MambaOut-CA for Multimodal EBSD and SEM Phase-Microstructure Segmentation in Additively Manufactured Alloys

Alloy additively manufactured parts often exhibit steep thermal gradients, elongated grains, melt-pool bands, and locally ambiguous phase contrast, making separate analysis of electron backscatter diffraction and scanning electron microscopy insufficient for robust phase segmentation. This study introduces MambaOut-CA, a multimodal segmentation model that integrates EBSD orientation descriptors and SEM morphology through a registration-aware encoder, MambaOut residual mixing, and coordinate attention. A curated dataset of 4,320 paired EBSD-SEM tiles from laser powder bed fused nickel-based alloy specimens was used to evaluate four phase classes and boundary-sensitive labels. The method combines reliability-aware input encoding, modality-specific shallow feature extraction, coordinate-guided context aggregation, and a boundary-assisted decoder to reduce errors around melt-pool traces, carbide particles, and uncertain interfaces. Compared with the UNet baseline, the proposed model increased mean intersection-over-union to 88.9%, improved boundary F1 to 84.8%, and reduced phase-fraction error to 1.6% across three build heights. The model retained 87.3% mean intersection-over-union under 20% EBSD confidence dropout and required 27.8 ms per 512 x 512 tile. Additional tests show that coordinate attention improves boundary continuity, while the reliability mask limits error propagation during modality degradation. These results indicate that explicit multimodal alignment and coordinate-guided context aggregation can improve phase continuity and preserve thin precipitate networks for quantitative metallographic analysis.