Scientific Machine Learning
PINNs and PINOs for DED process prediction
We have developed Scientific Machine Learning-based models for the prediction of thermal histories in Directed Energy Deposition (DED) processes, namely Physics-Informed Neural Networks (PINNs), eXtended Physics-Informed Neural Networks (XPINNs) and Physics-Informed Neural Operators (PINOs). These models were used to predict multi-layer thermal fields while preserving the underlying heat transfer physics and reducing the computational cost compared with conventional finite element simulations. The initial PINN framework demonstrated the feasibility of accurate meshless thermal prediction, while the XPINN implementation improved scalability and robustness through domain decomposition. More recently, the PINO framework enabled one-shot, generalisable prediction under varying geometries, toolpaths, material properties and process parameters without retraining, facilitating rapid thermal simulation and optimisation for changing manufacturing scenarios.
