Machine learning on white mica short-wave infrared (SWIR) spectral data in the Tengjia Au deposit, Jiaodong peninsula (Eastern China): A prospecting indicator for lode gold deposits

• OPLS-DA analysis of mineral SWIR spectral data can distinguish wall-rock samples and ore samples. • RF analysis on white micas spectral data can effectively distinguish orebodies. • Machine learning integrated with SWIR analysis can effectively guide the exploration of lode gold deposits. Lode gold deposits are the primary source of global gold resources and possess significant mineralization potential at depth, necessitating new strategies to locate deep concealed orebodies. The Tengjia Au deposit, a newly-discovered concealed altered rock-type lode gold deposit (50 t @ 3.89 g/t), is located within the Zhaoping metallogenic belt of the illustrious gold-rich Jiaodong peninsula in Eastern China. It is distinguished by pervasive phyllic alteration associated with gold mineralization, making it an ideal target for mineral geochemical exploration in lode gold deposits. The mineralization and alteration at Tengjia unfold across three distinct stages, delineated by mineral assemblages and textural relationships: K-feldspar-quartz (I), quartz-sericite-native gold-sulfide (Ⅱ), quartz-calcite-galena-sphalerite (Ⅲ) stages. Systematic analysis of short wavelength infrared (SWIR) spectra, coupled with petrographic observation, has unveiled an abundance of white micas (montmorillonite, muscovite, illite, paragonite, and phengite) within Stage Ⅱ at Tengjia. The Al-OH absorption feature wavelengths (Pos2200), as well as illite crystallinity (IC) values, exhibit a discernible shift towards longer wavelengths (>2204 nm) and higher values (>1.4) in the vicinity of ore deposition, which likely resulted from intense water–rock interaction between ore-forming fluid and wall rocks. Discriminant analysis of the orthogonal partial least squares method (OPLS-DA) shows that the absorption wavelengths corresponding to Water, –OH, and Al-OH effectively differentiate between ore and wall-rock samples. Additionally, analysis using the random forest algorithm (RF) demonstrates that spectral data from Tengjia white micas can reliably classify orebodies, achieving an accuracy of 83.2 %. Hence, the findings suggest that the unique SWIR spectral features of white micas offer a valuable tool for detecting the concealed Tengjia gold mineralization. This study proposes a novel approach that integrates machine learning technology with SWIR analysis for the identification of concealed lode gold deposits.