Abstract Thin sections provide geoscientists with a wealth of information about composition and diagenetic history of sedimentary rocks. From a practical perspective, the quantity of detrital clay minerals or percentage of porosity can play a large role in the quality of a reservoir. However, the quantitative analysis of thin sections often requires many hours of manual labor, which limits the number of samples a single person can analyze in a reasonable time frame. Here we apply a supervised machine-learning method that requires only traced grains as inputs, which eliminates the need for an expert to hand design input features. We also present a data-augmentation method to reduce the amount of tracing required. The traced grains form a multi-channel input that takes into account plane- and cross-polarized images, and a segmented image is output. Using a simplified grain categorization (quartz-feldspar-rock fragments-dense minerals) the statistical error for results on grain composition is comparable to a point count with 350 points. Once the model is trained, it can be applied quickly to additional images. In addition to providing component percentages, a segmented thin section can be used further to describe the morphology of grains (e.g., angularity, ellipticity) or serve as the basis for digital rock-physics experiments. This test of supervised machine learning does not reproduce the level of detailed component identification that is typical of manual point-counting, but it provides a clear indication that a diverse and fully representative data set will be required to achieve automated component identification that is both accurate and precise.

中文翻译：

---

用于薄截面中砂粒的点计数和分割的机器学习

摘要 薄片为地球科学家提供了大量关于沉积岩成分和成岩历史的信息。从实践的角度来看，碎屑粘土矿物的数量或孔隙度的百分比对储层的质量起着很大的作用。然而，薄片的定量分析往往需要很多小时的体力劳动，这限制了一个人在合理的时间范围内可以分析的样本数量。在这里，我们应用了一种监督机器学习方法，该方法只需要跟踪的谷物作为输入，从而无需专家手动设计输入特征。我们还提出了一种数据增强方法来减少所需的跟踪量。跟踪的颗粒形成考虑平面和交叉偏振图像的多通道输入，并输出分割图像。使用简化的颗粒分类（石英-长石-岩石碎片-致密矿物），颗粒成分结果的统计误差与 350 点的点数相当。一旦模型经过训练，就可以快速应用于其他图像。除了提供组分百分比外，分段薄片还可进一步用于描述晶粒的形态（如棱角、椭圆度）或作为数字岩石物理实验的基础。这种受监督的机器学习测试没有重现手动点计数典型的详细组件识别水平，但它清楚地表明需要多样化和完全具有代表性的数据集才能实现既准确的自动化组件识别和精确。