PIML-based Real-time Long-horizon temperature Fields prediction in metallic additive manufacturing

The real-time long-horizon temperature field prediction during the Wire Arc Additive Manufacturing (WAAM) process is crucial for controlling heat accumulation, optimizing process parameters, and ensuring the quality of the manufactured parts. However, due to the time-consuming nature of finite element methods (FEM) and the process-informed neglect by existing data-driven models, real-time long-horizon temperature field prediction remains a technical challenge for control systems in metallic AM. Despite the interpretability offered by Physics-Informed Machine Learning (PIML), which still suffers from relatively long training times and a lack of long-horizon prediction capabilities for dynamic systems. To address this issue, the study proposes a Physics-informed Geometric Recurrent Neural Network (PIGeoRNN) model that fuses geometric characteristic and physical information through an encoder, extracts spatiotemporal characteristic using convolutional long short-term memory cells, and ensures that the prediction results conform to physical laws by incorporating hard-encoding initial/boundary conditions and a PI loss function. The model is capable of predicting the temperature field for the future 1.25 s based on data from the current 1.25 s, and it has also been evaluated for more long-horizon predictions. Transfer learning was used for enhance the model's efficiency in practical applications. Results demonstrate that the proposed PIGeoRNN model achieves a maximum prediction error of only 4.5% to 13.9% for both simulation and experiment. The inclusion of geometric features and physical information reduces the maximum error by approximately 1%, while the integrated model can lower the maximum error by 4%. Furthermore, transfer learning training reduces the model's training time by approximately 50% while achieving the same loss level. These findings indicate that the proposed method holds significant potential for process feed-forward control and digital twin in WAAM. It also represents an important first step toward real-time long-horizon physical fields prediction in metallic AM.