Machine learning extraction of viscoelastic material properties from full-field deformation measurements

Characterizing the mechanical properties of viscoelastic materials is crucial in various engineering disciplines, from structural mechanics to biomedical applications. Traditional characterization approaches often rely on analytical or numerical models and iterative processes that are computationally expensive and labor-intensive. In this study, we propose two novel deep learning-based frameworks designed to efficiently extract viscoelastic material properties from diverse data sources, including analytical methods, numerical simulation results, as well as experimental observations. The first algorithm, Deep learning V iscoelastic M aterial modeling with known F ield governing E quations (or Deep-VM-FE), is designed for problems with known analytical models. For Kelvin-Voigt viscoelastic material models, it decouples stress fields into rate-independent and rate-dependent components, which are learned by two parallel neural network channels. A comprehensive loss function ensures accurate total stress predictions. This approach was successfully validated on analytically generated databases for various hydrogel materials, including polyacrylamide, agarose, and gelatin. For more general applications where explicit governing equations are unavailable, we propose a second algorithm, called Deep learning V iscoelastic M aterial modeling with F ield D ata (or Deep-VM-FD). This method learns stress evolution directly from point-based, rate-dependent field data. A novel gradient pooling technique is incorporated to integrate information from neighboring points, enriching the model's spatial awareness. The Deep-VM-FD algorithm was extensively validated using numerical data from 1D creep and relaxation tests, as well as 2D cavitation and single-point loading simulations. Furthermore, its practical utility was demonstrated by successfully analyzing experimental results from bubble cavitation tests. Across all validation cases, the predictions showed excellent agreement with ground truth data. In summary, we demonstrate that our developed two deep learning-based approaches offer a powerful and efficient alternative for viscoelastic material characterization, capable of handling diverse data types and complex physical scenarios.