significantly lower than the variance of the actual data. While this minimizes the global error variance, it introduces a conditional bias that can be catastrophic for selective mining: high-grade ore is diluted into waste (underestimation), and waste is processed as ore (overestimation). This paper proposes a Hybrid Geostatistical-Machine Learning Framework that bridges this gap. We introduce a methodology where Kriging provides a spatially coherent baseline, and an ML model acts as a "residual corrector," adjusting the grade based on the local geometric texture of the deposit. To achieve this, we developed a specialized computational engine in Rust, capable of calculating complex descriptors—such as Multifractal Detrended Fluctuation Analysis (MFDFA) and 3D Wavelet Transforms—on millions of block model centroids in minutes, a task previously computationally prohibitive. 2. METHODOLOGY The proposed workflow, termed "Geological Co-Stacking," consists of three integrated stages: Data Preprocessing, Geometric Feature Extraction, and Residual Stacking. 2.1 Data Description and Characterization The case study focuses on a real epithermal silver deposit (anonymized). The dataset comprises diamond drill hole (DDH) composites regularized to 2.0m lengths. Exploratory Data Analysis (EDA) reveals a highly skewed, log-normal distribution with a Coefficient of Variation (CV) of 3.24, indicating extreme variability. A critical characteristic of this deposit is the concentration of value: the top 1% of samples (grades > 205 g/t) contain 26.07% of the total metal content. In traditional workflows, such outliers are often subjected to aggressive top-capping to prevent variance smearing in Kriging, which artificially reduces the total metal inventory. 2.2 High-Performance Geometric Feature Extraction (The Rust Engine) To enable the ML model to "see" geology, we developed a parallelized engine using the Rust programming language. For every point to be estimated (block centroid), the engine queries the knearest neighbors and computes a vector of 26 geometric features. Key algorithms include: • Local Curvature (Laplacian Proxy): Calculates the second-order derivative of the grade field. This feature allows the model to distinguish between "peaks" (high-grade shoots), "valleys," and "slopes," effectively identifying the structural controls of mineralization. • 3D Wavelet Energy: We implement a separable 3D Haar Wavelet Transform on the local neighborhood. By decomposing the local grade distribution into high-frequency and lowfrequency components, this feature quantifies the "texture" of the mineralization. High wavelet energy typically correlates with chaotic, high-grade brecciation zones. • Multifractal Spectrum (MFDFA): We compute the mass exponent related to the generalized Hurst exponent for the drill hole data sequence. This measures the "roughness" and persistence of the grade series, helping the model distinguish between true high-grade continuity and random noise (nugget effect). • Convex Volume Density: Calculates the density of metal within the convex hull of the nearest neighbors, serving as a proxy for identifying massive vs. disseminated mineralization. 2.3 The Stacking Architecture The final estimation is performed by an XGBoost (Extreme Gradient Boosting) meta-model. The input vector X for the model includes: 1. The Ordinary Kriging estimate (Baseline). 2. The 26 geometric features extracted by the Rust engine.
RkJQdWJsaXNoZXIy MTM0Mzk2