Development of machine learning models and design of an analytical application for XRF-based nickel laterite classification

Nickel is a crucial metal with continuously increasing global demand, primarily driven by the stainless-steel sector and battery development for net-zero emission policies. Nickel laterite deposits, formed from the intensive weathering of ultramafic rocks in tropical climates, are the world's main source of nickel. Accurate classification between the Fe-rich limonitic zone, and Mg- and Si-rich saprolitic zone within these vertical profiles is vital for determining efficient processing routes, such as High-Pressure Acid Leaching (HPAL) for limonite or pyrometallurgy for saprolite. This research aims to develop and validate a machine learning-based methodology for the automatic and objective classification of nickel laterite ores using chemical composition data from Wavelength Dispersive X-Ray Fluorescence (WDXRF) analyses. The research methodology utilized a dataset of 524 laterite samples. Descriptive statistical analysis of the dataset revealed the heterogeneous chemical structure of the deposit, with average concentrations of 1.16% Ni, 32.11% Fe, and 4.15% Mg. Correlation heatmap analysis further indicated complex inter-element interactions, such as a very strong positive correlation between Fe and Cr (r = 0.94) and a very strong negative correlation between Mg and Fe (r = -0.95), confirming the necessity of a multivariate analysis approach. The data were analyzed using the unsupervised K-means algorithm and four supervised classification models: Random Forest (RF), Gradient Boosting Classifier (GBC), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). The results show that the Elbow method, used to determine the optimal number of clusters, clearly formed an 'elbow' at K=3. This finding, supported by the practical interpretation of the Silhouette score, confirmed the existence of three geochemically consistent clusters: the limonite, saprolite, and transition zone profiles. In the evaluation of supervised models, the ensemble methods (RF and GBC) demonstrated absolute superiority by achieving a perfect 100% accuracy, precision, and F1-score. The SVM and MLP models also performed very well, with accuracies of 96% and 94%, respectively, but showed minor errors in distinguishing between adjacent classes. Feature importance analysis consistently highlighted nickel content (%Ni) as the most dominant predictor across all models, confirming the geological validity of the machine learning process. As the culmination of the research, the top-performing RF model was successfully implemented into a functional web application. This study proves that machine learning offers a reliable solution to transform the complex ore classification process into a rapid, objective, and automated procedure, thus presenting a practical prototype with significant potential to enhance efficiency in the mining industry.