mask target. It reconstructs cavity surfaces from scans, maps geological and operational information to voxels, and normalizes the channels.

3.2.1. Cavity Surface Reconstruction and Voxelization

  • The selected watertight mesh is ray-traced to label each voxel as inside or outside the reconstructed cavity.Candidate meshes are filtered for watertightness and ranked by mean distance between point-cloud nodes and mesh surfaces.Source: §3.2.1 · pp. 7–8

3.2.2. Geological, Geotechnical, and Backfill Channels

  • Five geological and geotechnical properties are mapped from a mine-wide lattice to cavity-centered sub-lattices, with prior neighboring-cavity backfill represented as a binary mask.The properties are Rock Mass Rating, Volumetric Water Content, Clay Content, U3O8 Percentage, and Arsenic Percentage; missing or invalid voxel values are set to zero.Source: §3.2.2 · pp. 8–9

3.2.3. Drilling and Jetting Channels

  • Jetting time, rotation speed, and pump power are assigned to voxels within per-slice conical masks, while a distance field encodes each voxel’s shortest distance to the pilot hole.The conical masks extend to the voxel-domain boundary rather than stopping at the planned cavity diameter, so the planned geometry is not explicitly encoded.Source: §3.2.3 · pp. 9–10

3.2.4. Data Normalization

  • Each channel is min-max scaled to [0, 1] using training-data minima and maxima, with logarithmic normalization first applied to quantities spanning several orders of magnitude.Incomplete voxel values are assigned zero.Source: §3.2.4 · p. 11

3.2.5. Summary of Model Channels

  • The preprocessing workflow produces 11 input channels and a single-channel binary cavity mask on the same nx × ny × nz Euclidean grid.Table 1 summarizes the channels, data sources, voxelization processes, and normalization schemes.Source: §3.2.5 · p. 11, Table 1 · p. 11

3.3. Deep Learning Model

The model maps voxelized geological, drilling, jetting, and backfill inputs to a binary cavity mask using a 3D U-Net. Training combines segmentation losses and ensemble learning, while evaluation uses imbalance-aware metrics and Integrated Gradients for feature attribution.

3.3.1. Network Architecture

  • The 3D U-Net reconstructs cavity geometry through encoder–decoder skip connections, while a separate backfill branch projects backfill information directly into the output.The branch is intended to preserve backfill location information that may be lost during the encoder–decoder’s downsampling and upsampling.Source: §3.3.1 · pp. 12–13

3.3.2. Training Design

  • The model is trained as an ensemble of independently initialized models whose predictions are averaged, with ensemble size set to 20.Training uses Adam for 225 epochs, and the authors report no further performance gains from increasing the ensemble size.Source: §3.3.2 · pp. 14–15
  • Training combines soft IoU and binary cross-entropy losses, with augmentation intended to improve generalization from limited cavity data.The authors note that spatially uncorrelated multiplicative noise cannot fully represent geological-channel measurement uncertainty.Source: §3.3.2 · pp. 14–15

3.3.3. Performance Evaluation

  • Evaluation emphasizes IoU and F1-Score for consumed voxels because cavity masks are highly imbalanced, and also reports precision, recall, and calibration measures.Predictions are thresholded at 0.5 for binary metrics; PR AUC and ROC AUC are computed from continuous outputs.Source: §3.3.3 · pp. 16–17

3.3.4. Feature Attribution via Integrated Gradients

  • Integrated Gradients assigns input-feature attribution by integrating output gradients along a path from a zero baseline to the actual input.The path integral is evaluated numerically using Gauss–Legendre quadrature.Source: §3.3.4 · p. 18

4. Results

The results section reports ensemble performance across dataset splits and includes representative test cavities. It also uses Integrated Gradients to quantify how input channels influence predicted cavity geometry.

4.1. Aggregated Performance

  • Performance is assessed primarily with IoU, F1-Score, precision, recall, and PR AUC, with calibration assessed using ECE.Reported metrics are for the ensemble model.Source: §4 · p. 18

4.2. Representative Test Cavities

  • Representative test cavities are presented to illustrate the model’s predictive accuracy.The section overview identifies these as test-set examples.Source: §4 · p. 18

4.3. Feature Attribution

  • Integrated Gradients is used to quantify each input channel’s spatial influence on predicted cavity geometry.Source: §4 · p. 18

4.1. Model Performance

The ensemble is evaluated against a cylindrical baseline using overlap, voxel-wise, tonnage, cavity-type, calibration, and segmentation-benchmark analyses. Its test overlap scores exceed the baseline, while performance varies by cavity type and calibration metrics have class-imbalance and spatial-correlation caveats.

4.1.1. Baseline Model

  • The cylindrical baseline achieved IoU = 0.491 ± 0.111, F1-Score = 0.651 ± 0.108, and mean absolute U3O8 tonnage error of 38.0%.It assumes a domain-expert-defined diameter and spans the pilot hole's top and bottom jetting intervals.Source: §4.1.1 · p. 19, Table 2 · p. 19

4.1.2. Overlap and Voxel-Wise Metrics

  • PR AUC was 0.842 against a random-classifier baseline of 0.063, and test IoU varied more across cavities, partly because small cavities magnify relative overlap error.Pooled voxel-wise MCC was 0.737; precision exceeded recall (0.783 vs. 0.723).Source: §4.1.2 · pp. 19–20
  • Test IoU of 0.601 ± 0.088 and F1-Score of 0.747 ± 0.077 improved by 22.4% and 14.7%, respectively, over the baseline.Table 3 summarizes ensemble performance across the training, validation, and test datasets.Source: §4.1.2 · p. 19, Table 3 · p. 20

4.1.3. Tonnage Error and Performance by Cavity Type

  • Per-cavity mean signed U3O8 tonnage error was 11.2% (mean absolute 25.6%), while pooled signed tonnage error was −7.7%.The per-cavity mean absolute error was 25.6%, compared with the 38.0% baseline error.Source: §4.1.3 · pp. 20–21
  • Backfilled cavities scored IoU 0.610 ± 0.076 and mean absolute tonnage error 21.7%; on virgin cavities, the baseline outperformed the proposed model.For virgin cavities, the proposed model achieved IoU 0.525 ± 0.160 and mean absolute tonnage error 62.0%, versus baseline values of 0.592 ± 0.095 and 33.0%.Source: §4.1.3 · p. 21, Table 4 · p. 22

4.1.4. Calibration

  • ECE and Brier score were 0.020 and 0.023; reliability analysis found underconfidence at predicted probabilities ≤0.42 and overconfidence above 0.5.Standard ECE is dominated by voxels with predicted probability below 0.1, and the voxel-wise calibration metrics do not account for spatial correlation.Source: §4.1.4 · pp. 22–23, Figure 9 · p. 23

4.1.5. Comparison with Segmentation Benchmarks

  • The test IoU of 0.601 falls within the range considered acceptable for comparable segmentation problems, despite geological, data, alignment, and three-dimensional-task uncertainty.The comparison passage describes IoU values of 0.6 to 0.7 as more commonly reported under challenging segmentation conditions.Source: §4.1.5 · p. 24

4.2. Representative Test Cases

The representative test cavities span poor to high predictive performance, with the weakest case being a small virgin-ground cavity. The examples also show that overlap and tonnage outcomes can differ.

Representative test cases

  • 0.601 ± 0.088 mean IoU summarizes test-split performance, with per-cavity IoU ranging from 0.223 to 0.747.The three representative cases span poor, average, and high-accuracy predictions; Cases 1 and 2 have backfilled neighbours, while Case 3 is in virgin ground.Source: §4.2 · p. 24

Test Case 1: Average Prediction

  • 0.609 IoU and 0.757 F1-Score characterize Test Case 1, where false positives slightly exceed false negatives.The U3O8 tonnage error is −2.5%; volumetric over-segmentation can coexist with under-predicted ore recovery.Source: §4.2 · p. 25, Figure 13 · p. 26

Test Case 2: High-Accuracy Prediction

  • 0.747 IoU and 0.855 F1-Score make Test Case 2 the highest-IoU example, with −0.4% U3O8 tonnage error.Its predicted boundary is consistent with the backfill interface.Source: §4.2 · p. 26, Figure 14 · p. 26

Test Case 3: Poor Prediction

  • 0.223 IoU, 0.224 precision, and 0.984 recall characterize Test Case 3, whose prediction over-segments the small virgin-ground cavity.The predicted volume is 11,994 voxels versus 2,732 ground-truth voxels, and U3O8 tonnage error is +236.3%; the case is out of distribution and virgin-ground cavities are underrepresented.Source: §4.2 · p. 27, Figure 15 · p. 27

4.3. Feature Attribution Scores

Ensembled Integrated Gradients maps attribute predictions to spatial and operational inputs relative to a zero baseline. The pilot-hole distance field has the largest attributions, while other channels show distinct spatial patterns.

Attribution method

  • Integrated Gradients scores are averaged across independently initialized models, and slices show signed attribution relative to a zero baseline.Positive attribution corresponds to increased predicted output; negative attribution corresponds to decreased output.Source: §4.3 · p. 28

Pilot-hole distance field

  • The pilot-hole distance field has the largest attributions, with a negative band around the predicted wall and weakly positive attribution inside the cavity.Increasing distance from the jet decreases predicted consumption.Source: §4.3 · pp. 28–29, Figure 17 · p. 29

Operational recipe channels

  • Pump power, jetting time, and rotation speed share an attribution pattern, with positive values extending radially through the conical jetting region.Negative attribution inside the cavity is associated with saturation; positive attribution beyond the predicted boundary suggests influence on radial erosion extent.Source: §4.3 · p. 29, Figure 18 · p. 29

Nozzle angle

  • Nozzle angle has high aggregated attribution but no spatial correlation with cavity geometry, consistent with its uniform broadcast across the domain.The authors suggest it influences the conical extent of erosion at each slice.Source: §4.3 · pp. 29–30

Backfill

  • Backfill attribution is localized to the backfilled region, with negative values at the interface where backfill reduces predicted output.The passage relates this pattern to backfill's lower erodibility.Source: §4.3 · p. 30, Figure 19 · p. 30

Geological channels

  • U3O8 attribution is mostly positive, strongest around the predicted boundary, while VWC and clay attribution is diffuse and arsenic influence is negligible.RMR attribution transitions from positive to negative across the predicted cavity, which the authors associate with differences in rock erodibility.Source: §4.3 · pp. 30–31, Figure 20 · p. 30, Figure 21 · p. 31

5. Conclusions

The study presents an ensemble 3D U-Net workflow for voxel-level cavity prediction and reports improvements over the cylindrical baseline. Its weakest performance is on underrepresented small cavities in virgin ground, and the work evaluates predictive performance only.

Workflow

  • A 3D U-Net predicts cavity geometry as binary voxel segmentation from geology, geotechnical, and operational data projected onto a common grid.The workflow uses an ensemble of 20 independently initialized models.Source: §5 · p. 31

Results

  • The ensemble achieves test IoU of 0.601 ± 0.088 and F1-Score of 0.747 ± 0.077, improving over the baseline by 22.4% in IoU and 14.7% in F1-Score.Mean absolute U3O8 tonnage error decreases from 38.0% to 25.6% per cavity.Source: §5 · pp. 31–32

Scope and conclusion

  • Performance is weakest on small virgin-ground cavities, which are underrepresented in the training data; the work evaluates predictive model performance only.The authors describe the model as a promising tool to improve cavity planning, without evaluating production or economic outcomes.Source: §5 · p. 32

Declaration of Competing Interest

The declaration reports Cameco Corporation financial support for Robert Gracie and John Quilty, and a Cameco employment relationship for Peter Cheng.

Financial support

Employment relationship

Other authors

Related papers

Loading 2610.36303v1…