The core problem the researchers set out to solve is one that has haunted mineral exploration for decades: dimensional heterogeneity. Surface geochemical surveys—measurements of trace elements in soils, stream sediments, or rock chips—are inexpensive, spatially extensive, and readily available almost anywhere on Earth. They carry faint chemical fingerprints of buried hydrothermal systems, because ore-forming fluids leak elements upward and outward along faults and fractures. Three-dimensional geological models, by contrast, capture the deep architecture of faults, intrusions, and alteration zones where gold actually precipitates, but they are built from limited boreholes and geophysical inversions, making them sparse and uncertain. Putting these two data types together is not straightforward: one describes a two-dimensional skin over the landscape, the other a three-dimensional volume of the subsurface, and traditional machine learning methods struggle to reconcile data that live in different dimensional spaces and encode different physical processes.

The team’s answer is a multiscale mixture-of-experts graph convolutional network, or multiscale MoE-GCN. Graph neural networks are a class of deep learning architectures that operate not on regular grids or images but on graphs—collections of nodes connected by edges—which makes them naturally suited to irregularly sampled geoscience data. In this framework, the researchers constructed a specifically designed heterogeneous graph in which two fundamentally different kinds of nodes coexist: spatial voxels, the small three-dimensional cells that discretize the subsurface volume, and multi-dimensional predictive variables, including geochemical indicators measured at the surface. Edges in the graph encode both spatial relationships and genetic associations—the explicit hypothesis that a geochemical anomaly at the surface may be connected, through fluid pathways, to alteration and mineralization at depth. Because all variables and voxels are unified within a single topological system, information can propagate between surface and subsurface nodes during training, allowing the network to learn how chemistry above relates to structure below.

Within this heterogeneous graph, the model employs a dual-scale graph convolution module that extracts representations at two distinct levels of granularity. At the fine scale, convolutions operate on the 3D structural geometry and on outputs of ore-fluid simulations—numerical reconstructions of how hydrothermal fluids flowed through the fault networks of the deposit. At the coarse scale, the network processes the surface geochemical data, which naturally describe broader, smoother patterns of elemental dispersion. Separating these scales prevents the fine-grained, high-dimensional voxel features from drowning out the lower-resolution geochemical signals, and vice versa. Each scale effectively acts as a specialized lens on the mineralization system, capturing features that a single-scale model would either blur or discard.

The second architectural innovation is the cross-modal mixture-of-experts fusion mechanism. Mixture-of-experts is a technique from machine learning in which multiple specialized subnetworks—experts—are trained in parallel, and a gating function learns to route each input to the expert or experts best suited to handle it. The concept has powered some of the largest language models ever built, but applying it to geology required rethinking what the “experts” should represent. In the multiscale MoE-GCN, the experts specialize in deciphering the intrinsic correlations between surface geochemical indicators and the deep structural–hydrothermal mineralization processes that produced them. Rather than forcing a single monolithic network to model every possible relationship between geochemistry and geology, the MoE mechanism adaptively assigns different combinations of experts to different spatial contexts, and adaptively weighs the heterogeneous features to mitigate the inherent uncertainty that pervades three-dimensional deep prospectivity prediction.

The testing ground for the framework was the Xiadian gold deposit, a member of the world-class Jiaodong gold province in eastern China—the country’s premier orogenic gold district, where mineralization is structurally controlled by major fault zones, including the Zhaoping fault that runs through the Xiadian-Jiangjiayao sector. Jiaodong-type deposits formed through hydrothermal fluid circulation deep in the crust, with gold precipitating along faults and their damage zones, typically accompanied by sericitic and other alteration halos. This geological setting makes Xiadian an ideal case study for the data-scarcity problem: decades of mining and exploration have produced detailed 3D structural knowledge, but the deep extensions of the deposit remain poorly constrained, and surface geochemical anomalies hold clues that conventional 3D models ignore.

When the team benchmarked the multiscale MoE-GCN against baseline methods that depend solely on sparse 3D structural models, the improvement was striking. By incorporating surface geochemistry as a constraint, the fused model reached an area under the curve of 0.97, a near-perfect discrimination between prospective and non-prospective voxels. More telling than the aggregate metric was the qualitative result: the fusion of surface geochemical data enabled the model to identify mineralization targets that had been overlooked by models built on single-source 3D structural information alone. Critically, those newly flagged targets showed high consistency with known ore-controlling geological structures, suggesting that the model was not merely fitting noise in the geochemical data but genuinely recovering the physical linkages between surface chemistry and deep ore formation.

The implications extend well beyond one deposit in Shandong Province. As easily discovered, near-surface orebodies are progressively exhausted, the global mining industry is being pushed toward increasingly deep and concealed targets, a trend that makes brownfields exploration—the systematic search for new ore around existing mines—both economically vital and technically demanding. Yet deep exploration is chronically starved of direct data: drilling to significant depths is expensive, and geophysical methods often lack the resolution to distinguish ore-bearing structures from barren ones. Surface geochemical exploration data, by contrast, are “easily accessible, have extensive spatial coverage, and contain crucial signals indicative of buried hydrothermal mineralization,” as the authors note. A method that can reliably translate those shallow signals into deep three-dimensional predictions effectively converts one of the cheapest datasets in exploration into a constraint on one of the most uncertain problems.

The study also contributes to a broader movement within computational geoscience toward multimodal and heterogeneous machine learning. Earlier work by members of the same group demonstrated 3D prospectivity modeling with convolutional neural networks, transfer learning, hidden Markov models, and end-to-end transformers for inferring deep alteration zones from surface geochemistry. Other researchers have applied graph attention networks and self-supervised graph contrastive learning to geochemical anomaly recognition. What distinguishes the present contribution is its explicit treatment of dimensional heterogeneity as a graph-topology problem: rather than interpolating 2D geochemistry into 3D space or flattening 3D geology into 2D maps, the heterogeneous graph allows both data types to retain their native structure while still exchanging information. The mixture-of-experts fusion then handles the fact that different regions of the study area may express different relationships between the two data domains—a flexibility that fixed-weight fusion schemes lack.

There are practical caveats and paths forward. The exploration datasets used in the study are confidential and not publicly available, which limits external validation for now, though the authors have released their complete source code on GitHub under the repository Jiaxuan0908/MMGCN-Prospectivity, inviting replication and adaptation by other groups. The method also presupposes a reasonably well-constructed 3D geological model as one of its inputs, so its value scales with the quality of conventional geological mapping and modeling. And like all machine learning systems in the geosciences, its predictions are probabilistic guides rather than guarantees; the 0.97 AUC was achieved on known data and will need to be tested prospectively in genuinely unexplored ground. Still, the authors argue that the multiscale MoE-GCN can incorporate surface geochemical data as a vital complement to deep 3D mineral prospectivity modeling, thereby alleviating the persistent challenge of data scarcity in deep-seated exploration.

For a discipline in which a single deep borehole can cost hundreds of thousands of dollars and misallocated drilling campaigns can sink exploration budgets, an algorithm that squeezes more information out of already-collected surface samples is more than an academic curiosity. If the approach generalizes to other structurally controlled hydrothermal deposits—gold, copper, tungsten, and beyond—the quiet chemistry of the ground surface may become one of the most powerful windows geologists have into the hidden architecture of the deep crust. The research was supported by the National Major Science and Technology Projects of China, the National Natural Science Foundation of China, and the Science and Technology Innovation Program of Hunan Province, with correspondence directed to Zhankun Liu at Central South University.