The Polymer Society of Korea

Abstract

Induction welding of fiber-reinforced polymer composites is fast and clean, but temperature uniformity is hard to achieve due to nonlinear coupling among process parameters.  Here, we develop a deep learning surrogate model to predict spatial temperature distributions in induction-heated joints and support inverse design of process parameters. A CNN is trained on validated induction-heating simulations that systematically vary coil geometry, power, frequency, and dwell time. The model accurately reconstructs temperature fields and key metrics.  Coupled with multi-objective optimization, the framework identifies process windows that reach target temperatures quickly while improving uniformity and limiting adherend overheating.  Compared with baseline settings, the optimized conditions reduce heating time and enhance joint-region temperature uniformity. The approach can be readily extended to include susceptors for additional heat-generation control.