基于机器学习的磷灰石源岩识别方法  

A Machine Learning-based Method for Identifying Apatite Source Rocks

作　　者：刘嘉情 钟世华 李三忠[1,2] 戴黎明 索艳慧[1,2] 郭广慧 牛警徽[1,2] 薛梓萌 黄宇 LIU Jiaqing;ZHONG Shihua;LI Sanzhong;DAI Liming;SUO Yanhui;GUO Guanghui;NIU Jinghui;XUE Zimeng;HUANG Yu(College of Marine Geosciences,Ocean University of China,MOE Frontiers Science Center for Deep Ocean Multispheres and Earth System,MOE Key Lab of Submarine Geosciences and Prospecting Techniques,Qingdao 266100,Shandong,China;Laboratory for Marine Mineral Resources,Qingdao Marine Science and Technology Center,Qingdao 266237,Shandong,China)

机构地区：[1]中国海洋大学海洋地球科学学院,深海圈层与地球系统教育部前沿科学中心,海底科学与探测技术教育部重点实验室,山东青岛266100 [2]青岛海洋科技中心海洋矿产资源评价与探测技术功能实验室,山东青岛266237

出　　处：《大地构造与成矿学》2025年第2期467-480,共14页Geotectonica et Metallogenia

基　　金：国家自然科学青年基金项目(42203066);山东省自然科学青年基金项目(ZR2020QD027)联合资助。

摘　　要：正确识别花岗岩成因类型一直是国内外地质学领域关注的焦点,对研究大陆壳的形成和演化过程以及认识金属矿床成因、指导找矿勘查等工作均具有重要意义。本次研究在汇编I型和S型花岗岩的磷灰石微量元素数据的基础上,借助两种有监督机器学习算法——支持向量机与随机森林,建立了基于磷灰石微量元素特征区分I型和S型花岗岩的方法。研究选取磷灰石的La、Ce、Pr、Nd、Sm、Eu、Gd、Tb、Dy、Ho、Er、Tm、Yb、Lu、Sr、Y、δEu、Sr/Y、La/Yb共19种特征用于机器学习训练,获得的分类准确率均不低于0.99,证实运用磷灰石成分可以有效识别花岗岩类型。除此之外,基于准确率更高的支持向量机模型,提出了9种花岗岩成因类型二元判别图解,这些图解在识别I型和S型花岗岩时准确率均高于0.90。本研究不仅为花岗岩成因类型识别提供了新的途径,还为利用其他副矿物开展花岗岩成因研究提供了思路和方法参考。相关机器学习代码已上传至GitHub,地址为https://github.com/ShihuaZhong/Apatite2023MLcode。Correct identification of the genetic types of granites has long been a focal point of geological research,as it is of great significance for understanding continental evolution,ore deposit genesis,and mineral exploration strategies.Using compiled apatite trace element data of I-type and S-type granites around the world combined with two supervised machine learning algorithms-Support Vector Machine and Random Forest,this study established identifiers for I-type and S-type granites based on trace element characteristics of apatite.Nineteen features of apatite,including La,Ce,Pr,Nd,Sm,Eu,Gd,Tb,Dy,Ho,Er,Tm,Yb,Lu,Sr,Y,δEu,Sr/Y,and La/Yb,were used in model training,with a classification accuracy greater than 0.99.To further evaluate the classification performance of the two models,external data were also considered as an independent validation set.The classification accuracy of the independent validation set was also high.It is 0.99 for the Support Vector Machine model and 0.97 for the Random Forest model.The results confirmed that the apatite composition can be used to identify granite types effectively.To facilitate the use by researchers,nine binary diagrams for identifying granite genetic types were proposed based on a trained support vector machine model with better performance.These binary diagrams also worked well,and the classification accuracy was greater than 0.90.The machine learning methods proposed in this study provide a new method for the identification of I-type and S-type granites.This study also provides a methodological and conceptual reference for investigating granite genesis through the application of other accessory minerals.The machine learning code has been uploaded to GitHub at https://github.com/ShihuaZhong/Apatite2023MLcode.

关 键 词：磷灰石 S型花岗岩 I型花岗岩 支持向量机 随机森林 机器学习 

分 类 号：P628[天文地球—地质矿产勘探]