Self-supervised learning for grain size prediction in Fe-Si alloy steel sheets: A TabNet-based approach with explainable AI validation

Conventional grain size measurement methods for Non-Oriented Electrical Steel (NOES), such as chemical etching and optical microscopy, are destructive and unsuitable for full-scale inspection. In addition, the high cost of acquiring labeled grain size data constrains the generalization of conventional supervised machine learning models in this application. To address these limitations, three Self-Supervised Learning (SSL) models—VIME, SCARF, and TabNet—were employed with Accumulated Local Effects (ALE) and Counterfactual Explanations (CE) to non-destructively predict grain size from magnetic properties measured using a single-sheet tester. The study used 0.1 mm-thick Fe-Si NOES sheets (Si 0.61 wt%, Al 0.34 wt%) with grain sizes spanning 42.1–116.8 ㎛. TabNet-SSL achieved the highest predictive accuracy of R² = 0.81 and RMSE = 12.20 ㎛, outperforming conventional supervised baselines (Random Forest, Gradient Boosting, MLP, XGBoost) by +0.29 in R². A lightweight model utilizing only 30 input variables—a 70% reduction in input dimensionality—maintained R² = 0.73. ALE and CE analyses confirmed that the predictive mechanisms are consistent with magnetics theory, including domain wall pinning effects and permeability behavior. These findings demonstrate that non-destructive grain size prediction is achievable from magnetic measurements. The XAI-driven analysis establishes an interpretable basis for measurement condition selection, enabling reduction of required input variables. Within the validated range of alloy composition, sheet thickness, and grain size, the proposed framework shows potential for XAI-driven measurement protocol design and supports quality monitoring of NOES sheets. This offers a viable route toward data-driven process management with reduced measurement cost.