Identification of Fluid-Filled Inclusions of Arbitrary Quantities and Shapes in Solids from Measured Elastodynamic Responses Using CNN

This study presents a data-informed convolutional neural network (CNN) for identifying fluid-filled (e.g., water) inclusions in two-dimensional (2D) solids. To generate training data, we simulate the propagation of elastic waves in structures containing fluid inclusions using Ansys Mechanical. To provide structural details, we introduce a gridwise classification method that reconstructs the targeted domain containing arbitrary fluid-filled inclusions by creating a 2D grid-structured image. This work produces training data comprising input-layer features (i.e., measured wave responses) and output-layer features (i.e., gridwise maps labeling each grid element as solid or fluid). The CNN is trained on these simulated data sets to classify grid elements based on wave responses recorded at surface sensors, and clusters of predicted fluid grid elements are used to reconstruct fluid-filled inclusions. Numerical results show that the proposed CNN can effectively detect fluid inclusions of various locations, sizes, and shapes. We also present parametric studies of factors such as the number of sensors, the size of the training data sets, and excitation frequency, providing a comprehensive analysis of their effect on detection performance.