A Framework for Bayesian Calibration of Fiber-Reinforced Composite Material Properties

Abstract

Fiber-reinforced composite materials continue to rapidly improve in terms of structural performance. These materials offer the promise of significant weight reduction in many products—particularly aerospace structures, where weight minimization is critical to both cost and performance. However, the introduction of new materials is a slow and expensive process that requires dependable material property data early in the design cycle, and it takes a rigorous, comprehensive testing program to fully characterize the properties first at the ply level, then at the laminate level, and finally at the product or component level. The ability to reliably predict laminate material properties using limited ply-level material property data early on— and an understanding of the accuracy of those predictions—can dramatically shorten the design cycle. We propose a data-driven approach to obtain material properties for newly developed composite materials. The proposed approach uses limited ply-level property test data to generate finite element model (FEM) parameters that best fit the statistical distributions from test data. The material model correlation process uses neural networks as surrogates for high-fidelity FEMs. The surrogate model is then used to perform Bayesian calibration to directly find the parameters that produce a best fit for a given set of test data. At the end of the process, statistical distributions of all the parameters of interest are provided. This process provides a tractable and systematic approach for the selection of FEM material parameters for a novel material with limited test data. In recent work, the approach has been validated using test data derived from a composite material typically used in space applications. This technology could potentially enable accelerated development of new materials, and it could be extended to other kinds of materials beyond composites—especially those materials that exhibit high levels of statistical variability.