A physically interpretable deep hybrid artificial intelligence framework integrating fuzzy logic with dual-branch neural networks for magnesium alloy classification

Abstract

Rapid and rigorous classification of magnesium alloys is essential for industrial manufacturing and recycling, yet it remains challenging due to the subtle spectral variations and severe matrix effects inherent in complex metallurgical environments. This study proposes a novel analytical paradigm synergizing femtosecond laser-ablation spark-induced breakdown spectroscopy with a physically interpretable deep hybrid learning framework for the high-fidelity identification of five magnesium alloy grades. To mitigate instrumental vagueness and non-linear spectral fluctuations, the proposed artificial intelligence framework introduces a Gaussian-based fuzzy logic layer to map signals into a robust semantic space. The stabilized representations are subsequently processed through a parallel dual-branch architecture comprising an attention-augmented bidirectional long short-term memory network for decoding long-range sequential dependencies, and a one-dimensional convolutional neural network for isolating high-frequency localized atomic emissions. Evaluated under a rigorous, leakage-free sample level partitioning scheme, the proposed framework achieved an exceptional average accuracy of 99.97% across four independent test sets, decisively outperforming all baseline models, including support vector machines at 99.62%, Residual Network at 98.75%, random forests at 96.42%, and Transformer at 89.04%. McNemar tests confirmed that the superiority of the proposed framework over every baseline is statistically significant with all p values below 0.001. Comprehensive ablation studies confirm that the synergistic fusion of fuzzy logic, spatiotemporal deep learning, hybrid statistical features, and normalization preprocessing is indispensable for superior spectral disentanglement. The model exhibits a sub-millisecond inference speed of 0.0489 ms per sample with 7,366,438 trainable parameters, suitable for edge computing deployment, while Integrated Gradients attribution confirms its decision-making relies on genuine physical emission lines. Ultimately, this synergistic approach establishes a highly accurate, interpretable, and scalable artificial intelligence solution for real-time alloy sorting.