Microstructure reconstruction of short fibre reinforced composites via transient elastic wave and Deep Learning

Reconstruction of fibre orientation fields in short fibre reinforced composites using non-destructive measurements remains a challenging inverse problem due to complex interactions between elastic waves and heterogeneous microstructures. This study investigates the feasibility of reconstructing spatially varying fibre orientations from transient elastic-wave responses using a Deep Learning framework. A physics-based finite-difference time-domain model is used to generate a large synthetic dataset mapping discretised fibre orientation fields to multi-sensor wave signals. A multilayer perceptron is trained to learn the inverse relationship between measured responses and underlying orientation distributions. Robustness and limitations are evaluated through independent finite-element validation, comparison with alternative Deep Learning architectures, identifiability analysis, dataset-size sensitivity, temporal-window selection, noise robustness, and generalisation to random fibre lengths and fibre clustering. Results demonstrate that transient elastic-wave measurements contain sufficient information to reconstruct spatial orientation fields under the simplified assumptions adopted. Reconstruction performance depends strongly on sensor configuration, temporal signal information, and the statistical characteristics of the training dataset. This proof-of-concept study establishes the fundamental relationship between wave responses and fibre orientation fields, providing a foundation for future investigations involving realistic microstructures, experimental measurements, and physics-guided inverse reconstruction strategies.