Prediction of 3D as-deposited layer geometry with sequential SDF-Based spatio-temporal networks in wire arc additive manufacturing

Wire arc additive manufacturing (WAAM) enables efficient fabrication of large-scale metallic components. However, the geometry of each as-deposited layer (ADL) is highly sensitive to coupled thermofluid dynamics and process fluctuations. These complex interactions make it difficult to maintain geometric stability, underscoring the need for accurate layer-wise profile predictions to support process planning, dimensional control, and defect prevention. This study presents ILPP-Net, an inter-layer profile prediction network that leverages a sequential signed distance field (SDF) representation to model the spatio-temporal evolution of multi-layer WAAM deposition. The point cloud of each layer is converted into a sequence of SDF cross-sections, which are combined with the process metadata, including global position, wire-feeding speed, deposition speed, arc current, and lateral torch offset, to predict the next-layer contour. A temporal context block (TCB) integrates geometric and process information through parametric fusion and bidirectional long short-term memory (BiLSTM), whereas a FiLM-modulated U-Net performs spatial regression to reconstruct a continuous SDF distribution. A WAAM system integrated with a 3D vision module was developed to acquire real layer datasets for training and evaluation. ILPP-Net achieves high prediction fidelity across different deposition strategies. For early thermally unstable layers (Layer 2 - 3), the model achieves an mloU of 0.85 - 0.86 with root mean square errors (RMSEs) of 0.192 - 0.187 mm, while performance improves for the more stabilized upper layers (Layers 7–10) to 0.89–0.90 mloU and 0.165–0.141 mm RMSE. Ablation studies show that TCB contributes to an accuracy gain of ≈14.36%, highlighting the importance of temporal modeling. The proposed model was further deployed in a prediction-driven feed-forward compensation experiment, where the ILPP-based parameter adjustment significantly improved the dimensional accuracy. Relative to the open-loop deposition, the compensation strategy reduced the final height RMSE by 56.57% and the mean error by 62.4%, demonstrating the effectiveness of the model for geometry-aware adaptive control.