Process-informed Spatial Conditioning within Data-driven Framework for High-fidelity Reconstruction of Residual Stress and Distortion Fields in Wire Arc Additive Manufacturing

Published: 1 September 2026| Version 1 | DOI: 10.17632/rv4tdtskhw.1
Contributor:
Han Yan

Description

Residual stress and distortion in wire arc additive manufacturing (WAAM) severely compromise dimensional accuracy and service performance of components. Existing deep-learning models mainly consider geometry and deposition paths while insufficiently accounting for process parameters, limiting prediction accuracy and adaptability across varying conditions. To address this issue, a Process-Conditioned Spatial Modulation Y-Net (PCSM-YNet) multimodal deep learning framework is proposed for high-fidelity reconstruction of residual stress and distortion fields. A Process-Conditioned Residual Spatial Modulation (PCRSM) module couples process-parameter mappings with path-sensitive weights to modulate deep features; a Thermo-Mechanical Response-Guided Cross-Level Feature Fusion (TMR-CFF) module selectively reorganizes encoder features during decoding; and a Process-Conditioned Spatial Reweighting (PCSR) module enhances critical local responses in skip connections. PCSM-YNet achieves R² above 0.9 for all prediction tasks and captures stress–distortion trends under varying heat-source powers and deposition speeds. Compared with 3D U-Net and 3D ResUNet, this model maintains more consistent predictive accuracy across varying process conditions, reducing the maximum RMSE of Y-direction distortion by approximately 25% and 16%, respectively, and the peak absolute error of residual von Mises stress by up to approximately 25% and 20%. The framework reduces reliance on repeated experiments and costly simulations, supporting rapid WAAM process evaluation, optimization, and quality control.

Files

Institutions

Categories

Mechanical Engineering, Multimodal Deep Learning

Licence