Machine learning-enhanced modelling and experimental analysis of foam-core thermoplastic composites produced via pultrusion

Foam-core thermoplastic composites manufactured by pultrusion offer lightweight, recyclable structural solutions but require precise control of coupled thermal and curing phenomena to ensure uniform properties. While physics-based models can capture these thermochemical interactions, their computational cost limits their use for rapid prediction and process optimisation. This study presents an integrated experimental–numerical–machine learning framework for foam-core thermoplastic pultrusion using Elium® resin. Cure kinetics are characterised by DSC and incorporated into a validated 3D multiphysics model coupling heat transfer and polymerisation. Microscopy confirms limited resin penetration into the foam surface, forming a mechanical interlocking mechanism at the skin–core interface. A large parametric simulation campaign is used to train machine-learning surrogate models (neural networks, random forests, and gradient boosting), achieving 𝑅² > 0.998 and enabling millisecond-level predictions with over 10⁴× speed-up compared to finiteelement simulations. These surrogates are employed for rapid prediction and process optimisation to identify operating windows that balance throughput, thermal control, energy efficiency, and complete curing.