A domain-adaptive deep learning strategy for cross-material recognition of laser textured surface microstructures for adhesion enhancement

Annals of Forest Science · Springer Nature · 2026-10-27 · DOI: 10.1007/s10845-026-03001-3

Аннотация

Laser texturing is a non-contact, flexible and eco-friendly pretreatment without the application of toxic chemicals. As the depth of laser texturing increases, the geometry of microgrooves formed on the metal surface transitions from an ‘open’ type to an ‘interlock’ type, the latter improving the fracture toughness of adhesively bonded interfaces most remarkably. In order to precisely process surface interlocking microstructures and facilitate its intelligent manufacturing, surface texture recognition based on deep learning has been conducted on Ti6Al4V titanium alloy. A cross-material transfer (CMT) strategy is proposed to transfer the pre-trained model on the Ti6Al4V dataset via an optimal ResNet architecture to the texture recognition of DP590 steel and Aluminium alloy AA6061 using two novel functions: (1) a domain-adaptive attention mechanism to modulate the contribution of inherited attention, and (2) a three-stage gradual optimization strategy to address the asymmetric labelling and domain shift to ensure stable convergence and global feature alignment across domains. The results show that the ResNet50 integrated with the attention mechanism of CBAM exhibits superior performance in the recognition of laser-manufactured microstructures with varying depths. The CMT framework proposed achieves outstanding performance on cross-material recognition of surface textures for small-sample datasets compared to training from scratch and ImageNet-pretrained model, proving that laser-processed features constitute transferable laser-induced morphological descriptors of high transferability. The CMT framework also improved training efficiency, with the reduction in training time depending on the labelled target-domain data ratio. These make the CMT highly suitable for smart manufacturing.

Читать у издателя →

Похожие статьи