Data-driven reconstruction of spatially varying material properties from elastic wave signals

Abstract Accurate characterization of microstructures in inhomogeneous materials is essential for understanding and predicting their macroscopic mechanical behavior. In this work, we propose a deep learning based inverse framework for reconstructing complex, nonlinear, and realistic microstructures of inhomogeneous materials directly from elastic wave responses. Compared with previous studies based on classical machine learning classifiers, the present approach employs advanced deep neural networks capable of capturing highly nonlinear relationships between elastic wave propagation and heterogeneous microstructural features. The elastic wave response serves as the input to the model, while the corresponding microstructure is predicted as the output. Numerical simulations are used to generate training and test datasets for inhomogeneous materials with increased material complexity and spatial heterogeneity. The results demonstrate that the proposed deep learning model can accurately reconstruct realistic microstructures from elastic wave data, significantly outperforming traditional machine learning approaches in terms of both accuracy and robustness. This study highlights the potential of elastic wave driven deep learning as a powerful nondestructive inverse tool for microstructure characterization of engineering inhomogeneous materials.