Integration of wave propagation physics into data-driven models for structural health monitoring in composite materials

Structural Health Monitoring (SHM) based on guided elastic waves offers a powerful framework for the detection and characterization of damage in advanced engineering structures. However, the complexity and heterogeneity of modern composite materials and structures pose significant challenges for conventional physics-based models, which are often constrained by computational costs and simplifying assumptions. On the other hand, purely data-driven approaches rely heavily on large experimental datasets, which are costly to obtain and may lack interpretability and adherence to physical laws. To address these limitations, a novel hybrid paradigm -physics-informed machine learning (PIML)-, has emerged. In PIML, physical knowledge is embedded directly into the training process by incorporating governing physical laws as constraints within the neural network’s loss function. This approach, supported by synthetically generated data from finite element simulations, enhances model fidelity, generalization, and alignment with realistic physical behaviour. The dual integration of data and physics strengthens the learning process and mitigates deviations from true physical phenomena. This work presents a hybrid framework that embeds wave propagation physics both in synthetic data generation and in the training of machine learning models. Synthetic data are generated by finite element solutions of mechanical problems, which naturally enable considering the role of various physical aspects, such as wave dispersion, dissipation and constitutive nonlinearities to name a few. As such, the proposed method significantly enhances the interpretability and generalization capabilities of the model. The ultimate goal is to advance SHM systems that enable sensitive damage detection and reliable diagnostics under diverse environmental and operational conditions.