Physics-informed machine learning-based real-time long-horizon temperature fields prediction in metallic additive...

Area of Science:

• Materials Science
• Manufacturing Engineering
• Computational Science

Background:

• Accurate real-time temperature prediction is crucial for process control and quality assurance in wire arc additive manufacturing (WAAM).
• Traditional finite element methods (FEM) are computationally intensive, while existing data-driven models struggle with error accumulation and adaptability.
• Developing efficient and accurate predictive models for WAAM thermal behavior remains a significant challenge.

Purpose of the Study:

• To develop a physics-informed geometric recurrent neural network (PI-GRNN) for real-time, long-horizon temperature prediction in WAAM.
• To integrate geometric features and physical constraints into a deep learning framework for improved prediction accuracy and adaptability.
• To leverage transfer learning to enhance model efficiency for practical WAAM applications.

Main Methods:

• A physics-informed geometric recurrent neural network was proposed, incorporating convolutional long short-term memory (ConvLSTM) cells for spatiotemporal feature extraction.
• Physical consistency was enforced by hard-encoding initial/boundary conditions and utilizing a physics-informed loss function.
• Transfer learning techniques were applied to optimize model training efficiency.

Main Results:

• The PI-GRNN model demonstrated effective real-time temperature field prediction for future time horizons (e.g., 1.25s) using current data.
• The model achieved maximum prediction errors ranging from 4.5-13.9% on both simulation and experimental WAAM data.
• Integrating geometric and physical information reduced maximum error by approximately 1%, while the full PI-GRNN model lowered it by 4%.
• Transfer learning reduced training time by about 50% without compromising prediction accuracy.

Conclusions:

• The proposed physics-informed geometric recurrent neural network offers a computationally efficient and accurate solution for real-time temperature prediction in WAAM.
• The integration of geometric characteristics and physical laws significantly enhances model performance compared to purely data-driven approaches.
• Transfer learning provides a practical pathway for deploying advanced predictive models in industrial WAAM settings, reducing development time and costs.