A role-separated multimodal foundation-model framework for cross-modal verification of process-microstructure-B50 relations in non-oriented electrical steel

Le résumé fourni par la source

Non-oriented electrical steel (NOES) is a core soft magnetic material for electric motors, its flux density B50 indicating efficiency. Because B50 is governed mainly by texture and grain morphology, relating it to microstructure and verifying that relation is more informative than prediction alone. A multimodal foundation-model framework combining three role-separated tools to interpret, generate, and cross-modally verify process–microstructure–B50 relations is presented. Si 3.27 wt% NOES (0.35 mm) was annealed under 40 conditions (750–1000 °C, 5–60 min), yielding 205 micrographs with B50 spanning 1.529–1.608 T (σ = 0.023 T). With a 25-D input (process variables, Mask R-CNN grain descriptors, 16-D Swin latent), B50 was already well explained by process variables alone (TabPFN R² = 0.758, MAE = 0.0086 T); the microstructure modalities did not improve accuracy, as annealing governs microstructure and renders morphological descriptors largely redundant; instead they serve interpretation, generation, and verification. A segmentation-trained Mask2Former, never given magnetic labels, yielded a latent z₁ aligning with independent grain descriptors at R² = 0.87–0.98, exceeding the inter-segmenter floor. Its cross-attention shifted from grain boundaries to interiors with rising temperature, tracking grain coarsening; boundary focus co-varied with B50 (ρ = +0.73), an association carried by annealing temperature. A process-conditioned generator reproduced grain size at R² = 0.926 in-distribution and 0.874 under interpolation. At the recommended condition (T⁎ = 800 °C, t⁎ = 45 min), predictor, generator, and measurement agreed within 10%. This framework shows promise as a route to interpreting, generating, and verifying process–microstructure–property relations, beyond predictive accuracy alone.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.