Sampling by Exploration and Replication for Estimating Experimental Strength of Composite Structures

Abstract

The industry has to resort to experiments for practical design of composite laminates when physical models/simulations are inadequate for desirable accuracy or require excessive computational resources. Surrogates are used to predict strength of composite laminates in the design space by conducting an array of tests. It has been reported that an exploration strategy to test as many different configurations as possible is more effective than replication of fewer points for reducing test noise. The observation was based on analytical test functions and synthetic Gaussian noise. This paper first studies real experiments to check whether the previous observation stands. Test results of open-hole tension for composite plates that included 18 replicates per configuration were examined. A resampling procedure was developed that compared exploration and replication, and it was found that exploration proved to be more accurate for prediction than replication. Second, the major source of uncertainty for surrogate prediction was examined, which is variability of strength. The distribution of experimental open-hole tension strength was found to be not unambiguously independent and was identically distributed normal distribution as commonly assumed. The variation of specimen strength is correlated rather than independent at different configurations due to the between-batch variability. Consequently, the influence of distribution type was then investigated on an analytical test function with three synthetic distributions. The exploration strategy proved to be better than the replication strategy for all three distributions. It was found that the exploration strategy allows for higher-order polynomial surrogate to be used, which is a key point to improve characterization of a function with complex dependence on design parameters.