机器学习在矿床学中的应用进展

Advances in the Application of Machine Learning to Ore Deposit Research

摘要: 机器学习方法具有处理复杂高维数据的能力，且具有非线性模式识别优势，因此，在矿床学多个核心领域的应用日益深入。本文从矿床成因与类型判别、蚀变矿物光谱识别以及矿产远景预测三个方面，系统梳理了近年来机器学习在矿床学研究中的主要进展。在矿床成因与类型判别方面，以黄铁矿、磁铁矿、闪锌矿和磷灰石等矿物的微量元素数据为研究对象，研究者利用随机森林、支持向量机、XGBoost等算法构建的分类模型准确率普遍超过90%，SHAP等可解释性工具的引入也推动了“黑箱”模型向地质可解释方向的转变。在蚀变矿物光谱识别方面，短波红外(SWIR)光谱、高光谱遥感和蚀变矿物化学与机器学习的结合，将蚀变信息从定性描述转变为定量成矿指示工具，为多尺度、多方法协同的隐伏矿体精细定位提供了技术支撑，在深部金矿、斑岩铜矿等多类矿床勘查中均取得显著进展。在矿产远景预测方面，多源数据融合、和三维成矿预测成为当前的重要发展方向，预测模式正从传统的二维平面靶区圈定快速向三维深部靶区定位转变。目前，机器学习在矿床学中的应用仍受数据质量参差、地质知识耦合深度不足等因素制约。未来，将地质先验知识有效嵌入数据驱动模型，有望成为推动这一领域实现突破的关键路径。

Abstract: Machine learning methods possess the capability to process complex, high-dimensional data and offer advantages in nonlinear pattern recognition, leading to their increasingly deep application across several core areas of ore deposit research. This paper systematically reviews recent advances in the application of machine learning to ore deposit studies from three perspectives: ore genesis and deposit type discrimination, alteration mineral spectral identification, and mineral prospectivity mapping. In the area of ore genesis and deposit type discrimination, researchers have used trace element data from minerals such as pyrite, magnetite, sphalerite, and apatite to build classification models using algorithms including Random Forest, Support Vector Machine, and XGBoost, with accuracies generally exceeding 90%. The introduction of interpretability tools such as SHAP has also facilitated the transition of “black-box” models toward geologically interpretable frameworks. In alteration mineral spectral identification, the integration of shortwave infrared (SWIR) spectroscopy, hyperspectral remote sensing, and alteration mineral chemistry with machine learning has transformed alteration information from qualitative descriptions into quantitative metallogenic indicator tools. This provides technical support for multi-scale, multi-method synergistic precise localization of concealed ore bodies, with notable progress achieved in the exploration of various deposit types including deep gold deposits and porphyry copper deposits. In mineral prospectivity mapping, multi-source data fusion and three-dimensional metallogenic prediction have emerged as important current development directions, with prediction paradigms rapidly shifting from traditional two-dimensional planar target delineation to three-dimensional deep target localization. At present, the application of machine learning in ore deposit research remains constrained by factors such as uneven data quality and insufficient depth of coupling with geological knowledge. In the future, effectively embedding geological prior knowledge into data-driven models is expected to become a key pathway for achieving breakthroughs in this field.