Source-linked AI summary
Bayesian deep learning integration of geophysical and drilling data for 3D prediction of copper mineralization and drill targeting: a case study from the Kogodai prospect, Rudny Altai
Margarita Veshchezerova, Egor Barashov, Evgenii Gusev, Michael R. Perelshtein, Arlan Kasymzhan, Bolat M. Kabaziev, Nurlan Y. Askarov
TL;DR
Sparse sampling, heterogeneous datasets, and ambiguous geophysical inversions complicate drill targeting in structurally complex brownfield terranes. The study develops a Bayesian deep-learning workflow that learns continuous 3D Cu-grade and geophysical fields with epistemic uncertainty from integrated exploration evidence. At Kogodai, predictions recover known mineralized patterns, identify IP-coincident candidate zones, and support uncertainty-aware drill prioritization rather than resource estimation.
Problem
Brownfield drill targeting must integrate sparse, heterogeneous exploration evidence while managing uncertainty in costly follow-up drilling.
Method
A Bayesian multi-head deep-learning model jointly learns continuous 3D Cu-grade, chargeability, and apparent-resistivity fields while estimating epistemic uncertainty from sampled neural-network weights.
Results
The model delineates a principal mineralized trend and localized zones coinciding with elevated IP responses, with Cu predictions separating ore-bearing from ore-free sampled blocks at an AUC of 0.93.
Takeaways & Limitations
Continuous Cu predictions can be thresholded into prospectivity maps, while uncertainty mapping supports robust-target identification and drill selection for improved confidence.
Takeaways & Limitations
Incomplete geophysical provenance metadata and heterogeneous historical sampling reduce confidence in detailed anomaly geometry and constrain interpretation of legacy-driven predictions.
Abstract
from arXiv · showhide
Exploration drill targeting in structurally complex terranes is hindered by sparse sampling, heterogeneous datasets, and the ambiguity of geophysical inversions. Here, we present an uncertainty-aware 3D workflow for the acceleration of time-to-discovery in brownfield explorations and apply it to the Kogodai prospect in the Rudny Altai metallogenic province. We jointly analyse existing drilling and geophysical data in a comprehensive approach, revealing hidden patterns in already available data. Drillholes and trenches were desurveyed to a common 3D reference frame, and assays were composited to a consistent spatial support to facilitate joint modelling with geophysical inputs. We develop Bayesian deep-learning models to predict 3D fields of Cu grade together with chargeability and apparent resistivity while quantifying epistemic uncertainty via Monte Carlo sampling. The original contribution of this work is to treat the problem not as pointwise regression between co-located observations, but as joint learning of spatially continuous 3D fields from sparse, heterogeneous exploration evidence. The resulting 3D predictions delineate a principal mineralized trend and several localized candidate zones that coincide with elevated induced polarization (IP) responses, while uncertainty mapping highlights where predictions are robust versus where additional drilling would be most informative. The continuous Cu-grade field can also be thresholded to produce binary prospectivity maps, allowing the sensitivity of target delineation to the chosen cutoff to be evaluated. The outputs are intended for qualitative interpretation and risk-aware drill targeting rather than resource estimation, and we discuss key limitations arising from incomplete provenance metadata for geophysical products and heterogeneity of historical sampling.
1. Introduction
The study addresses brownfield drill targeting by integrating sparse drilling, trench, geological, and geophysical evidence into uncertainty-aware 3D models for Kogodai. It positions the workflow between prospectivity mapping and resource estimation while retaining a targeting-focused scope.
- Motivation: Brownfield targeting must integrate sparse drilling, historical maps, trenches, and indirect geophysical measurements to guide costly follow-up drilling.The Kogodai dataset includes 45 drillholes, trenches, and IP and resistivity profiles.
- Motivation: AI-based workflows can provide an independent second opinion by validating known trends, highlighting overlooked anomalies, and ranking targets in heterogeneous legacy datasets.
- Study setting: Kogodai is a VMS-related copper occurrence in eastern Kazakhstan, where mineralization is localized near amphibolite–gneiss or terrigenous-rock contacts and controlled by structural-tectonic factors.
- Contribution: The study combines elements of mineral prospectivity mapping and resource estimation but predicts continuous 3D copper-grade and geophysical fields for drill planning rather than resource-grade or grade-tonnage estimation.
- Contribution: Bayesian weight sampling estimates epistemic uncertainty from limited and uneven observations, separating robust targets from weakly constrained anomalies.
- Contribution: The workflow supports two targeting uses: prioritizing high-probability copper targets to accelerate discovery or selecting drillholes expected to improve subsurface confidence.
2. Geological background
Kogodai is a VMS-related Cu prospect in the structurally complex Rudny Altai province, hosted by high-grade metamorphic rocks and associated with folds, contacts, and deformation. Its mineralization includes layered pyrite-rich and superimposed Zn–Cu styles, with underexplored flanks motivating integrated targeting.
- Regional and structural setting: Kogodai lies in the Rudny Altai VMS province within the high-grade metamorphic Kurchum block and the Irtysh shear zone.
- Regional and structural setting: The prospect is associated with the gently northwest-plunging closure of the Kogodai syncline, while minor longitudinal and transverse faults occur without dominant large offsets.
- Mineralization controls: Mineralized zones are broadly lens- to ribbon-shaped, conformable with host-rock fabric, and preferentially localized along amphibolite–gneiss contacts, schistosity, deformation, and brittle structures.
- Exploration context: The southern altered-mineralized zone extends more than 500 m along strike, whereas the flanks remain comparatively underexplored and motivate geophysical integration.
- Mineralization types and weathering: The dominant mineralization is layered pyrite-rich sulfur–pyrite mineralization, locally overprinted by economically more important disseminated or veinlet Zn–Cu mineralization.
- Mineralization types and weathering: Near-surface oxidation transforms sulfides into limonite–quartz assemblages with malachite and leached cavities, typically extending about 15 m and locally deeper along disturbed structures.
3. Methods
The workflow jointly models sparse drillhole and trench assays with indirect IP and resistivity observations as continuous 3D fields. A Bayesian multi-task neural network produces Cu, chargeability, and resistivity predictions together with epistemic uncertainty for interpretation and drill targeting.
- Joint modelling: Non-co-located drillhole, trench, IP, and resistivity observations are formulated as joint learning of spatially continuous 3D fields rather than pointwise regression.The datasets differ in spatial support: assays follow 1D sampling paths, whereas geophysical products represent 2D sections or inversion-derived observations in 3D.
- Data preparation: Before modelling, datasets are transformed to a common 3D coordinate frame, assays are composited to consistent downhole support, and variables are normalized for joint training.Geophysical profile or inversion products are represented as spatial observations with associated 3D coordinates.
- Bayesian uncertainty: Monte Carlo sampling of Bayesian weight configurations yields predictive means and variances, with variance interpreted as epistemic uncertainty from limited knowledge.The model treats parameters as distributions rather than deterministic values, producing multiple plausible field realizations.
- Architecture: A shared Bayesian neural-network trunk learns a latent 3D representation, while task-specific heads predict copper grade, chargeability, and apparent resistivity.The multi-head structure supports heterogeneous label availability and uses shared information among the three predicted fields.
- Architecture: The gated Modified MLP computes input-dependent modulation vectors and applies them through the shared trunk, while SiLU activations are used in hidden layers and tanh bounds head outputs.The gating is described as an adaptive interpolation between learned representations at each layer and is intended to improve gradient flow in coordinate-based networks.
- Training: A masked multi-task loss lets each target contribute only where its observations are available, avoiding a strict co-location requirement.Copper assays, chargeability, and apparent resistivity can therefore be trained jointly despite incomplete and uneven label coverage.
4. Induced Polarization and Resistivity data
IP chargeability and apparent resistivity were prepared as spatially continuous constraints for 3D targeting, using quality control, inversion, and heterogeneous legacy products. The resulting geophysical evidence is model-dependent rather than uniquely geological.
- IP chargeability and apparent resistivity were treated as primary constraints for targeting beyond existing drilling and trenching.Their use is motivated by associations between electrical responses, sulfides, alteration, fluid pathways, and lithological contrasts.
- Ground surveys used pole–dipole electrical tomography with 25 m electrode spacing and 200–400 m profile spacing.
- Quality control retained average off-time chargeability because it is less sensitive to transient noise than individual time-window measurements.
- ZondRes2D performed 2.5D resistivity and IP inversion to approximately 350 m depth using regularized iterative optimization and an Occam algorithm.The inversion assumes lateral invariance perpendicular to each profile and produces comparatively smooth solutions.
- Because legacy products had incomplete provenance and heterogeneous anomaly expression, modelling tested horizon-wise slices and a 3D chargeability voxel representation after excluding non-physical values.
5. Drilling and trench sampling data
Kogodai drilling and trenching provided sparse, selectively sampled copper assays that were reconstructed in 3D and composited to match the coarser geophysical support. These preprocessing choices enable joint modelling but constrain how the assay data represent mineralization.
- Recent exploration contributed 4,949 m of drilling across 45 drillholes and 1,207 m of trenches across 35 trenches, supplemented by legacy records.
- Sampling targeted visible sulfide intervals, so unsampled core is not missing at random and commonly represents intervals without recorded mineralization.Sample lengths ranged from 0.3 to 2.0 m, averaging approximately 1.0 m.
- Only 1,876 of 7,136 recorded core metres were assayed, representing 26.3% sampled core; 37 of 44 drillholes contain assays.
- Collars, downhole surveys, trench coordinates, and assay intervals were compiled and drillhole trajectories reconstructed through a 3D desurvey workflow.
- Assays were composited into 10 m downhole intervals to harmonize their support with geophysical observations of approximately 15 m spatial sampling.The compositing was intended for machine-learning integration rather than resource estimation.
- Copper-grade distributions were examined on both original and log-transformed scales before modelling.
6. Predicted 3D mineralization model
The Bayesian Modified MLP predicts a continuous 3D Cu field with uncertainty from integrated assay and geophysical evidence. Its outputs recover the main mineralized pattern, identify localized IP-associated candidate zones, and support threshold-based prospectivity assessment for follow-up drilling.
- The model delineates a principal southern mineralized trend and smaller features coinciding spatially with elevated IP responses, especially in inverted chargeability products.Figure 10 highlights blocks exceeding the 0.15% Cu visualization threshold.
- Candidate features occur preferentially where chargeability anomalies are clearest, while uncertainty increases in sparsely sampled or weakly constrained volume parts.The localized features are interpreted as targets for follow-up validation.
- The predictions recover the principal known mineralized zones and provide a geological comparison using five reference zones separated with a 0.15% Cu cutoff.
- No prediction was made outside the geophysical voxel-data coverage.
- The continuous Cu field can be thresholded into binary prospectivity maps, allowing target delineation sensitivity to the chosen cutoff to be evaluated.
- A 0.08% predicted-Cu threshold gives AUC = 0.93, approximately 0.75 true positive rate, and approximately 0.09 false positive rate for sampled-block classification.Threshold curves retain many ore-bearing blocks while rejecting most ore-free blocks near the reference threshold.
- After confidence prefiltering, 84.1% of ore-bearing blocks and 32.5% of confident positive blocks fall within the 0.08% predicted-Cu threshold.The prefilter retained 15,674 blocks, or 3.2% of all modelled blocks.
7. Discussion
The workflow integrates drilling, trench assays, and heterogeneous IP/resistivity products into a Bayesian 3D decision-support model for ranking brownfield drilling hypotheses. Its uncertainty estimates distinguish robust targets from areas where additional data would be most informative, while legacy-data limitations constrain interpretation.
- 7.1. Model outputs as a decision-support layer for drill targeting: The workflow combines copper assays with IP chargeability and apparent resistivity in a single 3D predictive model for prioritizing follow-up drilling.It is intended as a decision-support layer rather than a replacement for geological interpretation or a formal resource estimate.
- 7.1. Model outputs as a decision-support layer for drill targeting: Epistemic uncertainty separates targets supported by multiple constraints from poorly sampled or single-source anomalies where additional drilling would be especially informative.Higher uncertainty does not make a target uninteresting; it identifies where new data could reduce model uncertainty.
- 7.3. Geological meaning of the predictions: All recognized mineralized zones within geophysical coverage were identified, and additional targets including Zones 2 and 5 were recommended for verification.Subsequent geological review found these additional areas prospective, but the outputs remain target-ranking and hypothesis-generation results rather than resource-continuity claims.
- 7.4. Limitations: Historical maps were excluded as quantitative inputs because incomplete georeferencing, metadata, and spatial consistency could underrepresent lithological and structural controls.They were retained only as qualitative geological context.
- 7.4. Limitations: Incomplete geophysical provenance reduces confidence in detailed anomaly geometry, while selective assays and inconsistent core logging bias or limit the geological information available for training.The model therefore relies primarily on relative geophysical contrasts and spatial patterns rather than absolute values from one product.
- 7.5. Future work: Future improvements center on additional IP/resistivity coverage, standardized provenance-controlled data management, and more consistent validated core logging.These changes are presented as priorities for strengthening target confidence and reproducibility.
8. Conclusions
The study presents a Bayesian deep-learning workflow that learns continuous 3D copper and geophysical fields from sparse, heterogeneous exploration evidence while estimating epistemic uncertainty. At Kogodai, it reconstructs known mineralized zones, identifies IP-coincident candidate areas, and supports threshold-sensitive, risk-aware target prioritization within clear legacy-data limitations.
- 8. Conclusions: The workflow learns mutually constrained, spatially continuous 3D fields from incomplete drilling, trench, and geophysical evidence rather than treating observations as simple co-located regression.It integrates IP/resistivity data and assays within a Bayesian deep-learning framework for drill targeting.
- 8. Conclusions: Cu prediction achieved an AUC of 0.93 in separating ore-bearing from ore-free sampled blocks, with greatest confidence around more densely sampled prospect areas.Threshold-sensitivity and ROC analyses support the reported predictive performance.
- 8. Conclusions: The 3D model delineated a principal mineralized trend and localized candidate zones coinciding with elevated IP responses, while uncertainty mapping indicated where additional data were required.Continuous Cu predictions can also be thresholded into binary prospectivity maps to examine sensitivity to the selected cutoff.
- 8. Conclusions: The workflow combines target generation, geological validation, and risk-aware prioritization, but its interpretation is limited by incomplete geophysical metadata and inconsistent historical core-logging conventions.The authors frame the approach as a contribution to drilling planning rather than resource estimation.
CRediT authorship contribution statement
The authors’ contributions span conceptualization, methodology, software, analysis, investigation, visualization, supervision, data curation, and writing.
- CRediT authorship contribution statement: Margarita Veshchezerova led conceptualization, methodology, software, formal analysis, and original-draft writing.
- CRediT authorship contribution statement: Egor Barashov contributed methodology, software, investigation, visualization, and original-draft writing.
- CRediT authorship contribution statement: Evgenii Gusev contributed software, validation, and data curation, while Michael R. Perelshtein contributed supervision, project administration, and review and editing.
- CRediT authorship contribution statement: Arlan Kasymzhan contributed resources, investigation, and data curation, while Bolat M. Kabaziev contributed resources and supervision.
- CRediT authorship contribution statement: Nurlan Y. Askarov contributed resources and data curation.