A physics‑guided Bayesian‑weighted loss‑function framework for robust geometry‑to‑process inversion in wire arc additive manufacturing
Аннотация
Industrial application of wire arc additive manufacturing (WAAM) requires a digital process chain analogous to computer-aided manufacturing (CAM), yet its core bottleneck is the absence of a digital process kernel capable of robustly inferring process parameters from target geometries. The mapping between WAAM parameters and deposited-bead geometry is highly stochastic and nonunique, while the complexity and incomplete understanding of WAAM physics, together with limited industrial data, hinder the construction of universal constraints based on physical-equation residuals and the training of complex generative models. Here, we propose a physics-guided Bayesian weighting and model-guided loss-function framework, termed BWMG. Through physics-guided neighborhood augmentation, BWMG represents difficult-to-quantify process perturbations as an explicit data distribution and combines a Gaussian neighborhood prior with Bayesian weighting to convert noisy-label fitting into posterior inference. Unlike generative approaches that reconstruct complete multimodal posteriors, BWMG formulates multi-solution inversion as a decision-oriented Bayesian point estimation problem, yielding representative process solutions with high posterior credibility and local robustness. Experiments on 5083 aluminum-alloy and Q235 steel datasets show that BWMG reduces mean error by 29.2 and 25.5% and error standard deviation by 42.0 and 27.4%, respectively, relative to the baseline model, while maintaining low error fluctuations under data distribution shifts. Mechanistic analysis demonstrates that physics-guided gradient and curvature regulation improves inversion stability under limited data. Forming cases involving acute-angle filling, variable-curvature structures, and cavity further validate its potential for process-parameter decisions in path planning and geometric fidelity, providing a new route toward a digital process kernel for high-fidelity WAAM manufacturing.