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

Real-time long-horizon temperature prediction in wire arc additive manufacturing is critical for process control and quality assurance. However, finite element methods are computationally expensive, and the existing data-driven models suffer from error accumulation and poor adaptability. Here we propose a physics-informed geometric recurrent neural network that integrates geometric characteristics and physical constraints, captures spatiotemporal characteristics via convolutional long short-term memory cells, and enforces physical consistency through hard-encoding initial/boundary conditions and physics-informed loss function. The model can predict the temperature field for future 1.25 s based on current 1.25 s data, and has also been evaluated for more long-horizon predictions. Transfer learning was used to enhance the model's efficiency in practical applications. Results demonstrate that the proposed model achieves 4.5-13.9% maximum prediction error in simulations and experimental data. Including geometric characteristics and physical information reduces maximum error by about 1%, while the integrated model lowers it by 4%. Furthermore, transfer learning reduces the training time by approximately 50% while achieving the same loss level.