Global lithospheric thickness reconstruction using machine learning

The lithosphere, as the outermost solid layer of our planet, preserves a progressively more fragmentary record of geological events and processes from Earth's history the further back in time one looks. Thus, the evolution of lithospheric thickness and its cascading impacts on Earth’s tectonic system are presently unknown. Herein, we track the lithospheric thickness history using machine learning based on lithogeochemical big data of basalt. Our results demonstrate that four dramatic lithospheric thinning events occurred during the Paleoarchean, early Paleoproterozoic, Neoproterozoic, and Phanerozoic with intermediate thickening scenarios. These events respectively correspond to supercontinent breakup and assembly periods. Causality investigation further indicates that crustal metamorphic and deformation styles are the feedback of lithospheric thickness. Cross-correlation between lithospheric thickness and metamorphic thermal gradients record the transition from intra-oceanic subduction systems to continental margin plus intra-oceanic in the Paleoarchean and Mesoarchean, and a progressive emergence of large thick continents that allow supercontinent growth, which promoted assembly of the first supercontinent during the Neoarchean.

Methods
Our compiled global basaltic rocks database (see data S1) contains 24,194 chronological and oxide analyses of whole-rock, aged between 3.8 Ga to the present. The chronological and oxide data were obtained from the EarthChem data repository (http://portal.earthchem.org/). The lithospheric thicknesses are caclulated by the method in the paper. 

Our compiled experimental basaltic database (see data S2) contains 1,392 experimental basaltic liquid compositions in equilibrium with olivine and orthopyroxene, compiled from the LEPR database (https://lepr.earthchem.org/) and published literature. Samples were filtered to include only those basalts with SiO2 (43–55 wt%) and MgO (7–17 wt%). We also manually removed the samples with total oxides below 95 wt%, and missing pressure and temperature.

Our compiled global metamorphic thermal gradient database (see data S3, n= 564) is after Brown et al. (2020, Geology).

data S4 is the lithogeochemical data of the Cenozoic basalts (after Guo et al., 2020, Geology) and lithospheric thickness reconstruction results in eastern China.

data S5 is the metamorphic thermal gradient after Neogene and corresponding lithospheric thickness obtained by Litho1.0.

data S6 is the normalized reconstructed PF and T/P series, results for the global and local Pearson correlation coefficient (PCC), time-lagged cross-correlation coefficients (TLCC), and windowed TLCC analyses.

Codes 1-7 are the computational source codes used in this paper.