Prediction of compressional wave velocity of cement-reinforced soil from core images using a convolutional neural network regression model

This study aims to predict the compressional wave velocity (V-p) from the photographic images of cylindrically cored cement-reinforced soils using a convolutional neural network (CNN) model. The experimentally measured V-p values and corresponding surficial core images were subjected to the CNN regression model based on a backbone network pre-trained by transfer learning. The model was retrained by fine-tuning and optimized with regularization strategies and data augmentation. The results showed that the trained network model reliably predicted V(p)( )with reasonable performance of R-2 = 0.78. Three-dimensional X-ray computed tomographic imaging of both overestimated and underestimated specimens revealed that surficial core images did not sufficiently reflect the internal structures. The slightly scattered prediction seemed attributed to the insufficient dataset size and invisible internal structure. Nevertheless, the proposed approach allowed not only estimating V-p at unmeasured spots in cores based on core images by fully leveraging artificial intelligence but also obtaining consecutive V-p profiles. Furthermore, this study established the hardly seen correlation between core image and V-p by the proposed CNN regression model and can be extended to estimation of other geophysical and geomechanical properties to construct a sufficient dataset for subsurface geostatistical modeling.

Automated summary

This study predicts the compressional wave velocity (V-p) of cylindrically cored cement-reinforced soils using a convolutional neural network (CNN) model. The trained network reliably predicts V(p) with reasonable performance, and it can also obtain consecutive V(p) profiles.