Copper alloys represent a crucial category of industrial materials extensively employed in manufacturing applications. By introducing various metallic elements into a pure copper matrix, these alloys can be engineered to achieve diverse microstructures and enhanced properties. The mechanical characteristics and functional performance of copper alloys are fundamentally governed by their compositional elements, elemental distribution patterns, and internal microstructural features. For example, trace additions of elements such as tin and antimony can significantly improve the physical properties of these alloys. While traditional metallographic analysis methods remain effective for alloy classification, they suffer from practical limitations including time-consuming procedures and operational inefficiency. Laser-induced breakdown spectroscopy (LIBS) has emerged as an innovative solution, offering rapid analytical capabilities through plasma spectroscopy. This technique combines micro-destructive sampling with simultaneous multi-element detection, finding widespread applications in fields ranging from metallurgical analysis to environmental monitoring and biomedical research. The integration of LIBS technology with pattern recognition algorithms shows a particular promise for alloy classification tasks. However, current implementations of machine learning models in LIBS spectral analysis often function as opaque “black boxes,” exhibiting critical limitations such as poor model interpretability, training data overfitting, and inadequate generalization performance when handling new spectral data. To address these challenges, interpretable deep learning approaches are being developed. These next-generation algorithms employ explainable artificial intelligence (AI) techniques to demystify modelling decision-making processes, thereby enhancing classification accuracy while maintaining computational efficiency. This dual focus on performance optimization and decision transparency represents a significant advancement in analytical methodology for material characterization. In this study, target-enhanced orthogonal dual-pulse LIBS (DP-LIBS) is combined with interpretable deep learning algorithms for the rapid and accurate classification of copper alloy samples.

In the experiment, DP-LIBS is utilized to collect spectral data from copper alloy samples. This approach achieves enhanced spectral line information while ensuring minimal sample damage, resulting in richer feature information. The application of interpretable deep learning algorithms aids in understanding the decision-making process of the classification model, thereby improving the execution efficiency of the algorithm. Each spectral dataset covers wavelengths ranging from 196.8 nm to 500.1 nm, containing a total of 6144 spectral features. For data processing, spectral wavelength features are selectively extracted based on their quantified contribution to classification outcomes, enabling dimensionality reduction of the spectral dataset. This process is achieved through attribution analysis using the integral gradients method, which systematically evaluates the influence of individual wavelength features on model predictions. Then, the decision logic of the classification model is clarified by quantifying the impact of individual spectral wavelengths on the categorical predictions through an attribution analysis. Critical spectral bands are identified by ranking features according to computed importance metrics, while redundant or non-informative wavelength regions are eliminated. The application of interpretable deep learning algorithms aids in understanding the decision-making process of the classification model, thereby improving the execution efficiency of the algorithm.

Significant enhancement in spectral signals and richer spectral line information are observed by DP-LIBS excitation. Compared with those under the LIBS condition, the signal intensities of Cu I 324.75 nm, Cu I 327.40 nm, and Zn I 334.50 nm spectral lines are enhanced under the target-enhanced DP-LIBS condition by a factor of 37.4, 34.8, and 31.8, respectively (Fig. 3). The choice of optimizer for the convolutional neural network (CNN) classification model significantly impacts training speed and network performance. Four optimizers are tested during training. The Adam optimizer demonstrates the highest accuracy and optimal performance (Fig. 5), making it the best choice for the DP-LIBS system. Through integral gradient analysis, the feature points of samples 1?3 with top 100 integral gradient (IG) values are mapped to local regions of the experimental spectra, highlighted in red in the correspondence diagrams (Fig. 6). These high-intensity points reflect spectral features that contribute most significantly to prediction outcomes. The feature-spectra alignment diagrams illustrate the specific spectral details prioritized by the classification model during internal data processing. To validate the performance of the IG-based feature selection algorithm, comparisons are made with manual feature extraction and principal component analysis (PCA) for dimensionality reduction. Classification accuracies for manual feature selection, PCA, and IG-based methods are 84%, 92%, and 99%, respectively. The IG-based method reduces redundant memory usage by 83% and compresses training time by 91% compared to the pre-screening approaches (Fig. 7). This analysis substantiates that interpretable feature selection transcends conventional dimensionality reduction by synergistically enhancing both computational efficiency and physical interpretability.

This study presents a novel methodology for rapid and precise classification of copper alloy specimens by integrating DP-LIBS with interpretable deep learning algorithms. The DP-LIBS technique substantially amplifies elemental signal intensities, yielding enhanced spectral signatures with richer characteristic information. Feature importance analysis through interpretable deep learning reveals that not only the elemental emission lines but also the selected baseline-derived spectral features significantly influence model performance, thereby elucidating the CNN internal decision-making mechanism for the input feature processing. Subsequent feature selection guided by importance metrics prior to model retraining markedly improves the computational efficiency. The proposed approach maintains high classification accuracy while substantially reducing computational resource demands, simultaneously providing explicit explanations of internal data processing dynamics. Demonstrated generalizability and micro-destructive sampling compatibility make this methodology particularly suitable for rapid alloy classification in micro-ablation scenarios.