A robust workflow for regional porphyry targeting; Application to the QUEST-South project area in southern British Columbia

This study utilizes geophysical and geochemical datasets from the QUEST-South project (Geoscience BC, 2009), a geoscience initiative covering the southern Quesnel Terrane in British Columbia, Canada. The location of the project is presented on Figure 1. The geochemical dataset is further complemented by the regional compilation from the British Columbia Geological Survey (Hans, 2020).

The ground gravity data comprises the Natural Resources Canada (NRCAN, 2024) gravity compilation. Gravity data presents a variable station distribution over the project area, is fully processed to free air, complete Bouguer and isostatic corrected data, and interpolated onto 2 km resolution grids.

The QUEST-South airborne gravity survey (Geoscience BC, 2009) was flown in 2009 with a fixed-wing aircraft along 2 km spaced E-W oriented flight lines. The terrain clearance was defined by a pre-planned drape surface designed to guide the aircraft over the topography in a consistent manner as close to target clearance as possible. The achieved mean terrain clearance is 658.8 m.

A key component of this work is the focus on the petrophysical characteristics of the lithologies found within the project area and to use these properties to build a petrophysically-based geological model. Geophysical inversions of the magnetic and gravity data are thus implemented to generate physical rock property models of magnetic susceptibility and density. Given the scale of the project area the inversion approach is simplified to models producing apparent magnetic susceptibility and density models.

The initial inversions are unconstrained and informed on spatial relationships between magnetic susceptibility, density and regional geology (Cui et al., 2019). These models contribute to the interpretation of a series of key petrophysical domains shown in Figure 3 (a).

These interpreted domains are further leveraged with constrained inversions, which have their initial model values defined by the computed per-domain means of the unconstrained models. The resulting constrained apparent magnetic susceptibility and density models are shown in Figure 3 (b) and (c) with their anomalous values i.e. the difference between the constraining model starting value and the resulting value after inversion.

We propose that the magnetic susceptibility and density anomalies derived in this way may be more closely related to alteration than the absolute value of magnetic susceptibility and density in either constrained or unconstrained inversions.

Geochemical rock analysis compilations from the British Columbia Geological Survey (Hans, 2020) have been extracted and compiled, with only the most recent values retained.

The geochemical values have been interpolated using geophysical data as secondary inputs. The interpolation method is a “Neural Inverse Distance Weighting” approach, illustrated on Figure 5 which trains a rotation matrix based on the distances between points and then assigns weights according to trainable range and nugget parameters.

The data is first standardized using the Centered Log-Ratio (CLR) transformation (Aitchison, 1986) to ensure compositional coherence and comparability. Then, the transformed values are interpolated and projected onto the geophysical data grid.

Machine learning consists of statistical models that aim to define relationships within a dataset and apply them to new, unseen data. However, in the context of mineral targeting, the ability to infer these relationships is limited by high variability in geological setting and data availability across regions. Therefore, the primary expectations for machine learning in mineral targeting should be (1) identifying correlations within the training dataset and (2) projecting those correlations on a map.

Mineral targeting datasets present statistical challenges that can lead to overfitting or other biases. The issues include (1) spatial correlation, (2) highly imbalanced positive targets, and (3) the presence of inherent false negatives in the dataset. To address these challenges, we apply the following strategies:

1. Regional Data Partitioning: The dataset is divided into geographical regions, as presented in Figure 6, with each deposit area considered separately. One region is designated for testing, while the others are used for training and cross-validation to determine optimal model parameters.
2. Handling Imbalanced Data: We use an Imbalanced Random Forest, presented in Figure 7, to mitigate the issue of highly imbalanced classes (Lemaitre et al., 2017). This implementation is particularly robust for such datasets, as each tree in the random forest is trained on a different balanced subset, randomly selected from the original data.
3. A major cause of overfitting is the improper selection of input properties. To address this, we used the Optuna algorithm (Akiba et al., 2019), which integrates hyperparameter tuning directly into the training process. To determine which properties to include or exclude, we define the activation of each property as an input parameter (0 or 1) and allow the model to converge, as illustrated in Figure 8. This results in the automatic selection of relevant properties based on the training dataset.

This approach effectively eliminates properties that introduce bias or provide little value for model training. However, some properties can lead to overfitting because they are naturally correlated with the target variable (e.g., copper concentration when predicting copper deposits). These can obscure other meaningful and sought-after relationships. A geologist must manually discard such properties using domain knowledge to ensure a balanced and interpretable model.

This study demonstrates how machine learning techniques, combined with geophysical and geochemical datasets, can enhance mineral exploration targeting in the southern Quesnel Terrane. By leveraging publicly available data in the QUEST-South project area, the study successfully identified new prospective areas while validating known deposits.

Key findings include:

• The robust magnetic susceptibility and density models generated from constrained inversions that leverage interpreted petrophysical domains,
• The successful interpolation of geochemical values using geophysical data as secondary inputs,
• Three machine learning model, demonstrating strong predictive capabilities,
• The delineation of several potential targets, including anomalies in the north of the area, and trends in the continuity of the Copper Mountain deposit.

A crucial aspect of this study is proper data preparation, extracting meaningful information from geophysical surveys and interpolating geochemical values using geophysically based models. Morevover, by integrating structured data preparation, balanced training, rigorous spatial validation, and automatic properties selction this study provides a reliable framework for mineral prospectivity modelling.