An Optimization Strategy Based on Machine Learning and LayerWise Models to Minimize Process-Induced Deformations in CFRP Parts

Virtual manufacturing of composites has become increasingly important over the last two decades as a tool to improve the quality and sustainability of composite structures. However, the manufacturing process simulation is complex and multidisciplinary and may require high computational overheads. This paper aims to the virtual curing of composite parts and use machine learning techniques to optimize the stacking sequence and minimize process-induced deformations. Such a methodology combines 1D finite element (FE) models, experimental testing, and machine learning models. The FE model exploits higher-order layer-wise theories to compute accurate through-the-thickness distributions of shear and peeling stresses. 1D elements lead to better efficiency by requiring a fraction of the computational cost usually needed by 3D finite element models. DSC and DMA tests characterize relevant material properties, e.g., the degree of cure, viscoelastic moduli, and free strains. Furthermore, a cure-hardening instantaneously linear elastic (CHILE) constitutive model is adopted. The FE model allows the rapid evaluation of residual stresses, spring-in angles, and warpage. Due to its numerical efficiency, various lay-up combinations can be evaluated, and the results can feed a dataset to train an Artificial Intelligence system. This study uses Gaussian process regression (GPR) to fit probabilistic response surfaces to numerical deformation predictions, explore the design space, and find optimal lay-ups minimizing defects. Furthermore, mitigation strategies are developed using specific lay-ups around geometry transition points such as sharp corners. Finally, optimal lay-ups are validated by fabricating L-shaped parts with similar conditions.