Data-driven viscoelasticity multiscale method: Integrating deep neural networks and proper orthogonal decomposition in the frequency domain

초록

Viscoelastic materials, which are known for their combined viscous and elastic properties, exhibit complex mechanical behaviors that are challenging to analyze. This study introduces a multiscale analysis technique using data-driven computational mechanics to investigate the behavior of viscoelastic composites in the frequency domain. Utilizing finite element method (FEM) data, proper orthogonal decomposition was first applied for dimensionality reduction. Subsequently, a deep neural network was developed to predict the stress fields at microscale integration points while simultaneously analyzing the macroscale structures. This approach facilitates the direct resolution of mechanical behaviors without relying on traditional material models, thereby reducing issues such as model bias and information loss. This data-driven viscoelasticity multiscale method enables simulations in the frequency domain, thereby advancing the understanding of viscoelastic material behavior across various scales. This is one of the first attempts to apply multiscale analysis in the frequency domain. The findings of this study demonstrate the significant benefits of combining machine learning techniques with computational mechanics, highlighting the efficiency and accuracy of the proposed data-driven viscoelastic multiscale method in handling real-world complexities and uncertainties in composite material research.

키워드