Mineralogical Characteristic, Formation Mechanism and Origin Determination of Peridot from Yiqisong Nanshan, Jilin Province, China

A field investigation on Yiqisong Nanshan peridot deposit in Jilin Province, China were carried and several reliable basalts and peridot crystals were collected for mineralogical, gemmological and origin determination studies. The deposit is predominantly comprised of lherzolitic basalt formed during Neogene volcanic activity, with olivine crystallisation temperatures around 919 ℃. Mantle-derived olivine megacrysts were entrained by ascending basaltic magma, and fragmented during ascent, while concurrent crystal fractionation led to the crystallisation of diopside, enstatite, and porphyritic olivine phenocrysts. Only the larger mantle olivine fragments that survived the magmatic transport have the potential to be gem-grade peridot. Gem-grade peridot is classified into five grades based on crystal dimensions, colour, clarity and cutting quality. It is green to yellowish green, typically>3 mm in size, and exhibits high forsterite content (Fo>90) and Ni concentrations (exceeding 0.3 wt.%). In contrast, most of non-gem-grade olivine (yellowish brown to brown, typically < 3 mm) corresponds to basaltic phenocrysts with lower Fo values (~80) and Ni concentration (below 0.1 wt.%). Using peridot geochemical signatures, a systematic comparison was conducted to evaluate the classification performance and model robustness of conventional Linear Discriminant Analysis (LDA) versus several machine learning algorithms for five origins: Yiqisong Nanshan (Jilin, China), Damaping (Hebei, China), Jangyon (Hwanghaenam-do, North Korea), Kohistan (Khyber Pakhtunkhwa, Pakistan), and San Carlos (Arizona, USA). Under strictly nested cross-validation, most machine learning algorithms significantly outperformed conventional LDA in accuracy. The bagging tree model achieved the highest mean accuracy (97.65% on outer-loop test sets), demonstrating superior capability for determining unknown samples. Meanwhile, the boosting tree model showed the best stability (standard deviation: 0.012 8 on outer-loop test groups) while maintaining high classification accuracy.