Machine learning of geophysical data: An innovative methodology to understand the nature and tectonic evolution of cratonic lithosphere

Abstract

The lithology of the deep lithosphere is crucial to providing an understanding of its evolution and related dynamic processes. However, the physical properties of rocks that make up these lithologies depend on various factors, such as mineral and chemical composition, confining pressure, temperature, and pore fluid pressure. Uncertainties in rock composition and pressure–temperature conditions at their depth of formation lead to non-uniqueness in geophysical exploration data. This study employs machine learning methods to accurately predict the lithological composition of the lithosphere using experimental data including P-wave velocity (V_P), S-wave velocity (V_S), density (ρ), and derivative parameters from 36 rock types. This achieves quantitative evaluation of resulting classifications and improves precision and reproducibility. Taking the South China Block as an example, this study shows that, although the Sichuan Basin preserves a thick cratonic lithosphere, its mantle lithosphere exhibits significant re-fertilization related to the plume activity in the Emeishan Large Igneous Province. More importantly, in the central and eastern South China Block, geological evidence combined with existing geodynamic models allows the reconstruction of post-Mesozoic lithosphere evolution to be divided into four major stages. These are: (1) subduction stage, (2) dehydration and eclogitization of crust in subducting lithosphere, (3) slab rollback and lithosphere delamination stage, and (4) extension stage and mantle-derived magmatic underplating. This study identifies a series of key geological parameters and provides new insights into methodology to understand the architecture and evolution of cratonic lithosphere systems globally provided the terranes have sufficient high-resolution geophysical data.