Classification of Digital Rocks by Machine Learning to Discover Micro-to-Macro Relationships and Quantify Their Uncertainty

多孔材料高分辨率成像的最新进展导致数字地下岩石样品的收集急剧增加，并刺激了对岩石微观结构建模和通过数值模拟计算宏观尺度输运、机械和声学特性的能力的发展。这些导致一种建模方法，提供了极大地扩展我们的材料特性数据库的潜力，而不依赖于昂贵的，或在某些情况下，不可能的，实验室测量。许多人设想，当用验证实验室测量和关注的微观尺度物理学增强时，这种方法可以扩展到发现空隙/固体的微观尺度布置与宏观尺度性质之间的预测关系，沿着量化它们的不确定性。这种方法提供了一个潜在的解决方案，许多应用程序中的关键类型的岩石的属性必须估计从几个样本。废物处理、二氧化碳储存和地下水合物勘探都是NERC职权范围内的好例子。在这些应用中，细粒岩石是关键问题，因为它们被认为是防止物质逃逸到大气层/生物圈的屏障。不幸的是，这些材料的取样成本很高，在实验室中测量也极其困难和昂贵。因此，可靠和有力地预测细粒岩石特性的能力将有助于更好地模拟宏观尺度的物理行为，评估行为的不确定性，并了解这种应用对环境和公众健康的影响。当物理学变得复杂时，微观到宏观的预测关系预计是高度非线性的。我们对3D显微图像的初步研究[1]表明，即使是单相流动性质，如渗透率，也显示出与几何和拓扑特征的强烈非线性相关性（图1）。此外，必须从大量样本中确定稳健的非线性关系，并针对新样本进行验证。机器学习（ML）提供了一个框架来执行自动化过程，其中可以以自我监督的方式从不断增长的样本集合中逐步学习非线性关系的知识。这样的过程适合这个目的，但必须由一组智能和高效的数据搜索和检索，数据分析和数据挖掘工具来支持。所有这些工具所基于的基础是根据其微观结构的不同特征以及测量和/或计算的属性对数字岩石样本进行分类的能力。该项目的目标是探索在NERC环境信息学主题内为数字岩石构建一套基于特征，内容感知和自我监督的ML分类技术的可行性。这将产生一个ML系统，能够根据预定义的控制特征对数字岩石样本和宏观尺度属性进行分类。最终，从这个试点项目中获得的知识和经验将使PI能够提出更全面的建议，以开发一套基于ML的技术，用于确定微观和宏观尺度特征之间的预测关系，并预测宏观尺度特性。有一个范围，将技术扩展到其他类型的天然多孔介质，并影响整个行业和研究界，以解决有关多孔材料的物理特性的工程和科学问题。

Recent advances in high-resolution imaging of porous materials have led to a dramatic increase in the collection of digital subsurface rock samples and have stimulated the development of a capability to model the rock microstructures and to calculate macro-scale transport, mechanical and acoustic properties by numerical simulations. These lead to a modelling approach that offers the potential for greatly expanding our database of material properties, without relying on expensive, or in some cases, impossible, laboratory measurements. It is envisaged by many that this approach, when augmented with validation lab measurements and micro-scale physics of concern, can be extended to discover predictive relationships between micro-scale arrangements of voids/solids and macro-scale properties, along with quantification of their uncertainty. This approach offers a potential solution to many applications where the properties of key types of rocks must be estimated from few samples. Waste disposal, CO2 storage and hydrate exploration in the subsurface are good examples within the NERC remit. In those applications, fine-grained rocks are of key concern, since they are assumed to function as barriers preventing substances from escaping into the atmosphere/biosphere. Unfortunately, such materials are expensive to sample and extremely difficult and costly to measure in the laboratory. Hence, an ability to predict fine-grained rock properties reliably and robustly would enable better modelling of macro-scale physical behaviours, assessment of the uncertainty of the behaviours, and understanding of the impacts of such applications to environment and public health. A micro-to-macro predictive relationship is expected to be highly non-linear when the physics becomes complex. Our preliminary investigations [1] on 3D micro images shows that even a single-phase flow property, like permeability, shows a strong non-linear correlation with the geometric and topological features (fig.1). Moreover, a robust non-linear relationship has to be identified from a large collection of samples and validated against new samples. Machine Learning (ML) provides a framework to carry out an automated process in which the knowledge of non-linear relationships can be learnt progressively from the growing collection of samples in a self-supervised manner. Such a process suits this purpose but must be underpinned by a set of smart and efficient tools for data search and retrieval, data-analysis, and data-mining. A basis on which all these tools are based is the ability to classify digital rock samples according to the diverse features of their microstructures as well as measured and/or calculated properties. The objective of this project is to explore the feasibility of constructing a suite of feature-based, content-aware and self-supervised ML classification techniques for digital rocks, within the NERC topic of environmental informatics. This will produce a ML system capable of classifying digital rock samples and macro-scale properties according to pre-defined controlling features. Ultimately, knowledge and experience gained from this pilot project will enable PIs to make fuller proposals to develop a suite of ML-based technologies for identifying predictive relationships between micro- and macro-scale features and predicting macro-scale properties. There is a scope for extending the technologies to other types of natural porous media and impacting across industries and research communities to address engineering and scientific questions about the physical properties of porous materials.