Convolutional Neural Network for Detection of Fluid-Filled Inclusions in Solids from Measured Elastodynamic Responses

We present a data-informed convolutional neural network (CNN) to detect fluid-filled inclusions within a two-dimensional (2D) solid domain using wave responses measured at surficial sensors. ANSYS Mechanical is utilized to simulate the propagation of elastic waves in the solid containing internal fluid inclusions. This work produces data comprising input-layer features (measured wave responses) and output-layer features (a contour map representing solid or fluid cells in the domain). The CNN is trained on synthetic data to classify each cell as solid or fluid based on the measured wave responses at sensors. Our numerical results show that the proposed CNN can effectively detect fluid inclusions of various locations, sizes, and shapes. We also present numerical experiments with respect to various parameters, including the number of sensors, the amount of training data, and frequency, and provide a comprehensive analysis of their effect on outcomes.