Multimodal deep learning and genetic algorithms for adaptive optimization of injection stretch blow molding with...

In blow molding with non-conformal temperature control, changes in environmental thermal conditions can affect product quality even when machine settings remain unchanged. Conventional studies often rely on offline quality evaluation or physics-based simulations, which are computationally expensive and difficult to deploy for in-situ process adjustment. This study proposes an in-situ adaptive optimization framework that integrates multimodal deep learning with a genetic algorithm. Thermal images, lamp temperature settings, and environmental variables are used to predict the post-heating preform temperature distribution. A weighted loss function is paired with dropout regularization to improve performance in quality-sensitive regions. The estimated temperature distribution then guides the search for lamp settings that satisfy target quality constraints. Condition-level 5-fold cross-validation showed that the optimized multimodal prediction model achieved MAEs of 2.299 ± 0.245 °C for full thermal images and 1.174 ± 0.223 °C for region of interest (ROI) centerline surface-temperature profiles. The proposed optimization framework suppressed representative defect modes in the evaluated validation cases, including whitening, incomplete molding, and combined whitening-bursting. During prolonged operation, thermal drift was addressed through re-optimization with updated environmental inputs, restoring stable product quality. The computational optimization was completed within 10 s in the main validation cases, including 7.94 s for adaptive re-optimization under thermal drift, which is substantially shorter than the available process window of 115.4 s before blowing. This reported time represents the calculation time required to identify updated lamp settings and should be distinguished from the physical response time required for the heating system and preform to realize the optimized temperature field. Overall, the study indicates that multimodal temperature prediction and genetic algorithm-based optimization can support supervisory adaptive process adjustment while remaining compatible with the practical process cycle of thermally sensitive manufacturing processes.