基于数据驱动的南岭地区花岗岩岩体含矿性判别  

Ore-bearing discrimination of granite rock masses in the Nanling area via data-driven models

作　　者：花旗 夏庆霖[1,2] 刘奇锋 HUA Qi;XIA Qinglin;LIU Qifeng(School of Earth Resources,China University of Geosciences(Wuhan),Wuhan 430074,China;.State Key Laboratory of Geology Process and Mineral Resources,China University of Geosciences(Wuhan),Wuhan 430074,China)

机构地区：[1]中国地质大学(武汉)资源学院,武汉430074 [2]中国地质大学(武汉)地质过程与矿产资源国家重点实验室,武汉430074

出　　处：《地质科技通报》2025年第1期332-345,共14页Bulletin of Geological Science and Technology

基　　金：国家自然科学基金项目(41672328);青海省有色地质勘查局科研项目(2022jgky02)。

摘　　要：花岗岩作为成矿作用的重要参与者,对它的研究有利于了解钨锡成矿作用的地球化学过程并区分岩体的含矿性。收集了南岭地区含钨花岗岩、含钨锡花岗岩和不含矿花岗岩的主量元素和稀土元素数据,共42个岩体466组数据。总结对比了3类岩体的地球化学特征,从数据驱动和机器学习的角度区分了3类岩体的含矿性和岩石地球化学特征之间的关联,运用受限玻尔兹曼机来训练自编码神经网络以消除主量元素和稀土元素之间量级的差别,并且提取中间特征,再将中间特征输入随机森林和多层BP神经网络,建立AE-RF和AE-BP岩体含矿性分类模型。通过随机森林输出了分类特征重要性。结果表明,含钨花岗岩的演化程度最高,含钨锡花岗岩次之,不含矿花岗岩最低。2种模型在测试集上都有很高的正确率(平均>90%),并且在盲测试集上AE-BP模型的实际运用效果更好。随机选择了6组岩体作为盲测试集,20个岩体中有13个岩体正确率>80%,有2个岩体正确率为[70%,80%],有2个岩体正确率为[50%,70%]。还有4个岩体正确率<50%。铁锰磷钙镁等主量元素和轻重稀土元素是区分3类岩体的重要特征。机器学习能够很好地反映出3类花岗岩的含矿性。地球化学特征的相似性会导致模型错误分类,陂头岩体有一定的成矿潜力。铁锰磷钙镁等主量元素决定了岩体能否含矿,而轻稀土元素是区分含钨岩体和含钨锡岩体的重要指标,表明岩浆的分异演化程度决定了岩体能否含矿,而幔源物质的加入是区别岩体含钨还是含钨锡的特征。[Objective]As a significant component of mineralization,granite plays a critical role in understanding the geochemical processes of tungsten-tin mineralization and distinguishing the ore-bearing of rock masses.[Methods]This study collected both major and rare earth element data from tungsten-bearing granite,tungsten-tinbearing granite,and non-ore-bearing granite in the Nanling area,with 466 groups of datasets of 42 rock masses in total.The geochemical characteristics among three types of rock masses were summarized and compared.A datadriven approach integrating with machine learning techniques was used to explore the relationship between orebearing properties and geochemical characteristics.The restricted Boltzmann model was employed to train an autoencoder neural network to eliminate dimensional differences between major and rare earth elements,allowing for feature extraction.Then,these features were subsequently input into random forests and multilayer BP neural networks to develop AE-RF and AE-BP classification models for ore-bearing evaluation.The importance of classification features was derived from random forests.[Results]Results indicate that tungsten-bearing granite exhibits a slightly higher evolution degree compared against tungsten-tin-bearing granite,and non-ore-bearing granite displays the lowest evolution degree.Both two models achieved high accuracy(>90%)on testing datasets,with the better application performance of AE-BP model on the blind testing set.Six rock masses were randomly selected as the blind test set,13 out of 20 groups of rock masses had an accuracy rate above 80%,two of them had accuracy between 70%and 80%,and two had accuracy between 50%and 70%,while rest four rock masses showed accuracy below 50%.Major elements such as iron,manganese,phosphorus,calcium,and magnesium,along with light and heavy rare earth elements,were important for distinguishing the three rock mass types.Machine learning effectively identified the ore-bearing properties of these granite types.[Conclusion]Results re

关 键 词：高分异花岗岩 数据驱动模型 机器学习 南岭钨锡多金属矿床 岩体含矿性评价 

分 类 号：P628.4[天文地球—地质矿产勘探]