A deep surrogate model for filling simulations in liquid composite moulding on unstructured 3D grids

Numerical filling simulations are widely used in Liquid Composite Moulding (LCM) to optimise injection strategies and reduce void formation. However, their high computational cost limits their applicability in real-time applications such as digital twins, where predictions must be delivered within milliseconds to seconds to enable effective process monitoring and control. Deep learning–based surrogate models offer a promising alternative by providing fast, approximate predictions suitable for manufacturing-relevant tasks. A key limitation of many existing approaches is their reliance on structured grids, which restricts their applicability to complex geometries and unstructured meshes typical of industrial LCM processes. In this work, we propose a unified framework that addresses these challenges. An interpolation-based grid mapping technique enables the use of state-of-the-art convolutional neural networks on unstructured meshes arising from complex boundary conditions and spatially varying flow features. Additionally, a multi-branched deep encoder–decoder architecture is introduced to predict pressure and impregnation fields in three-dimensional T-shaped stringers. The geometry is decomposed into planar regions, each processed by a dedicated CNN branch and recombined to ensure spatial consistency. The surrogate models are trained on synthetic datasets and validated against physical experiments, achieving high predictive accuracy (MAE < 1 % , 99th-percentile absolute error < 7 % ) while maintaining inference times below 0.2 s. This corresponds to a runtime reduction of four to five orders of magnitude compared to conventional simulations, enabling practical real-time digital twin implementations for LCM processes.