Digital rock physics is at the forefront of characterizing porous media, leveraging advanced tomographic imaging and numerical simulations to extract key rock properties like permeability. However, fully capturing the heterogeneity of natural rocks necessitates imaging increasingly larger sample volumes, presenting a significant challenge. Direct numerical simulations at these scales become either prohibitively expensive or computationally unfeasible due to limitations in resolution and field of view (FOV). This issue is particularly pronounced in carbonate rocks, known for their complex, multiscale pore structures, which exacerbate the resolution-FOV tradeoff. To address this, we introduce a machine learning strategy that merges multiscale imaging data from various resolutions with a 3D convolutional neural network (CNN) model. This approach is innovative in its ability to identify cross-scale correlations, thereby enabling the estimation of transport properties in larger volumes—properties that are difficult to simulate directly—using trainable proxies. The integration of multiscale imaging with deep learning allows for accurate permeability predictions at scales beyond those feasible with traditional direct simulation methods. By employing transfer learning across different scales during the training phase, our multiscale machine learning model achieves robust performance, with an R² exceeding 0.96 when evaluated on diverse lower-resolution domains with larger FOVs. Notably, this method significantly enhances computational efficiency, reducing the computational time by orders of magnitude. Originally developed for the intricate pore structures of carbonate rocks, our approach shows promise for application to a wide range of multiscale porous media, offering a viable solution to the longstanding tradeoff between imaging resolution and FOV in digital rock physics.

中文翻译：

---

通过耦合多尺度数字岩石物理和深度学习，对复杂多孔介质中的流动进行有效的计算建模：改善分辨率和视场之间的权衡


数字岩石物理学处于表征多孔介质的最前沿，利用先进的层析成像和数值模拟来提取渗透率等关键岩石特性。然而，充分捕获天然岩石的非均质性需要对越来越大的样本量进行成像，这是一个重大挑战。由于分辨率和视场 (FOV) 的限制，在这些尺度上进行直接数值模拟要么成本高昂，要么在计算上不可行。这个问题在碳酸盐岩中尤其明显，碳酸盐岩以其复杂的多尺度孔隙结构而闻名，这加剧了分辨率与视场角的权衡。为了解决这个问题，我们引入了一种机器学习策略，将各种分辨率的多尺度成像数据与 3D 卷积神经网络 (CNN) 模型合并。这种方法的创新之处在于它能够识别跨尺度相关性，从而能够使用可训练的代理来估计较大体积的传输特性（难以直接模拟的特性）。多尺度成像与深度学习的集成可以在超出传统直接模拟方法可行的尺度上进行准确的渗透率预测。通过在训练阶段采用不同尺度的迁移学习，我们的多尺度机器学习模型实现了稳健的性能，在具有较大视场的不同低分辨率域上进行评估时，R² 超过 0.96。值得注意的是，该方法显着提高了计算效率，将计算时间减少了几个数量级。 我们的方法最初是针对碳酸盐岩石的复杂孔隙结构而开发的，它有望应用于广泛的多尺度多孔介质，为数字岩石物理中成像分辨率和视场之间长期存在的权衡提供了可行的解决方案。