An Integrated Process–Failure Simulation Framework for Predicting Composite Allowables via Multi-Fidelity Stochastic Simulation and Machine Learning

Abstract

For material screening, qualification, and certification of aerospace composites, it is essential to establish statistical design properties that account for both inherent material variability and process-induced uncertainty. Regulatory standards require the determination of A-basis and B-basis statistical values as conservative lower bound design allowables. At present, allowables are determined through extensive experimental campaigns, which are costly and time-consuming, thereby limiting the exploration of new material systems. In this work, we present a multiscale, multi-fidelity process-failure modeling framework that enables the digital prediction of allowables. The framework integrates thermo-chemical-mechanical process simulations with stochastic multi-fidelity failure simulations to capture the effects of curing-induced properties, residual stresses, and microstructural variability on mechanical performance. Machine learning surrogate models, trained via transfer learning, accelerate uncertainty propagation across scales, while Bayesian inference leverages limited experimental data to quantify uncertain parameters. High-fidelity macroscale failure simulations then predict tensile strength distributions from which A- and B-basis allowables are estimated. Application to Hexcel AS4/8552 cross-ply laminates demonstrates close agreement with test data, with conservative predictions of statistical allowables. By combining stochastic process-failure simulations with machine learning, this framework establishes a predictive pathway for estimating allowables and reducing the extent of costly experiments required during composite qualification and certification.