A Hyper Physics-Informed Neural Network for Predicting Heat Transfer Patterns During the Curing Process in Aerospace Composite Manufacturing

Abstract The importance of fast and accurate process simulation is crucial in various manufacturing settings to enable both design experimentation and real-time monitoring and control. Traditional numerical simulations, however, are often too slow and computationally demanding for real-time applications. Physics-Informed Neural Networks (PINN) offer a promising alternative but struggle with dynamic prediction and generalization due to the need for retraining with any change in system configurations. To overcome this, Hypernetwork-based PINN (HyperPINN) has been developed. HyperPINN utilizes a secondary neural network to generate necessary weights for the primary PINN network based on input system configurations, thus avoiding retraining for different conditions. This paper showcases the use of HyperPINN in simulating the heat transfer patterns of a curing process for aerospace-grade composites. The process involves modeling transient heat transfer in a one-dimensional composite part. The Hypernetwork and the primary PINN are jointly trained across various system configurations, allowing the Hypernetwork to dynamically adjust PINN parameters as per the current conditions while maintaining adherence to physical laws. The zero-shot capability, along with near real-time inference and accuracy comparable to conventional simulations, makes HyperPINN ideal for digital twins, design optimization, and real-time process control. Additionally, the model can be adapted for other transient physical processes in manufacturing and process industries.