Abstract: The commonly used elemental mapping method in gemstone origin tracing exhibits inherent limitations, such as subjectivity in element selection, reliance on original samples, and overlapping distribution of multiple origins in two-dimensional mapping. Machine learning (ML) has been widely applied in classification scenarios, including medical diagnosis and crop traceability. While linear discriminant analysis (LDA) has been extensively studied for gemstone origin determination, other ML algorithms have received less attention. In this study, peridot samples from three origins (Damaping, Hebei; Yiqisong, Jilin; Changwon District, Democratic People's Republic of Korea) were analyzed using LA-ICP-MS and modeled with Python. The influence of element selection on LDA effectiveness was analyzed. Results showed that selecting elements with low correlation and significant origin distribution differences improved model accuracy. A linear discriminant model using 10 elements (Mn, Zn, Na, Al, Sc, V, Cr, P, Ti, REE) achieved 0.889 cross-validation accuracy, outperforming models with all detectable elements. Comparing different ML algorithms (LDA, SVM, Decision tree, Random forest, Back propagation neural network) based on these 10 elements, non-linear algorithms, especially SVM, showed better performance.