Adaptive sampling-based surrogate modeling for composite performance prediction

• Adaptive sampling reduces the experimental and simulation effort by efficiently sampling the high-dimensional design space. • A Gaussian Process Regression surrogate model is trained to guide the adaptive sampling process. • The trained surrogate model can accurately predict the woven composites’ stress and failure in presence of notches. Composite structures experience varying strain states across different locations during loading due to the inherent material inhomogeneities. These strain states can be accurately captured using Digital Image Correlation (DIC)-based strain maps and incorporated into composite modeling through Representative Volume Element (RVE) analysis, which is expensive. In this paper, we develop an adaptive sampling-based surrogate model approach to efficiently map the high-dimensional experimental strain-states to stress-states (obtained through limited physics-based RVE analysis) for a woven composite structure. In this approach, a Gaussian Process Regression (GPR) model is used to guide the adaptive sampling process and build the surrogate model that can predict the stress response of any composite sample within the design space. We used this trained model to predict the stress and failure of a woven composite samples with holes and compared with experiment. Our findings demonstrate that adaptive sampling reduced the number of experiments and simulations significantly while still predict to failure and stress effectively for a composite part. This combined experimental, numerical, and data-driven approach can accelerate the design process for architectural composite materials and enhance their performance.