3D ResUNet and PU Bagging Enhance Mineral Targeting in China

Artificial intelligence has quietly reshaped how geologists hunt for buried copper, and a new study from China’s Duobaoshan ore district shows just how far the technology has come. Researchers Xiumei Lv and Gongwen Wang of the China University of Geosciences in Beijing have developed a coupled deep-learning framework that merges generative adversarial networks with weakly supervised learning to map where copper and molybdenum deposits are most likely to hide deep underground. The work, published in the journal Natural Resources Research, tackles one of the most stubborn problems in modern mineral exploration: how to train a neural network to find ore when almost all of the training data is either missing or unreliable.

Duobaoshan, located in Heilongjiang Province in northeastern China, sits within the northeastern segment of the Central Asian Xing’an–Mongolia orogenic belt. The district is far from virgin territory. It hosts the Duobaoshan super-large porphyry Cu–Mo deposit and the Tongshan large-scale porphyry deposit, two of the most significant copper systems in the region, and geological evidence points to substantial metallogenic potential remaining at depth. But as exploration pushes deeper, the task becomes dramatically harder. The geological structures are complex, ore bodies are increasingly concealed, and the drilling data that once guided discovery grows sparse. Traditional two-dimensional prospectivity maps, which summarize evidence on the surface or at fixed depth slices, simply cannot capture the three-dimensional architecture of a porphyry system.

The answer, many exploration geologists have concluded, lies in three-dimensional mineral prospectivity mapping, or 3D MPM. Rather than treating the subsurface as a stack of flat maps, 3D MPM divides the rock volume beneath a district into thousands or millions of cubic voxels, each carrying a suite of evidence: lithology from 3D geological models, density and magnetic susceptibility from inverted gravity and magnetic data, and known mineralization from drill holes. A machine-learning model then assigns each voxel a probability of hosting ore. The approach integrates depth information and spatial continuity far more comprehensively than flat maps, and it can delineate targets that no surface anomaly would ever reveal.

Yet 3D MPM has a fundamental weakness that has limited its practical success. Supervised deep-learning models need large, well-labeled training sets, and ore deposits provide neither. Positive samples, meaning voxels known to contain mineralization, are scarce because only a handful of drill holes intersect ore. Negative samples, meaning voxels confirmed to be barren, do not really exist at all; geologists cannot prove that any given block of rock contains no mineral. Researchers have traditionally sidestepped the problem by randomly selecting unlabeled voxels and treating them as negatives, but this introduces label noise: some randomly chosen “negatives” may actually sit within or near undiscovered ore. The resulting models suffer from unstable training and poor generalization, and their predictions can shift unpredictably from one run to the next.

Lv and Wang’s framework attacks the problem from two directions simultaneously, in what the authors describe as a “generative-weakly supervised” coupled approach. The first component addresses the shortage of positive samples. The researchers employed a Wasserstein generative adversarial network with gradient penalty, known as WGAN-GP, to learn the latent distribution of the limited real positive samples and then generate synthetic new ones. Generative adversarial networks pit two neural networks against each other: a generator that fabricates samples and a discriminator, or in the Wasserstein variant a critic, that tries to distinguish real from fake. Standard GANs are notorious for training instability, but the Wasserstein formulation measures the Earth Mover’s distance between distributions, and the gradient penalty enforces the Lipschitz constraint that the critic must satisfy. The result is a generator that converges reliably and produces high-confidence synthetic positive samples that faithfully reproduce the statistical fingerprint of genuine mineralized voxels. This generative augmentation effectively multiplies the small positive sample set without simply copying it.

The second component confronts the unreliable negatives. The framework applies bagging-based positive–unlabeled learning, abbreviated BPUL, a weakly supervised strategy in which the vast pool of unlabeled voxels is treated not as confirmed negatives but as an ambiguous mixture containing both hidden positives and true negatives. BPUL works by estimating the positive class prior, the proportion of positives hiding within the unlabeled set, and then iteratively identifying which unlabeled samples are most confidently negative. By training ensembles of models on bootstrap resamples, the bagging procedure stabilizes these estimates and allows reliable negatives to be selected with far less contamination than random sampling. The resulting training set, the authors argue, more closely approximates the true underlying distribution of the subsurface.

The backbone of the prediction engine is a 3D ResUNet, a volumetric convolutional network that descends from the U-Net architecture originally developed for biomedical image segmentation. U-Net’s encoder–decoder design, with skip connections that preserve fine spatial detail across the network, adapts naturally to voxel classification. The residual variant adds identity shortcut connections, in the spirit of ResNet, that allow gradients to flow through very deep stacks of layers and mitigate vanishing-gradient problems. For the Duobaoshan model, the 3D ResUNet served as the primary predictor, consuming multi-channel voxel cubes built from the district’s integrated multi-source geoscience dataset, which combines geological, geophysical, and drilling information into a single 3D evidence stack.

To rigorously test whether each ingredient mattered, the researchers designed four experimental configurations. The first was a baseline combining random negative sampling with the 3D ResUNet. The second replaced random sampling with BPUL to isolate the effect of reliable negative selection. The third added WGAN-GP generative augmentation to the baseline to test the value of synthesizing positive samples. The fourth, the full coupled framework, combined WGAN-GP augmentation, BPUL negative selection, and the 3D ResUNet. Performance was evaluated using standard classification metrics, including the area under the receiver operating characteristic curve, the F1 score, and precision.

The results were unambiguous. The coupled WGAN-GP plus BPUL framework outperformed each individual strategy across all evaluated metrics. Models trained with random negatives were undermined by label noise, while models using only generative augmentation or only weakly supervised negative selection improved partially. Only when the two components worked together did the 3D ResUNet achieve its best performance, demonstrating that the two techniques are complementary rather than redundant. The generated positives expand the representation of what ore looks like in feature space, while BPUL’s reliable negatives teach the model what barren rock looks like, and together they stabilize training and sharpen the network’s decision boundary.

Equally important was the character of the predicted targets themselves. The exploration targets delineated by the coupled framework exhibited higher spatial coherence and volumetric connectivity than those produced by the baseline methods. In practical terms, the high-probability voxels clustered into geologically plausible, connected bodies rather than scattered, isolated hotspots. This matters enormously for exploration managers, because coherent 3D volumes translate directly into drill targets with realistic geometry, whereas fragmented predictions are difficult to act on and often reflect model artifacts rather than genuine mineralization. The authors note that this combination improves training stability and generalization under conditions of extremely limited positive samples and unlabeled negatives, providing a reusable modeling workflow for 3D exploration targeting in Duobaoshan and similar ore districts.

The broader significance extends well beyond one copper district in Heilongjiang. The sample scarcity problem is universal in mineral exploration: every prospective camp on Earth has few confirmed deposits relative to the volume of unexplored rock, and no one can certify barren ground. A workflow that treats unlabeled data honestly, extracts reliable negatives algorithmically, and manufactures trustworthy synthetic positives from limited real ones offers a template for any district where deep, concealed mineralization is the next frontier. As the global demand for copper accelerates with electrification and renewable energy infrastructure, and as shallow deposits become exhausted, methods like this generative-weakly supervised framework may determine which exploration programs find the next giant deposit first. For Duobaoshan itself, the study delivers a set of 3D targets with the spatial coherence that drill planners look for, and for the field of computational geoscience, it demonstrates that two techniques borrowed from the machine-learning literature, married thoughtfully to a volumetric segmentation network, can solve a problem that has long hampered 3D prospectivity mapping.