FALCON-Net: A hybrid fuzzy-attentive LSTM-convolutional neural network architecture for high-precision classification of steel alloys via femtosecond laser-ablation spark-induced breakdown spectroscopy

Background

Steel alloy classification demands advanced methodologies to address spectral noise, matrix effects, and nonlinear interactions in laser-induced breakdown spectroscopy (LIBS). This study proposes FALCON-Net, a hybrid architecture integrating fuzzy logic, LSTM, attention mechanisms, and CNN to enhance LIBS-based identification. Utilizing femtosecond laser-ablation spark-induced breakdown spectroscopy (fs-LA-SIBS), the model processes spectral data from nine steel alloys with distinct Cr, Mn, Ni, Cu, V, Mo, and Al compositions. The fuzzy layer employs Gaussian functions to suppress noise, while attention-guided LSTM captures temporal dynamics and critical wavelengths. CNN extracts spatial features, and fused representations enable robust classification.

Results

FALCON-Net achieves 100 % accuracy across four test sets, with ROC curves yielding perfect AUC scores (1.0), surpassing benchmarks (RF: 97.89 %, SVM: 97.23 %). Under 20 dB Gaussian noise, it maintains 97.76 % accuracy, validating noise resilience. Meanwhile, the feature fusion technique reduces wavelength interference and improves model stability.

Significance

This work presents the first integration of fuzzy logic and attention-enhanced deep learning for LIBS-based steel alloy analysis. The exceptional performance of FALCON-Net highlights the transformative potential of hybrid architectures in addressing real-world spectroscopic challenges. By offering high accuracy and noise resilience, the framework provides a viable solution for industrial quality control and material characterization, accelerating the practical implementation of fs-LA-SIBS technique in precision metallurgy.