Pool-ParamNet: A multi-modal fusion model for process parameter prediction in wire-laser directed energy deposition
Year: 2026
Type: article
Abstract: Wire-laser directed energy deposition (WL-DED) faces significant challenges in online process control owing to the complex coupling relationships between the process parameters and forming quality. Existing deep learning methods based on melt-pool images face a fundamental causal dilemma: melt-pool images can only be acquired after the process parameters have been applied, whereas the control objective is to determine the appropriate parameters before fabrication. To address this issue, this study investigated a novel prediction strategy based on the fusion of prior melt-pool images and conditional process parameters (CPPs). The prior melt-pool images were representative melt-pool images acquired under nominal process conditions. These served as morphological references for the desired deposition state. The CPPs provided constraints on the operating conditions, jointly narrowing the solution space for inverse mapping. A multimodal fusion model, Pool-ParamNet, was developed using a Swin Transformer for melt-pool image feature extraction and multilayer perceptron for CPP encoding. The experimental results demonstrated that Pool-ParamNet achieved mean absolute percentage error values of 0.43%, 0.36%, and 0.14% when predicting the laser power, scan speed, and wire feed speed, respectively. Ablation experiments validated the complementarity of the melt-pool images and process parameters in joint modeling. Offline validation experiments confirmed that the model output reasonable parameter adjustments for typical defects, with a single-frame inference time of approximately 12–13 ms, thus supporting the computational feasibility of future real-time deployment in WL-DED. This study developed an effective parameter-prediction approach that lays the foundation for intelligent process control in WL-DED.
Source: Additive Manufacturing Frontiers
Language: English
Subfield: Mechanical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Awards: G3051971753, G3740908404, G8524591857
OpenAlex ID: W7171826060
