FlowCastNet: A CNN-based surrogate model for the rapid prediction of flow filling patterns in VARTM processes

The manufacturing of fiber-reinforced composites by vacuum-assisted resin transfer molding (VARTM) commonly requires a distribution medium (DM) to facilitate resin impregnation. The DM significantly accelerates the filling process but must be carefully designed to ensure defect-free parts. Traditional physics-based simulations can accurately predict filling patterns, however, inherent computational cost limits their applicability for rapid decision-making in high-dimensional and large design spaces. This paper proposes an alternative surrogate modeling approach that employs a multilayer perceptron (MLP) for initial feature encoding and a convolutional neural network (CNN) as the backbone. The resulting model, FlowCastNet, was specifically developed for fast prediction of filling patterns associated with flat-panel manufacturing. A parametric study was first conducted through numerical simulation, considering several key process parameters such as preform properties, DM placement and race-tracking effects. The resulting synthetic dataset was then used to train the surrogate model and assess its predictive accuracy. Despite the significant size of the input feature domain, the model achieves satisfactory performance when evaluated on an independent dataset ( R 2 = 0 . 992 and MSE of approximately 1 0 − 3 ). Remarkably, the model provides extremely fast predictions with an inference time near 1 ms. To highlight the advantage brought by such computational speed, two multi-objective optimization scenarios were analyzed as representative examples. Compared to numerical simulation, FlowCastNet allows exhaustive search of the design space, which enables rapid identification of the Pareto frontier. While small quantitative discrepancies remain, the proposed approach shows promising potential to complement physics-based modeling for multi-criteria optimization in VARTM processes.