Scholarlink

<p>Carbonatite-related rare earth element (REE) deposits represent the primary global source of REEs, with their geochemical features providing key insights into mineralization processes. However, the complexity and nonlinearity of REE mineralization challenge traditional linear discrimination methods. This paper introduces an interpretable machine learning approach for the nonlinear discrimination of REE mineralization potential. A total of 178 whole-rock geochemical samples from 21 global carbonatite-related REE deposits were analyzed, employing four models: support vector machine, gradient boosting, backpropagation neural network, and K-nearest neighbors (KNN). The KNN model, trained on trace element data, demonstrated superior performance. Permutation importance analysis identified Nb/Y, Ba, REE, and Nd as key geochemical indicators. Nonlinear decision boundaries derived from the approach effectively delineated mineralization potential zones in the geochemical feature space. Independent validation based on 157 samples from the Bayan Obo REE deposit (North China Craton; Proterozoic to Mesozoic carbonatite complexes) confirmed the robustness and generalizability of the model, particularly when using the Ba–Nd, REE–Ba, and Nb/Y–Ba feature pairs. This approach provides a geologically interpretable tool for REE classification, overcoming the limitations of linear discrimination methods and offering new theoretical and methodological insights into REE mineralization assessment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Interpretable Machine Learning Approach for Nonlinear Discrimination of Carbonatite-Related REE Mineralization

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Xiaochun Lin,
  • Yan Liu

摘要

Carbonatite-related rare earth element (REE) deposits represent the primary global source of REEs, with their geochemical features providing key insights into mineralization processes. However, the complexity and nonlinearity of REE mineralization challenge traditional linear discrimination methods. This paper introduces an interpretable machine learning approach for the nonlinear discrimination of REE mineralization potential. A total of 178 whole-rock geochemical samples from 21 global carbonatite-related REE deposits were analyzed, employing four models: support vector machine, gradient boosting, backpropagation neural network, and K-nearest neighbors (KNN). The KNN model, trained on trace element data, demonstrated superior performance. Permutation importance analysis identified Nb/Y, Ba, REE, and Nd as key geochemical indicators. Nonlinear decision boundaries derived from the approach effectively delineated mineralization potential zones in the geochemical feature space. Independent validation based on 157 samples from the Bayan Obo REE deposit (North China Craton; Proterozoic to Mesozoic carbonatite complexes) confirmed the robustness and generalizability of the model, particularly when using the Ba–Nd, REE–Ba, and Nb/Y–Ba feature pairs. This approach provides a geologically interpretable tool for REE classification, overcoming the limitations of linear discrimination methods and offering new theoretical and methodological insights into REE mineralization assessment.