3D Convolution Lightweight Vision Transformer to Progressive Semantic Focusing for Magnetic Property Prediction of Additively Manufactured Components

Objective/background: Predicting the magnetic performance of selective laser melting (SLM)-fabricated components is essential for quality assurance; however, the complex nonlinear dynamics of Fe-50Ni soft magnetic alloys under high-frequency excitation (400–800 Hz) remain poorly modeled by existing deep learning approaches. This study aims to overcome the significant accuracy degradation observed in conventional models when addressing these frequency-dependent hysteresis behaviors. Method We propose a lightweight conv-enhanced progressive sampling vision transformer (CVPSViT), which can synergize 3D spatial feature extraction with physics-informed process parameters. The architecture incorporates three methodological innovations: (1) it ingests stacked layer-wise imagery as 3D volumes to capture interlayer continuity and microdepth textures; (2) it introduces a conv-enhanced progressive sampling module (CPSM), which employs a coarse-to-fine strategy to dynamically update sampling coordinates, focusing attention on semantically discriminative regions akin to the human visual system; (3) it executes a deep cross-modal fusion by embedding critical manufacturing parameters, specifically laser power and oxygen concentration, directly into the global representation prior to inference. Results Extensive experiments on five key magnetic targets demonstrate that CVPSViT consistently outperforms conventional machine learning baselines and the standard CvT architecture. The model exhibits exceptional robustness in high-frequency scenarios: for coercivity (\(\:{H}_{c}\)) at 800 Hz, it achieves an\(\:\:{R}^{2}\) of 0.981, significantly surpassing the 0.876 of CvT. Furthermore, for iron loss \(\:\left({P}_{cv}\right)\), the most frequency-sensitive indicator, CVPSViT maintains a high accuracy of 0.934 compared to 0.909 for CvT. Ablation studies confirm high efficiency, with the model requiring 24% fewer parameters (38.5M) and 12% fewer GFLOPs (22.64) than the PSViT baseline. Conclusions This work presents a robust, computationally efficient framework for the real-time quality monitoring of additively manufactured components. By effectively balancing high-frequency prediction accuracy with low model complexity, CVPSViT offers a viable solution for intelligent manufacturing systems requiring precise feedback on material properties.