Answers are grounded in this paper and cite their sources.
Impact-induced subsurface damage is challenging to characterize because transient wave measurements contain spatially distributed and strongly coupled information associated with damage location, extent, and intensity. This study presents a physics-based wave–deep-learning framework for simultaneous three-dimensional reconstruction and quantitative characterization of impact damage. Random surface impacts with varying impactor properties and contact locations were applied to a homogeneous solid. Rather than prescribing artificial damage geometries, a Hertzian-contact-based surrogate was used to map the impact conditions to continuous subsurface damage fields. Transient elastic-wave responses were then simulated using distributed excitation and sensing, and differential signals between the damaged and intact states were used as inputs to the inverse model. A multitask convolutional neural network was developed with a receiver–time feature encoder, a volumetric damage decoder, and separate regression branches for damage volume and severity. To address the strong spatial sparsity of subsurface damage, the decoder explicitly separated damage occupancy from continuous damage intensity. The proposed framework successfully reconstructed damage location, spatial extent, morphology, and intensity distribution while simultaneously estimating damage volume and severity. Systematic studies of training-data size, model components, damage characteristics, impact surface, measurement noise, receiver loss, wave-speed mismatch, and damage-identification threshold further demonstrated the robustness and sensitivity characteristics of the learned inverse mapping. The results demonstrate the potential of combining physically conditioned data generation, transient elastic-wave measurements, and multitask deep learning for volumetric nondestructive characterization of subsurface impact damage. Additional frozen-model tests with alternative numerical damage surrogates revealed nonuniform out-of-distribution performance, emphasizing the importance of model-form uncertainty and independent physical validation. Independent baseline comparisons showed competitive spatial and damage-volume accuracy relative to alternative convolutional architectures, with some task-dependent trade-offs in severity and localization.
No takes yet. Share an insight, caveat, or question.
Sang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: