Abstract A procedure combining experiments and deep learning is demonstrated to acquire pore-scale images of oil- and water-wet surfaces over a large field of view in microfluidic devices and to classify wettability based upon these pore scale images. Deep learning supplants the manual, time-consuming, error-prone investigation and categorization of such images. Image datasets were obtained by visualizing the distribution of immiscible phases (n-decane and water) within in-house fabricated micromodels containing sandstone-type and carbonate-type pore structures. The reference dataset consists of 6400 color images binned into four classes for sandstone (water- or oil-wet surfaces) and carbonate (water- or oil-wet surfaces) pore-network patterns. There are 1600 images per class. During 10 sequential training and testing runs of the deep-learning algorithm, 3000, 100, and 100 images were randomly assigned per each rock pattern as the training, validation, and test sets, respectively. We trained and optimized both a Fully Connected Neural Network (FCN) and Convolutional Neural Network (ConvNet) using the image data. The ConvNet performs better as 5 and 8 layers are implemented, as expected. The FCN shows an average test set accuracy for binary surface wettability classification of 87.4% for sandstone rock type and 98.7% for carbonate rock type pore networks. Distinctive heterogeneity in the carbonate rock type and its relevant phase saturation profile resulted in a better prediction accuracy. The best ConvNet models shows an average test set accuracy of binary surface wettability classification of 99.4 ± 0.1% for both sandstone-type and carbonate pore networks. Heterogenous pore sizes and an abundance of small pores amplify the effects of wetting and aid identification. Overall, the test set accuracy for the simultaneous classification of four classes including both sandstone (water- or oil-wet) and carbonate rock pattern (water- or oil-wet) is 98.5% with an 8-layer ConvNet. Performance of the deep-learning model is further interpreted using saliency maps that indicate the degree to which each pixel in the image affects the classification score. Pixels at and adjacent to interfaces are most important to classification.

中文翻译：

---

用于自动表征孔隙尺度润湿性的深度学习

摘要 展示了一种结合实验和深度学习的程序，可以在微流体装置的大视场内获取油和水润湿表面的孔隙尺度图像，并根据这些孔隙尺度图像对润湿性进行分类。深度学习取代了对此类图像的手动、耗时、容易出错的调查和分类。图像数据集是通过在包含砂岩型和碳酸盐型孔隙结构的内部制造的微模型中可视化不混溶相（正癸烷和水）的分布而获得的。参考数据集由 6400 张彩色图像组成，分为砂岩（水或油湿表面）和碳酸盐（水或油湿表面）孔隙网络模式的四类。每个类有 1600 张图像。在深度学习算法的 10 次连续训练和测试运行期间，每个岩石模式分别随机分配 3000、100 和 100 张图像作为训练、验证和测试集。我们使用图像数据训练和优化了全连接神经网络 (FCN) 和卷积神经网络 (ConvNet)。正如预期的那样，随着 5 层和 8 层的实现，ConvNet 的性能更好。FCN 显示二元表面润湿性分类的平均测试集精度对于砂岩类型为 87.4%，对于碳酸盐岩类型孔隙网络为 98.7%。碳酸盐岩类型的独特非均质性及其相关的相饱和度剖面导致了更好的预测精度。最好的 ConvNet 模型显示二元表面润湿性分类的平均测试集精度为 99.4 ± 0。砂岩类型和碳酸盐孔隙网络均为 1%。不均匀的孔径和大量的小孔放大了润湿和帮助识别的效果。总体而言，使用 8 层 ConvNet 对包括砂岩（水或油湿）和碳酸盐岩模式（水或油湿）在内的四类同时分类的测试集精度为 98.5%。使用显着图进一步解释深度学习模型的性能，显着图指示图像中每个像素对分类分数的影响程度。界面处和邻近界面的像素对分类最重要。使用 8 层 ConvNet，同时分类包括砂岩（水湿或油湿）和碳酸盐岩模式（水或油湿）在内的四类的测试集精度为 98.5%。使用显着图进一步解释深度学习模型的性能，显着图指示图像中每个像素对分类分数的影响程度。界面处和邻近界面的像素对分类最重要。使用 8 层 ConvNet，同时分类包括砂岩（水湿或油湿）和碳酸盐岩模式（水或油湿）在内的四类的测试集精度为 98.5%。使用显着图进一步解释深度学习模型的性能，显着图指示图像中每个像素对分类分数的影响程度。界面处和邻近界面的像素对分类最重要。