Rapid Classification of Bauxite Ores from Different Mining Areas by BBO-XGBoost Combined with CC-LIBS Authors: Yang Yanwei, Zhang Jiaxin, Zhang Lili Published: 2026 ABSTRACT Focusing on the problems of complicated sample processing and long detection time of traditional bauxite ore composition testing in mining areas, this paper proposes a rapid bauxite ore classification method combining cavity-confined laser-induced breakdown spectroscopy (CC-LIBS) with Boruta feature selection and Bayesian optimization eXtreme Gradient Boosting (BBO-XGBoost), which can quickly and accurately classify bauxite ore samples from 9 different mining areas. First, the effects of different cavity configurations on the enhancement factor, signal-to-noise ratio (SNR), were investigated. When the cavity diameter is 5 mm, the enhancement factors for Al, Fe, and Ti reach 2.8, 3.1, and 2.6, respectively. Next, 900 × 12,248-dimensional spectral data under the D = 5 mm cavity configuration were preprocessed using the Boruta feature selection algorithm. Finally, the XGBoost classifier is optimized using Bayesian optimization to achieve accurate classification of bauxite. Compared with unoptimized approaches and other models, the classification method combining CC-LIBS and BBO-XGBoost achieves the best performance. The accuracy is 0.9852, the precision is 0.9861, the recall is 0.9852, and the F1-score is 0.9856. The proposed technique, combining CC-LIBS with BBO-XGBoost, not only plays a pivotal role in bauxite ore research but also provides a significant technical reference for the rapid identification of bauxite ore sources. Introduction As an important strategic mineral resource, the origin traceability and type identification of bauxite play a crucial role in mineral exploration, resource assessment and metallurgical process optimization. Accurate classification and detection not only help to ensure the quality of raw materials, but also provide a scientific basis for ore trade and supervision. However, traditional chemical analysis methods such as XRF (X-ray Fluorescence Spectrometer), ICP-MS (Inductively Coupled Plasma-Mass Spectrometry), , and NAA (Neutron Activation Analysis), , as well as ICP-AES (Inductively Coupled Plasma Atomic Emission Spectrometry), have high precision. Nevertheless, they face many problems such as complex sample pretreatment, long detection cycles, high costs, and difficulties in achieving rapid on-site analysis, which seriously restrict their application in large-scale screening and rapid monitoring. Therefore, a technology that offers simple operation and high analytical accuracy is urgently needed. Laser-Induced Breakdown Spectroscopy (LIBS) , technology is a new spectral detection technology that does not require complex sample preparation and offers fast response, multielement simultaneous detection, and in situ analysis. By irradiating the surface of the sample with a laser, local instantaneous heating with high energy density is generated. This triggers the avalanche effect of continuous collision ionization, causing the free electron density to increase exponentially, and finally forming a plasma with high temperature, high pressure, and high conductivity. The transition of the seed particles during the plasma cooling process forms a characteristic spectrum. Researchers can obtain the chemical composition of the sample by analyzing the wavelength and spectral intensity of the emission spectrum. At present, LIBS technology has been widely used in materials science, , medicine, , mineral exploration, , environmental testing, archeological identification, , metallurgy, and other fields for rapid classification and product quality evaluation of material samples. LIBS provides an effective solution for the rapid classification of bauxite. However, the traditional LIBS spectral data typically exhibit issues including weak intensity, low SNR, extremely high dimensionality (up to tens of thousands of features), which leads to ″dimension disaster″ and feature flooding during direct modeling, seriously affecting classification accuracy and model stability. To address the problems of weak LIBS spectral intensity and SNR, many scholars have studied spectral enhancement through various constraint methods. Shao et al. studied the effect of cylindrical cavity height on the spectral intensity in spatially constrained LIBS. They demonstrated that the cavity constraint can enhance the spectral intensity, reduce the detection limit, and improve the LIBS detection accuracy. Tang et al. studied the characteristics of plasma by placing Au nanoparticles (Au-NPs) on the surface of Cu samples in a magnetic field. Their results indicated that the spectral intensity of laser-induced plasma was greatly enhanced and the signal-to-noise ratio was also improved. Wang et al. studied the influence of laser pulse energy and discharge channel on spark discharge assisted LIBS, determined the optimal experimental conditions of LA-SIBS in the detection of aluminum alloy, brass and other samples, and reduced the detection limit of elements. Dan et al. studied the effect of heating sample combined with spatially confined LIBS on spectral intensity. They found that the spectral intensity increased with the increase of sample temperature. When the two external conditions worked together, the spectral enhancement effect collected by LIBS was the best. Under this condition, the intensity of the Al 396.2 nm spectral line increased by a factor of 2.1. These studies provide new ideas for solving the problems of weak LIBS spectral intensity and low signal-to-noise ratio of bauxite by using cavity constraints. In order to address the challenges posed by high spectral dimension and low classification accuracy, many research teams have used laser-induced breakdown spectroscopy (LIBS) technology to classify and identify substances with machine learning, thereby demonstrating strong material identification capabilities. Diaz et al. achieved effective classification of gold-bearing ores by combining laser-induced breakdown spectroscopy (LIBS) with principal Origin traceability. Sheng et al. demonstrated a method combining LIBS with the random forest algorithm to classify iron ore. Yang et al. applied machine learning techniques to classify LIBS data of iron ore, achieving rapid detection of mineral composition. Hao et al. successfully classified soils in different mining areas by optimizing neural networks and using advanced algorithms. Jin et al. combined laser-induced breakdown spectroscopy with high-frequency ultrasonic atomization system (AS-LIBS) and optimized grid search cross-validation (Opt-GSCV) to achieve high-precision and rapid classification of liquor in different dimensions. Chen et al. used cavity constraints combined with gray wolf optimization algorithm to optimize the bidirectional long short-term memory (GWO-BiLSTM) network to accurately classify the LIBS spectra of cigarettes. The above literature shows that using spectral constraint methods combined with advanced machine learning algorithms to model LIBS experimental data can effectively reduce and correct spectral intensity fluctuations caused by emission source noise and matrix effects, thereby improving measurement accuracy. At present, no research in the literature has addressed the rapid classification of bauxite using constraint methods and intelligent algorithms combined with LIBS. To this end, this paper constructs a laser-induced breakdown spectroscopy system based on cavity constraints (CC-LIBS). Additionally, it develops a qualitative analysis model using the Boruta feature selection algorithm combined with Bayesian optimization limit gradient boosting (BBO-XGBoost). The aim is to offer a fast and reliable method for classifying and identifying bauxite, thereby providing a scientific basis to ensure the quality of bauxite raw materials, ore trade, and supervision.The enhancement factor, SNR and electron temperature of LIBS spectral data at different cavity diameters were comprehensively compared to identify the optimal cavity confinement configuration for acquiring bauxite ore spectral data.The Boruta feature selection algorithm was implemented to dynamically screen the 12,248-dimensional spectral data. Unlike traditional rank-based methods, Boruta evaluates the statistical significance of each variable against shadow features, ensuring that only the most intrinsically relevant spectral features are retained. This process not only mitigates the “curse of dimensionality” but also purifies the input space for the subsequent optimization phase.A hierarchical optimization framework, integrating Boruta feature selection and Bayesian Optimization (BBO) with eXtreme Gradient Boosting (XGBoost), was established to navigate the complex hyperparameter landscape. The original contribution lies in the migration-driven search strategy of BBO, which enables the model to effectively escape local optimaa common limitation of PSO or GA when processing high-dimensional mineral spectra. By tightly coupling feature filtering with stochastic parameter migration, the framework ensures the global optimality of the classification model rather than merely tuning individual components. The enhancement factor, SNR and electron temperature of LIBS spectral data at different cavity diameters were comprehensively compared to identify the optimal cavity confinement configuration for acquiring bauxite ore spectral data. The enhancement factor, SNR and electron temperature of LIBS spectral data at different cavity diameters were comprehensively compared to identify the optimal cavity confinement configuration for acquiring bauxite ore spectral data. The Boruta feature selection algorithm was implemented to dynamically screen the 12,248-dimensional spectral data. Unlike traditional rank-based methods, Boruta evaluates the statistical significance of each variable against shadow features, ensuring that only the most intrinsically relevant spectral features are retained. This process not only mitigates the “curse of dimensionality” but also purifies the input space for the subsequent optimization phase. The Boruta feature selection algorithm was implemented to dynamically screen the 12,248-dimensional spectral data. Unlike traditional rank-based methods, Boruta evaluates the statistical significance of each variable against shadow features, ensuring that only the most intrinsically relevant spectral features are retained. This process not only mitigates the “curse of dimensionality” but also purifies the input space for the subsequent optimization phase. A hierarchical optimization framework, integrating Boruta feature selection and Bayesian Optimization (BBO) with eXtreme Gradient Boosting (XGBoost), was established to navigate the complex hyperparameter landscape. The original contribution lies in the migration-driven search strategy of BBO, which enables the model to effectively escape local optimaa common limitation of PSO or GA when processing high-dimensional mineral spectra. By tightly coupling feature filtering with stochastic parameter migration, the framework ensures the global optimality of the classification model rather than merely tuning individual components. A hierarchical optimization framework, integrating Boruta feature selection and Bayesian Optimization (BBO) with eXtreme Gradient Boosting (XGBoost), was established to navigate the complex hyperparameter landscape. The original contribution lies in the migration-driven search strategy of BBO, which enables the model to effectively escape local optimaa common limitation of PSO or GA when processing high-dimensional mineral spectra. By tightly coupling feature filtering with stochastic parameter migration, the framework ensures the global optimality of the classification model rather than merely tuning individual components. Experimental System Setup The instrument used for the experiments was a custom-built LIBS system. The laser is a Nd: YAG laser from Quantel, France, with a maximum frequency of 10 Hz, a maximum pulse energy of 200 mJ, a wavelength of 1064 nm, and a pulse duration of 10 ns. The spectrometer uses a seven-channel spectrometer from Avantes of The Netherlands, with a wide spectral range of 190–950 nm. The sample stage has a freely controllable XYZ axis. The stroke of the sample stage is 5 cm, which can achieve high-precision positioning control with micron-level precision (<1 μm). The laser beam was focused precisely on the geometric center of the sample surface within the cavity to ensure symmetrical shock wave reflection and consistent plasma confinement. The quality and type of bauxite are closely related to the smelting process and economic indicators of alumina. The main factor determining the quality of bauxite is chemical composition. Bauxite from different regions reflects the compositional characteristics specific to those regions. The accurate classification of bauxite is crucial for the acquisition and quality evaluation of raw materials in the alumina industry. The 9 different bauxite samples used in this study, listed in Table , were sourced from mining areas including Australia Weber, Australia Gove, Guinea, Aojiawan, Jinsui, Sunjiagou, Liulin, Xiaoyi, and Fugu, and provided by Shanxi Huaxing Aluminum Group Co., Ltd. All of them were taken from 9 representative aluminum production regions worldwide, and the bauxite ore sample collection sites are shown in Figure . After drying the bauxite ore to remove debris, it was ground by a high alumina ceramic ball mill to particles of 0.075 μm. Then, the powder was dried in a dryer for 24 h, weighed (3 g), and continuously pressed at 25 MPa for 150s to finally obtain bauxite ore pellets with a diameter of 30 mm and a thickness of 3 mm. 1Detailed Information of Bauxite Samples from Different Mining Areaslabelmining areaplace of originAJ001AojiawanAojiawan Town, Xingxian CityJS002JinsuiJinsui Town, Xingxian CitySJ003SunjiagouSunjiagou Town, Xingxian CityFG004FuguFugu City, Shanxi ProvinceLL005LiulinLiulin City, Shanxi ProvinceXY006XiaoyiXiaoyi City, Shanxi ProvinceGN007GuineaGuinea, West AfricaGF008GoveGove, Northern Territory, AustraliaWB009WeberWeber, Queensland, Australia Detailed Information of Bauxite Samples from Different Mining Areas Global distribution of bauxite sampling locations. In order to solve the problem of weak LIBS spectral intensity and low SNR, a cylindrical cavity with a height of 1 mm and a diameter of 2, 3, 4, 5, 6 mm was specially customized to study the influence of the cavity on the LIBS spectrum. In cavity-constrained LIBS, the cylindrical cavity is placed on the bauxite sample, and the influence of the cavity on the LIBS spectral intensity is explored by replacing the cavity of different sizes, so as to obtain the optimal cavity. Under the optimal cavity constraint, the LIBS spectral data are collected for subsequent qualitative analysis. The schematic diagram of CC-LIBS is shown in Figure . CC-LIBS experimental setup and system configuration. The laser energy is set to 80 mJ, the spectral delay time of the spectrometer is set to 1 μs, and the laser pulse repetition frequency is set to 10 Hz. Under the above experimental parameters, LIBS spectral data of Aojiawan bauxite samples were collected using different cavity sizes. The laser beam was focused on the sample surface through a focusing lens with a focal length of 100 mm to ablate the sample. Before data acquisition, the sample needs to be cleaned with the laser. The number of cleaning cycles is set to two, and data are collected five times. In this way, 30 (6 * 5 = 30) sets of LIBS spectral data are obtained under 6 kinds of cavities, which are used for LIBS spectral analysis of bauxite to obtain the optimal cavity. Under the optimal chamber, 10 samples were prepared for each bauxite sample. Data acquisition was carried out for different parts of 9 different bauxite samples, and 10 sets of data were collected for each sample. A total of 900 data sets were obtained for subsequent qualitative analysis. Data Analysis Methodology Boruta Feature Selection Algorithm Boruta is an all-relevant feature selection wrapper based on Random Forest (RF). Unlike minimal-optimal methods, Boruta identifies all features statistically related to the target variable by comparing original features with ″shadow features.″ The detailed workflow is implemented exclusively on the training data set to prevent data leakage: Step 1: Information Extension. The original feature matrix X n*p (where p = 12,248) is extended by adding shadow features Xshadow, created by randomly shuffling the values of each original feature across observations to remove their correlation with the target Y. Step 2: Importance Evaluation. A Random Forest is trained on the combined matrix [X,Xshadow]. The Mean Decrease Accuracy (MDA) or Gini importance Zi is calculated for each feature i. Step 3: Statistical Significance. Let Z max_shadow be the maximum importance score among all shadow features. A hit is recorded for an original feature if: Z original,i > Z max_shadow. This process is repeated for N iterations (e.g., N = 100). The cumulative hits follow a binomial distribution 1 P(X=k)=(Nk)qk(1−q)N−k where q = 0.5 represents the probability of a feature performing better than a shadow feature by random chance. Features that fall in the right tail of the distribution (typically p < 0.01) are marked as ″Confirmed.″ Bayesian Optimization Algorithm Bayesian optimization using the classification accuracy as the objective function to find the optimal configuration that meets the requirements of the classification task. To navigate the high-dimensional and nonlinear hyperparameter space of the classification model efficiently, Bayesian Optimization (BO) is implemented. Unlike traditional grid search, BO treats the hyperparameter tuning as a global optimization problem for a black-box function f(x)­ 2 x*=arg⁡maxx∈X⁡f(x) where x* represents the set of hyperparameters χ is the search space, and f(x) is the classification accuracy. The BO framework primarily consists of two core components:1.Surrogate model: A Gaussian Process (GP) is utilized to construct a probabilistic model of the objective function. Based on the observed results D 1:t = {(x1, y1),···, (xt, yt)} the GP provides a posterior distribution P(f | D{1:t}) that characterizes the uncertainty of the model across the entire search space.2.Acquisition function: To determine the next sampling point x t+1, an acquisition function (such as Expected α­(x) Improvement, EI) is employed to balance exploration (searching unsampled regions) and exploitation (refining known high-performance regions) 3 αEI(x)=E[max(0,f(x)−f(x+))] where f(x +) is the current maximum observed value. This iterative process ensures efficient convergence toward the optimal configuration while significantly reducing the computational overhead compared to exhaustive search methods. 1.Surrogate model: A Gaussian Process (GP) is utilized to construct a probabilistic model of the objective function. Based on the observed results D 1:t = {(x1, y1),···, (xt, yt)} the GP provides a posterior distribution P(f | D{1:t}) that characterizes the uncertainty of the model across the entire search space. Surrogate model: A Gaussian Process (GP) is utilized to construct a probabilistic model of the objective function. Based on the observed results D 1:t = {(x1, y1),···, (xt, yt)} the GP provides a posterior distribution P(f | D{1:t}) that characterizes the uncertainty of the model across the entire search space. 2.Acquisition function: To determine the next sampling point x t+1, an acquisition function (such as Expected α­(x) Improvement, EI) is employed to balance exploration (searching unsampled regions) and exploitation (refining known high-performance regions) Acquisition function: To determine the next sampling point x t+1, an acquisition function (such as Expected α­(x) Improvement, EI) is employed to balance exploration (searching unsampled regions) and exploitation (refining known high-performance regions) eXtreme Gradient Boosting (XGBoost) eXtreme Gradient Boosting (XGBoost) is selected as the primary classifier for bauxite ore classification due to its superior performance in handling high-dimensional and sparse spectral data. The model operates by constructing an ensemble of K additive trees to predict the output 4 ŷi=∑k=1Kfk(xi),fk∈F where F is the space of regression trees. To optimize the model, XGBoost minimizes a regularized objective function 5 L(ϕ)=∑i=1nl(yi,ŷi)+∑knΩ(fk) Here, L is a differentiable convex loss function that measures the difference between the prediction yi and the target yi . The second term Ω­(f k) is the regularization term designed to prevent overfitting by penalizing the complexity of the trees 6 Ω(f)=γT+12λ||w||2 where T is the number of leaves and w represents the leaf weights. A key technical advantage of XGBoost is the application of the second-order Taylor expansion of the loss function, which accelerates convergence by utilizing both the gradient gi and the Hessian hi 7 L(t)≈∑i=1n[l(yi,ŷ(t−1)))+gift(xi)+12hift2(xi)]+Ω(ft) By integrating these advanced optimization and regularization mechanisms, XGBoost provides a robust and interpretable framework for distinguishing the subtle geochemical signatures in bauxite spectra across different regions. Integrated Boruta-Bayesian XGBoost Framework To fully exploit the optimization potential within the high-dimensional CC-LIBS spectral space, this study develops a synergistic BBO-XGBoost framework. As illustrated in Figure , the architecture is organized into three hierarchical modules: Framework of the BBO-XGBoost model Statistical feature engineering: The Boruta algorithm is employed as an all-relevant feature selection wrapper. By generating shadow features and performing statistical significance testing, it effectively mitigates the ″curse of dimensionality″ while preserving discriminative spectral information. Bayesian hyperparameter optimization: To refine the classifier, a Gaussian process surrogate model is constructed to map the complex, nonlinear relationship between XGBoost hyperparameters and classification accuracy. An acquisition function is utilized to strategically balance exploration and exploitation, efficiently identifying the global optimal configuration. Model integration and validation: The optimized feature subset and hyperparameter configurations are integrated into the final BBO-XGBoost architecture. To rigorously assess the framework’s robust generalizability and mitigate the risk of overfitting, an 8-fold cross-validation protocol is implemented during the training phase. This systematic calibration, governed by the objective function’s fitness, ensures the selection of the most stable model configuration, thereby realizing high-precision, classification of bauxite ores. Accuracy is defined as the ratio of correctly predicted samples (including both true positives and true negatives) to the total number of samples in the data set. It is one of the most intuitive performance metrics, representing the overall proportion of correct classifications made by the model. Its formula is shown in eq . 8 Accuracy=(TP+TN)(TP+TN+FP+FN) Precision is defined as the ratio of the number of samples correctly predicted as positive classes to the number of samples predicted as positive classes, and is often used to indicate the degree of prediction of the correct sample outcome. Its formula is shown in eq . 9 Precision=TP(TP+FP) Recall is the ratio of the number of correct samples retrieved to the number of all correct samples in the database and is used to measure the ability of the classifier to identify the correct samples. Its formula is shown in eq . 10 Recall=TP(TP+FN) F1-score is the reconciled average of Precision and Recall, which is commonly used to evaluate the performance of machine learning models on unbalanced data sets. Its formula is shown in eq . 11 F1=(2×Precision×Recall)(Precision+Recall) Results and Discussion LIBS Spectroscopy under Cavity Confinement This study employed a cylindrical aluminum cavity constraint structure to enhance the laser-induced breakdown spectroscopy (LIBS) spectral intensity of bauxite samples. To determine the optimal size, the Aojiawan bauxite sample was used for the spectral enhancement test. Six different sizes of aluminum cavities (height of 1 mm, diameter of 2, 3, 4, 5, 6 mm) were used. For each cavity size, five sets of data were collected, resulting in a total of 30 (5 × 6) experimental data sets. The average value of the five sets of data under each cavity was taken as the data under this size. Al 396.15 nm, Fe 438.35 nm, and Ti 498.17 nm were selected as analytical lines. The effects of different sizes of cavities on LIBS spectral data were investigated by parameters such as enhancement factor and SNR. Enhancement Factor The enhancement factor described herein refers to the ratio of spectral intensity under cavity-constrained conditions to that under unconstrained conditions. It is used to characterize spectral intensity enhancement effects. As shown in Figure , when applying different cavities to constrain LIBS, the spectral intensities of aluminum, iron, and titanium first increase and then decrease. When the cavity diameter is 5 mm, the enhancement factor reaches its maximum values: 2.8 for Al, 3.1 for Fe, and 2.6 for Ti. Trend of Al, Fe and Ti enhancement factors under different cavity diameters. This phenomenon arises from the formation of shock waves accompanying the generation of plasma during laser irradiation of bauxite sample surfaces. The shock waves propagate faster than the plasma diffusion rate and contact the cavity walls before the plasma arrives. The confining effect of the cavity walls causes the shock wave to reflect toward the center of the plasma diffusion, thereby intensifying particle collisions within the plasma. This further excites the plasma, leading to an enhancement of the bauxite laser-induced breakdown spectroscopy. The enhancement factor reaches its maximum at a cavity diameter of 5 mm. As the cavity diameter continues to increase, the reflected shock wave energy becomes insufficient to sustain plasma enhancement, causing the enhancement factor to begin decreasing. Signal to Noise Ratio (SNR) SNR is another important parameter reflecting the quality of the spectral line, commonly used to evaluate the detection sensitivity of the spectral signal. According to eq , the SNR values calculated for the characteristic elements Al, Fe, and Ti are 42.8, 38.8, and 45.1, respectively, using a cavity configuration with D = 5 mm. 12 SNR=Ipσb where I p is the peak intensity and σb is the standard deviation of the background. Figure shows that the SNR varies with cavity diameter configurations and reaches its optimum at D = 5 mm. This enhancement is mainly due to two factors: the cavity confinement limits plasma expansion and concentrates the emission in a smaller volume, enhancing the spectral intensity of the elemental signal. On the other hand, the source of background noise is limited, the background radiation is weakened, thus reducing the probability of interference and improving the SNR. SNR of characteristic elements at different cavity diameters. By studying the influence of different cavities on LIBS spectral enhancement factor and SNR, we can conclude that the cavity with a diameter of 5 mm is optimal. Therefore, we collect bauxite spectral data under the optimal cavity for subsequent qualitative analysis. In order to further investigate the physical mechanism of the effect of the cavity confinement (CC) technique on the bauxite plasma radiation enhancement, the plasma electron temperature (T e) was calculated using the multiline spectroscopic method. The electron temperature has a significant effect on the atomization, excitation, and ionization of sample substances in plasma. By measuring the relative intensities of characteristic emission spectral lines of the Al, Fe, Si, and Ti atoms in the bauxite plasma emission spectrum, the corresponding electron temperatures were found according to eq Boltzmann plot method. Table shows the parameters of bauxite plasma spectra with data from NIST database. 13 ln(IkiλkigkAki)=−EkkBTe+C 2Spectroscopic Parameters of Characteristic Emission Lines for Bauxite Plasma Analysiselementwavelength/nm A k/108 s–1 E k/ev g k Al396.150.9853.142Fe438.350.5004.3111Si288.160.2175.083Ti498.170.6603.3413 Spectroscopic Parameters of Characteristic Emission Lines for Bauxite Plasma Analysis As can be seen from Figure , the electron temperature of the plasma varies when the cavity diameter is at 2, 3, 4, 5, and 6 mm, respectively. With the optimization of the cavity diameter, the electron temperature also has a small to large trend, and the electron temperature reaches 1.64 * 104 K at 5 mm, which rises significantly compared to the time when there is no cavity constraint. When the cavity diameter is increased to a certain extent, the electron temperature trend tends to level off, matching the conclusion above. The reason is that as the cavity diameter decreases toward the optimal size, the plasma is subjected to the wall reflection and compression, which restricts the expansion and motion of the plasma. The increase in the probability of collision between the particles makes the particles have different degrees of ionization, generating more new atoms, which ultimately makes the electron temperature rise. Electron temperature at different chamber diameter. The spatial synchronization of reflected shock waves, achieved by central ablation, is critical for maximizing the confinement effect. When the laser focuses on the geometric center, the reflected shock waves converge simultaneously toward the plasma core. This synchronized reflection intensifies the compression of the plasma, effectively converting the kinetic energy of the shock wave into internal plasma energy, which accounts for the significant rise in electron temperature and spectral intensity observed at D = 5 mm. Deviating from the center would lead to uneven energy distribution and diminished enhancement. Feature Selection and Classification Performance Under the optimal cavity, 10 samples are made for each bauxite areas, 10 sets of data are collected for each sample, and 100 sets of data can be obtained for each bauxite ore. To ensure the rigorous independence of experimental results and avoid information leakage, the 900 spectral data sets (comprising 9 mining areas, 10 physical pellets per area, and 10 locations per pellet) were partitioned at the pellet level. Specifically, 70% of the pellets from each area were randomly assigned to the training set, while the remaining 30% were equally split between the validation and test sets. This strategy ensures that each data set is strictly independent, with no overlapping spectra from the same physical sample across training and testing phases. As can be seen in Figure , it is difficult to directly recognize the bauxite ore elemental composition of the 9 different mining areas with small variability and high similarity of spectral information, so feature selection was performed using the Boruta algorithm on the full spectral information on the bauxite ores from different mining areas. The Boruta algorithm identified the most relevant spectral features by comparing real features against randomly permuted shadow features, ensuring statistical significance in feature selection. CC-LIBS spectra of 9 bauxite ore samples at D = 5 mm. To effectively mitigate spectral dimensionality and eliminate information redundancy, the Boruta algorithm was implemented to evaluate the statistical significance of 12,248 initial variables (Figure ). The results demonstrate that the vast majority of spectral variables were identified as stochastic noise or noninformative background and were subsequently categorized as Rejected Features. Through 100 iterations of Random Forest and rigorous binomial significance testing, 8 specific spectral variables were confirmed to have high importance scores exceeding the Z max_shadow decision threshold, categorized as Confirmed Features (p < 0.01). It is noteworthy that while Figure reveals a comprehensive range of atomic emission lines, the 8 confirmed variables align precisely with the characteristic emission transitions of Al, Fe, and Ti. This spatial consistency indicates that the selection process effectively filters out collinear redundancies and concentrates on the most robust spectral responses, thereby ensuring the generalizability and computational efficiency of the subsequent classification model. Boruta feature selection results showing feature. Algorithm Convergence and Hyperparameter Optimization To further enhance the generalization capability of the XGBoost model in handling high-dimensional (12,248 variables) LIBS spectral data, the Biogeography-Based Optimization (BBO) algorithm was employed for the global search of core hyperparameters (including learning_rate, max_depth, and n_estimators). The optimization process was conducted for 100 iterations. To verify the superiority of BBO, Genetic Algorithm (GA-XGBoost) and Particle Swarm Optimization (PSO-XGBoost) were introduced as baseline benchmarks under identical experimental conditions. Figure presents the convergence trajectories of the three optimization algorithms. It can be observed that all models start from a relatively high initial accuracy, benefiting from the high-quality spectral features selected by the Boruta algorithm. Comparison of the fitness evolution and optimization trajectories for three algorithms. While PSO-XGBoost demonstrates a rapid initial ascent, it quickly encounters a bottleneck at an accuracy of 97.85%, indicating premature convergence into local optima. In contrast, the proposed BBO-XGBoost exhibits superior global search capability. Through its unique migration mechanism, BBO undergoes a phase of stochastic exploration between 30 and 55 iterations, successfully surpassing the other benchmarks and stabilizing at a peak accuracy of 98.52%. This comparison confirms the effectiveness of BBO in refining hyperparameters within a narrow and complex search space for bauxite classification indicating premature convergence into local optima. In contrast, the proposed BBO-XGBoost exhibits superior global search capability specifically tailored for the intricate spectral features of bauxite. Given that bauxite spectra from different mining areas exhibit high similarity and narrow discriminative margins, the hyperparameter loss landscape is prone to rugged local minima. Through its unique migration mechanism, BBO undergoes a phase of stochastic exploration between 30 and 55 iterations. This allows the framework to effectively ’migrate’ optimal parameter traits across different high-dimensional feature subspaces identified by Boruta, successfully surpassing other benchmarks and stabilizing at a peak accuracy of 98.52%. This confirms that BBO is not merely a general optimizer but an essential adaptive strategy for refining the hyperparameter space in complex mineralogical classification tasks. Ablation Study and Modular Contribution Analysis To systematically verify the individual contributions of cavity confinement (CC), Boruta feature selection, and BBO optimization, a comprehensive ablation study was conducted. As shown in Table , five configurations were evaluated using 8-fold cross-validation, with each experiment repeated 10 times to determine the mean accuracy and standard deviation (SD). 3Ablation Study Results of the Proposed Framework for Bauxite ClassificationindexCC-LIBSBorutaBOaccuracy (mean ± SD)Δ accuracy1 84.21% ± 1.32%–14.31%2√ 88.56% ± 0.98%–9.96%3√ √91.25% ± 0.72%–7.27%4√√ 97.10% ± 0.48%–1.42%5√√√98.52% ± 0.11% Ablation Study Results of the Proposed Framework for Bauxite Classification The results indicate that Cavity Confinement provides a physical foundation for signal enhancement, increasing the baseline accuracy from 84.21 to 88.56% (+4.35%). The integration of Boruta (Row 4) yields a significant performance leap, pushing the accuracy to 97.10% (+8.54%), confirming that eliminating redundant variables is crucial for handling the 12,248-dimensional spectral data. Finally, the BO algorithm further refines the hyperparameter subspace, reaching the peak accuracy of 98.52% (+1.42%) with the lowest SD (±0.11%), which demonstrates superior model robustness. Wilcoxon signed-rank tests confirmed that the synergy of these three modules provides a statistically significant advantage (p < 0.01) over any partial configuration. Classification Results and Performance Evaluation Based on the integrated Boruta-BBO-XGBoost framework, high-precision classification of bauxite ores from nine distinct mining areas was successfully achieved. To rigorously evaluate the reliability and generalization of the optimization process, the model was subjected to an 8-fold cross-validation protocol. As illustrated in Figure , the classification accuracy across different folds exhibited a minimal fluctuation range of only 0.901%, fluctuating between 98.214 and 99.115%. This marginal variance demonstrates the exceptional stability of the integrated framework and its robustness against data set partitioning bias. 8-fold cross-validation. The proposed analytical pipeline was implemented within a standardized computational environment using Python 3.9. To eliminate stochastic bias and ensure that the experimental results are fully reproducible, a fixed random seed of 42 was strictly applied across all computational stages. This rigorous control of randomness encompassed the RobustScaler preprocessing phase, the iterative Boruta feature selection process, and the entire Bayesian Optimization (BO) framework. By maintaining this consistent execution environment, the integrity and reliability of the reported classification metrics are preserved across different hardware platforms. The optimal configuration of the XGBoost classifier was determined through a rigorous hyperparameter tuning process employing a Bayesian Optimization (BO) framework. This classifier was integrated with Boruta-selected features to enhance performance. Figure illustrates the performance landscape as a two-dimensional contour surface, showing the intricate nonlinear interactions between the learning rate and maximum depth within the defined search space (e.g., learning_rate in [0, 0.3], max_depth in [3, 10]). Model performance, measured by the target score using the Expected Improvement (EI) acquisition function, is visualized using a viridis colormap, where the gradient from deep purple to yellow-green represents the transition from suboptimal regions to the high-accuracy plateau. Note that while the Bayesian optimization landscape (Figure ) presents a continuous surface to illustrate the convergence trend and parameter interactions, the max_depth variable is strictly mapped to the nearest integer during actual model execution. This ensures the structural validity of the decision trees while allowing the BBO-XGBoost framework to navigate the hyperparameter space effectively. Optimization diagram of Bayesian parameters. Following 55 iterations of systematic exploration and exploitation, the global optimum (identified by the red asterisk) was found at a learning rate of 0.2591 and a maximum depth of 3. This configuration achieved an outstanding target score of 98.52%, demonstrating the model’s high fidelity in identifying complex bauxite ore classifications. The optimization trajectory, marked by the distribution of search points, reveals that a constrained maximum depth paired with a moderate learning rate strikes an ideal balance between structural complexity and generalization. This tuning effectively mitigates the risk of overfitting, ensuring that the BBO-XGBoost framework maintains robust predictive stability and superior discriminative power for large-scale CC-LIBS spectral data sets. The results corresponding to the optimal parameters of each classification model are selected for comparison with the integrated approach proposed in this paper. Figure a,b presents the confusion matrices for the training and test sets, respectively, which characterize the classification performance across bauxite ore samples from nine distinct mining areas. The training accuracy reached a stable plateau at approximately 99.52%, while the test set exhibited an exceptional accuracy of 98.52%. This minimal discrepancy between training and testing performance highlights the model’s strong precision and validates the effectiveness of the parameter set derived from the Bayesian optimization process. (a) Confusion matrix for train data. (b) Confusion matrix for test data. To further evaluate the robustness of the proposed CC-LIBS combined with the BBO-XGBoost framework, a comparative analysis was conducted involving several baseline machine learning architectures, including Support Vector Machine (SVM), Random Forest (RF), a standard XGBoost model without hyperparameter optimization, PSO-XGBoost, and GA-XGBoost. All models were trained using the Boruta-selected feature set to maintain fairness in comparison. The results, summarized according to the optimal parameter configurations of each candidate model, demonstrate that the integrated approach proposed in this study significantly outperforms traditional methods in both predictive accuracy and discriminative stability for complex bauxite ore classification tasks. The experimental results demonstrate that cavity confinement (D = 5) significantly enhances the overall classification accuracy across all evaluated models. The physical mechanism behind this improvement is that the cavity effect enhances the characteristic spectrum of bauxite relative to the background noise. As a result, the spectral SNR improved, and the consistency and repeatability of the experimental data were enhanced. Through this enhanced signal quality, the integrated BBO-XGBoost framework demonstrates superior discriminative accuracy compared to traditional machine learning models. As shown in Table , Statistical evaluation reveals that the proposed framework improves the classification accuracy on the test set by 0.67% relative to PSO-XGBoost, by 1.10% relative to GA-XGBoost, by 1.42% relative to standard XGBoost, by 2.98% relative to Random Forest, and by 7.42% relative to SVM. These results demonstrate that the synergistic combination of CC-LIBS and the intelligent optimization framework enable high-precision identification of bauxite ores from various mining regions. This methodology provides a robust and rapid analytical solution for mineralogical analysis and high-quality Automated identification in the aluminum industry. 4Classification Performance Comparison under D = 5 mm Cavity Confinementmodelhyperparameter optimizationaccuracyprecisionrecall F1-scoreBBO-XGBoostBayesian (55 iterations)0.98520.98610.98520.9856PSO-XGBoostPSO (25 iterations)0.97850.98020.97810.9791GA-XGBoostGA (70 iterations)0.97420.97550.9740.9747XGBoostnone (default)0.97100.97180.97100.9714RFgrid search0.95540.95620.95540.9558SVMgrid search0.91100.91240.91100.9117 Classification Performance Comparison under D = 5 mm Cavity Confinement Conclusion In this study, an integrated framework combining cavity-constrained laser-induced breakdown spectroscopy (CC-LIBS) with Boruta feature selection and Bayesian Optimized XGBoost (BBO-XGBoost) was developed for the rapid and accurate classification of bauxite ores from nine different mining areas. By studying the influence of different sizes of cylindrical cavities on the spectrum, the cavity with a diameter of 5 mm was identified as optimal. Using the optimal cavity, the enhancement factors of Al, Fe and Ti reached 2.8, 3.1, and 2.6, respectively. The original contribution of this work lies in the synergistic coupling of the physical enhancement and the algorithmic refinement. To address the challenges of high-dimensional spectral data, the Boruta algorithm was employed to reduce the dimensionality from 12,248 features to 8 core features effectively. Crucially, to overcome the stagnation issues common in traditional heuristic methods, the BBO algorithm’s unique migration mechanism was utilized to navigate the hyperparameter space. The optimized XGBoost classifier achieved an exceptional classification accuracy of 0.9852, with a precision of 0.9861 and an F1-score of 0.9856, significantly outperforming traditional models. These results demonstrate that the synergistic combination of CC-LIBS and the BBO-XGBoost framework provides a robust and highly precise analytical solution for bauxite ore.