Accelerating curing simulations of composites using Graph Neural Networks

Abstract

In thermoset composite manufacturing, process simulation is essential for identifying optimal cure cycles to avoid temperature overshoots, non-uniform curing, or overly conservative cure cycles. Deep Learning based techniques have emerged as promising tools to accelerate modelling compared to conventional numerical methods. This work presents a Graph Neural Network with a MeshGraphNet style architecture applied to the 2D thermochemical modelling of the curing of carbon-fibre reinforced composites. The results of the model in predicting the transient evolution of temperature and degree of cure for rectangular and L-shaped cross-sections under two-dwell cure cycles are presented, showing that the predictions agree closely with finite element results. The areas where the model’s predictions struggle are identified. When coupled with a genetic algorithm for cure cycle optimisation to reduce peak exothermic temperature and cure time, the surrogate reduces runtime by about 96.3% compared to finite element simulations. Additionally, the model’s flexibility to anisotropic conductivity, as well as other geometries such as T- and U-shapes when such variability is included in the training, is presented. The limitations regarding transferability to other boundary conditions are also identified while recognising directions for future development towards more generalisable surrogates for manufacturing processes.