Aligning Expert Knowledge With Microstructural Features: A Cross‐Modal CLIP Framework for Continuous Spheroidization Grading of 12Cr1MoV Steel

The microstructural degradation of 12Cr1MoV heat‐resistant steel, primarily pearlite/bainite spheroidization, is a critical indicator of material aging. Conventional automated metallography relies on visual‐only “black‐box” CNNs, treating degradation levels as discrete labels and ignoring physical rules, continuous evolution, and expert morphological descriptions (e.g., DL/T 773). This paper proposes a cross‐modal framework, PEO‐DenseNet‐CLIP, to align multiscale visual perception with domain text semantics. A customized image encoder captures carbide dimensional variations (from fine lamellae to coarse spheres) and suppresses preparation noise, while a text encoder processes expert diagnostic descriptions as prompts. Via Information Noise‐Contrastive Estimation (InfoNCE) contrastive learning, visual features are strictly aligned with metallurgical semantics. An Ordinal Distance‐based Label Smoothing (ODLS) strategy further addresses ambiguous “half grades” (e.g., Grade 2.5), modeling the continuous nature of phase transformations. Experiments show that embedding expert knowledge into the visual space delivers exceptional fine‐grained discrimination and robustness on transitional microstructures, providing an effective cross‐modal solution for expert‐knowledge‐informed quantitative metallography.