Deep learning surrogates for resin impregnation of 3D preforms with complex geometry and non-uniform material properties

The production of large, integrated aerospace structures employing Resin Transfer Moulding (RTM) is facilitated by high-fidelity simulations to predict resin flow behaviour in the presence of race tracking and permeability variations. Conventional physics-based RTM simulations are computationally expensive, which constrains real-time monitoring and digital twin deployment. Advances in machine learning and data assimilation enable accurate and fast surrogate models to aid the creation of digital twins for composite manufacturing. This paper presents a Recurrent Neural Network (RNN) surrogate for RTM simulation applied to a realistic 3D RTM geometry. The RNN surrogate approximates pressure–time curves more accurately than surrogates with alternative architectures and achieved a mean error in predicted fluid pressure below 400 Pa. The surrogate is employed within a Bayesian inversion framework to infer local reinforcement permeabilities from synthetic pressure data. Owing to the fast surrogate, updating local permeabilities based on pressure data required only 0.77 s on average. The presented surrogate and inversion framework provide the speed and accuracy required for near real-time applications such as accelerated post-process non-destructive evaluation and active process control.