Masked Mineral Modeling: Continent-Scale Mineral Prospecting via Geospatial Infilling

TL;DR

Confronts the challenge of continent-scale mineral prospecting with incomplete records. Proposes Masked Mineral Modeling (M3), a ResNet U-Net infilling framework that fuses mineral maps with geophysical and agronomic layers. Demonstrates Dice coefficient of $0.31 \pm 0.01$ and recall of $0.22 \pm 0.02$ on test data at $1\times 1\text{ mi}^2$ spatial resolution, and shows auxiliary data improve performance and enable evaluation in regions without minerals. Analyzes inter-mineral dependencies and scalability, and discusses future extensions like SRMM.

Abstract

Minerals play a critical role in the advanced energy technologies necessary for decarbonization, but characterizing mineral deposits hidden underground remains costly and challenging. Inspired by recent progress in generative modeling, we develop a learning method which infers the locations of minerals by masking and infilling geospatial maps of resource availability. We demonstrate this technique using mineral data for the conterminous United States, and train performant models, with the best achieving Dice coefficients of $0.31 \pm 0.01$ and recalls of $0.22 \pm 0.02$ on test data at 1$\times$1 mi$^2$ spatial resolution. One major advantage of our approach is that it can easily incorporate auxiliary data sources for prediction which may be more abundant than mineral data. We highlight the capabilities of our model by adding input layers derived from geophysical sources, along with a nation-wide ground survey of soils originally intended for agronomic purposes. We find that employing such auxiliary features can improve inference performance, while also enabling model evaluation in regions with no recorded minerals.

Figures (22)

• Figure 1: General model architecture and data pipeline. For a location $(x_j,y_k)$ in the $i^{\text{th}}$ (50 mi)$\times$(50 mi) context window, the availability of resource $\ell$ is marked with a per-pixel binary flag $z^{(i)}_{jk\ell}\in\{0,1\}.$ The mineral data layers are masked, stacked with geophysical and agronomic inputs, and jointly passed through either a ResNet of depth $H$ (shown) or a ViT. Masked minerals are infilled by the network, and the output is a per-pixel probability $p(z^{(i)}_{jk\ell})$ of the presence of resource $\ell$. We train to maximize Dice coefficients computed over the masked region. Due to their combination of spatial sparsity and feature richness, the agronomic data for a given pixel pass through a feature reduction network $f$ which maps them to vectors of length $K$ at that location.
• Figure 2: Example input features and model outputs. The displayed predictions are generated by the model trained in Experiment \ref{['sec:exp-georac']} with $N=10$K, $H=152$, and $K=64$. The left panel shows all possible forms of input data for this context window. The first 10 layers consist of the availability of 10 minerals [Gold, Silver, Zinc, Lead, Copper, Nickel, Iron, Uranium, Tungsten, Manganese] on (50 mi)$\times$(50 mi) regions. The next 6 layers consist of geophysical inputs [Rock type, Fault presence, Fault slip rate, Geological age (upper/lower limit), Elevation]. Finally, the agronomic data is embedded and included. The panel on the right shows the two possible masking strategies, mineral and spatial, and example predictions. Mineral masking removes an entire mineral layer, while spatial masking removes a random rectangular region. The blue points are ground truth outside of the masked region, the green points are ground truth within the masked region, and the red is prediction.
• Figure 3: Comparing M3 to prior work. Our proposed method, M3, generally outperforms prior work from both learning-based approaches (ViT), and classical geostatistical approaches (Kriging). We observe that ViT struggles to fit the data as effectively due to a lack of spatial inductive bias, and that a ResNet-based approach is better-suited to inference over sparse clustered mineral deposits. Kriging fits well to training data due to its use of a Gaussian Process, but completely fails to generalize to unseen samples. These conclusions hold especially for ID (left) generalization. For OOD generalization (right), ViT required 5$\times$ as many gradient steps to converge to equivalent validation and test performance, with significantly lower training performance. The OOD region is a 90K mi$^2$ square centered at the McDermitt Caldera, and the validation region is the surrounding 50-mi annulus.
• Figure 4: Scaling properties of M3. results: We study how the M3 model performance scales as a function of architecture depth $H$ (Left) and dataset size $N$ (Right). Dice coefficients (Top) over the masked regions, averaged over each mineral layer, and the recall on test data (Bottom) over the masked regions in each considered scenario. We observe favorable scaling, with improving performance with bigger models and more data.
• Figure 5: Adding auxiliary data to M3 (Left). Dice coefficients on training and test data as a function of input layer combinations ($N=10$K, $A=0.8$, $H=152$, $K=64$, 30K gradient steps). The inclusion of auxiliary data leads to faster training and significant generalization improvements, driven by geophysical data. Visualizing M3 predictions over the continental US (Center). An M3-generated map of resources in the conterminous United States (EPSG:5070). The model including agronomic and geophysical input layers in Experiment \ref{['sec:exp-georac']} was evaluated 150 times over a visualization dataset whose (50 mi)$\times$(50 mi) tiles covered the United States. Model outputs from the masked regions (red) were summed over mineral layers, with ground truth overlaid in blue. The model predicts 4 mainland clusters (Right) of iron and zinc resources where no prospecting records on the top 10 minerals exist; 3 match existing sand, silica, stone, and clay mines. (See Appendix \ref{['app:extramaps']} for details.)