Real‐Time Spatiotemporal Prediction of Flow‐Front Evolution in Vacuum‐Assisted Resin Infusion via ConvLSTM

ABSTRACT

Real‐time flow‐front tracking in vacuum‐assisted resin infusion (VARI) is critical for mitigating composite manufacturing defects. However, traditional numerical solvers suffer from prohibitive computational latency. This study presents a deep learning‐based spatiotemporal framework for faster‐than‐real‐time flow‐front prediction. Training datasets are generated via a parallelized control‐volume finite element method (CVFEM) integrated with Karhunen‐Loève (KL) expansions to capture stochastic preform permeability fields. A Convolutional Long Short‐Term Memory (ConvLSTM) network is developed to autoregressively predict two‐dimensional transient fluid advancement while strictly preserving spatial topologies. To evaluate industrial deployment, an in situ visual‐predictive closed loop is established by deploying a YOLOv11‐seg instance segmentation model on real manufacturing video feeds. Results demonstrate that the integrated network executes a 10‐step future spatiotemporal rollout within single‐digit millisecond computational latency, maintaining an Intersection over Union (IoU) of 95.33% at Step 10 on physical experimental data. This paradigm provides a robust data‐driven strategy for proactive quality control and online digital‐twin implementations in liquid composite molding.