Physics-Informed Neural Networks and Sequence Encoder: Application to heating and early cooling of thermo-stamping process
Summary: arXiv:2603.26245v1 Announce Type: cross
In a previous work by Elaarabi et al. (2025b), a novel approach combining Sequence Encoder for online dynamical system identification and Physics-Informed Neural Networks (PINN) was introduced. This combination, referred to as PINN-SE, was tested on both synthetic and real data scenarios. The Sequence Encoder effectively encodes time series into feature vectors, which the PINN utilizes to map to the dynamical behavior of a system, enabling predictions of system responses under varying parameters, initial conditions (ICs), and boundary conditions (BCs).
While previous tests on real data were confined to simple 1D problems, the current study delves into a more complex and realistic scenario: the heating and early cooling stages of the thermo-stamping process. This process is crucial in forming continuous fiber-reinforced composite materials with thermoplastic polymers.
Key Findings
The investigation focuses on several critical aspects:
- Application of PINN-SE: The study explores the feasibility of applying the PINN-SE approach to the thermo-stamping process, which involves intricate thermal dynamics.
- Multimodal Data Inputs: The potential to extend the inputs of the PINN-SE to incorporate multimodal data, such as sequences of temporal 2D images, is evaluated.
- Variable Geometries: Scenarios involving variable geometries are also considered, showcasing the adaptability of the method.
Results and Implications
The results of this study indicate that combining multiple encoders with the previously proposed methods is not only feasible but beneficial. Training the model on synthetic data generated from experimental data proved instrumental in enhancing the model’s ability to generalize well to real experimental data that was not included during the training phase.
This advancement holds significant implications for industries reliant on thermo-stamping processes, such as aerospace and automotive manufacturing. By improving the predictive capabilities of the models used in these processes, manufacturers can optimize their production strategies, reduce waste, and enhance the quality of the final products.
Conclusion
Overall, the integration of Physics-Informed Neural Networks with Sequence Encoders presents a promising avenue for advancing the understanding and optimization of complex thermal processes in manufacturing. Future work will aim to refine these models further and explore additional applications in various industrial settings.