Abstract:Bauxite is the primary raw material for alumina production, and its alumina content is a key indicator for evaluating ore grade and controlling smelting efficiency. Although conventional chemical analysis methods offer high accuracy, they involve complicated procedures, long detection cycles, and lack on-site analytical capability. Near-infrared (NIR) spectroscopy, characterized by rapid, non-destructive, and environmentally friendly measurement, provides a promising alternative for quantitative bauxite analysis. However, in portable detection scenarios, challenges such as light source instability, signal noise, and modeling robustness still restrict its analytical accuracy and generalizability. To address these issues, this study proposes a rapid alumina detection method by integrating portable NIR spectroscopy with a Radial Basis Function Kolmogorov-Arnold Network (RBF-KAN). A total of 778 bauxite samples from Australia, Guinea, Indonesia, and other major sources were analyzed. The alumina content, ranging from 38.54% to 57.66%, was determined following the YS/T 575.1–2007 standard. NIR spectra were acquired using a VIAVI MicroNIR Pro spectrometer covering 908.1-1676.2 nm at about 6 nm resolution. Each sample was scanned five times under constant geometry and averaged to minimize local mineral heterogeneity. To reduce baseline drift and scattering caused by grain size and moisture variation, standard normal variate preprocessing was applied before modeling. The proposed RBF-KAN model is developed based on the Kolmogorov-Arnold representation theorem, which decomposes multivariate nonlinear mappings into compositions of univariate functions. Unlike conventional KAN structures that employ B-spline functions, RBF-KAN replaces them with learnable radial basis functions whose centers and bandwidths are optimized during training, enabling adaptive characterization of localized nonlinear spectral responses. L1 regularization is further introduced to enhance sparsity and guide the model to focus on chemically relevant wavelengths. The model was trained using the RAdam optimizer with five-fold cross-validation. The predictive performance was comprehensively evaluated using four metrics: root mean square error (RMSE), mean absolute error (MAE), Pearson correlation coefficient (R), and coefficient of determination (R^2). Experimental results show that the RBF-KAN achieved an average RMSE of 1.1338%, MAE of 0.6506%, R of 0.9706, and R^2 of 0.9414, outperforming traditional models such as support vector regression, random forest, and partial least squares regression. Compared with classical deep learning networks (MLP, CNN, Transformer, and Unet variants), RBF-KAN achieved higher accuracy with only 10,840 parameters, reflecting its structural efficiency. Ablation studies further confirmed that applying SNV preprocessing reduced the RMSE by 35.5%, replacing the B-spline with RBF increased the R^2 by 1.5%, and incorporating L1 regularization improved feature sparsity and enhanced the model’s robustness across different data partitions. Visualization of the first-layer weights revealed a distinct activation peak at 1434.62-1459.40 nm, corresponding to the first overtone of O-H stretching vibration. This band corresponds to the Al-OH stretching overtone in hydroxyl-bearing aluminosilicate minerals, indicating that the model captures physically meaningful spectral features. Overall, the proposed RBF-KAN framework effectively integrates the interpretability of functional decomposition with the adaptiveness of radial basis mapping. It achieves high-precision and physically interpretable alumina quantification from portable NIR spectra. The method offers a reliable and efficient approach for on-site mineral analysis and provides a theoretical foundation for intelligent, field-deployable quantitative assessment in the aluminum industry.