镁铁质-超镁铁质岩浆岩中单斜辉石的智能分析研究

Intelligent analysis of clinopyroxene in mafic-uhramafic magmatite

区分不同构造环境中岩浆作用是认识地幔中岩浆形成过程的基本手段。目前较为成熟的是利用全岩去区分判别，而利用造岩矿物去判别构造环境、演绎岩浆演化的研究还不够深入。本文运用机器学习的方法，以全球新生代大洋中脊、洋岛以及岛弧构造背景中的镁铁质一超镁铁质岩浆岩中单斜辉石的地球化学数据为研究对象，试图区分这3种不同构造环境的单斜辉石。通过机器学习方法中K-邻近（K—Nearest Neighbor，KNN）和随机森林（Random Forest，RF）的计算和比较，认为RF是一种有效的地球化学区分方法，它的结果不仅可用来判别构造环境，同时还能够提取特征元素。同时我们发现，在镁铁质一超镁铁质岩浆岩单斜辉石构造环境判别图解中，Rb、La、Ba、Cr、Sr、Yb、V、Ti、Nd、Eu、Gd等微量元素具有较高的贡献率，而主量元素贡献率较低。在此基础上，我们结合前人的对单斜辉石的构造环境判别图的研究成果，提出几个判别效果较好的判别图解。但是整个研究由于缺少进一步可视化的成果，限制了机器学习方法的推广，这也是今后需要进一步研究的课题。

Differentiating the magmatism in different tectonic environments is the basic means to understand the magma formation process in the mantle. At present, it is more mature to use whole rocks to distinguish and discriminate. While, the use of rock-forming minerals to discriminate the tectonic environment and deduce magmatic evolution is still insufficient. In this paper, we use the geochemical data of clinopyroxene in the mafic-ultramafic magmatite, which is in oceanic mid-ocean ridge, oceanic island, and island-arc tectonic setting in the global Cenozoic as research object, trying to distinguish the three kinds of clinopyroxene in different tectonic environments with machine learning methods. By the calculation and comparison of K-Nearest Neighbor （KNN） and Random Forest （RF） in the machine learning method, it is considered that RF is an effective geochemical discrimination method, and its results can be used not only to discriminate tectonic environment, but also to extract characteristic elements. At the same time, we found that the trace elements such as Rb, La, Ba, Cr, Sr, Yb, V, Ti, Nd, Eu, Gd have a higher contribution rate in the discrimination of clinopyroxene from mafic-ultramafic magmatite, and the contribution rate of the major elements is low. Based on this, we propose several discriminative diagrams with better discriminant effect combining with previous research results. However, the lack of visualization results in the whole research limits the popularization of machine learning method. This is also a subject that we need to study further in the future.