Session: 11-03-01: Artificial Intelligence, Machine Learning and Data Science for Thermal Processes, Heat Transfer and Energy Systems

Paper Number: 144949

144949 - A Hyper Physics-Informed Neural Network for the Prediction of Thermochemical Curing Process in Aerospace Composites Manufacturing: Adaptive and Dynamic System Configuration-Aware Forecasting 

The significance of accurate and rapid process simulation is paramount across various manufacturing applications, particularly in the aerospace sector where composites curing is extensively used. Not only is this capability vital for design experimentation, but it is also crucial for real-time monitoring and control during the manufacturing process. Traditional numerical simulations typically have a larger computational footprint and longer inference time, rendering them unsuitable for online and real-time use cases.  

While Physics-Informed Neural Networks (PINN) have emerged as a promising alternative to traditional numerical simulations, they fall short in terms of dynamic prediction capabilities and lack generalization. Specifically, any change in the system configurations, including the initial and boundary conditions, entails retraining the PINN model. Hypernetwork-based PINN (HyperPINN) has been proposed as an alternative to address the above limitation. A hypernetwork is a neural network architecture that, given a set of input system configurations, generates the corresponding weights for another network, often referred to as the primary network, thereby eliminating the need to retrain the primary network for varying system conditions. 

This paper demonstrates the application of a hyperPINN for simulating an aerospace-grade composites curing process, addressing the challenge of generalization and dynamic prediction. The curing process of a one-dimensional composite part is considered, which involves transient heat transfer and internal heat generation due to the curing phenomenon. The Hypernetwork and the underlying PINN (primary network) are trained in a coupled manner based on a set of diverse system configurations spread over a specified range. The hypernetwork is trained to dynamically predict the parameters of the PINN for a given system configuration, while PINN on the other hand enforces the governing physical laws by embedding the governing equations in the loss function. The model’s architecture and training process enable ‘zero-shot’ prediction of composites’ thermochemical behavior under a wide range of conditions, including variations in the cure cycles (i.e., oven air temperature recipes), convective heat transfer coefficients, and dimensions of the components (e.g., composite and tooling). The zero-shot prediction capability of the HyperPINN eliminates the need to retrain or finetune the model for varying conditions. Equally important is the model’s near real-time inference capability without any loss of accuracy vis-a-vis conventional numerical simulations. This makes the model highly suitable for integration into digital twin applications, design optimization as well as real-time monitoring and control. Lastly, the proposed framework can be conveniently generalized for addressing other transient physical phenomena in manufacturing and process industries.

Presenting Author: Milad Ramezankhani Quantiphi

Presenting Author Biography: Milad Ramezankhani is a Senior Research Engineer at Quantiphi with over six years of experience developing advanced machine learning models for high-risk decision-making in engineering. He holds a Ph.D. in Mechanical Engineering from the University of British Columbia and has authored over 20 peer-reviewed papers in top-tier journals and conferences. In his current role, he leads efforts to develop state-of-the-art scientific foundation models and neural operators for complex engineering applications ranging from advanced manufacturing to aerospace.

Authors:

Anirudh Deodhar Quantiphi
Milad Ramezankhani Quantiphi
Rishi Parekh Quantiphi
Dagnachew Birru Quantiphi