Quantifying Continental Crust Thickness Using the Machine Learning Method.

Crustal thickness plays a key role in many geological processes. However, it remains challenging to quantify crustal thickness in the geological past. Here we propose an Extremely Randomized Trees algorithm‐based machine learning model to recover crustal thickness of old geological regions. The model is trained using major oxide and trace element compositions of 1,480 young intermediate to felsic rocks from global arcs and collisional orogens and geophysical measurements of crustal thickness. The model provides better estimations of crustal thickness than the commonly used methods based on Sr/Y and (La/Yb)N when applied to the testing data. The validity of this model is further demonstrated by its applications to the Kohistan–Ladakh, Gangdese and Talkeetna arcs, where paleocrustal thicknesses have been well constrained. We then use this model to construct the Mesozoic crustal thickness evolution of the Erguna Block in the southeast of the Mongol–Okhotsk suture belt. The closure time of the suture zone is still debated. Our results suggest that the crustal thickness of the Erguna Block increased from 43 ± 9 km at 210 Ma to 62 ± 7 km at 180 Ma, remained constant between 180 and 150 Ma, and then thinned to 36 ± 4 km at 120 Ma. These results suggest that the Mongol–Okhotsk Ocean closed in the Early–Middle Jurassic and the thickened crust was stretched during the Cretaceous. We show that the thick crust and compression‐extension transition seem to be favorable for the formation of porphyry copper deposits in the Erguna Block during the Late Jurassic. Plain Language Summary: Crustal thickness influences surface processes, magmatic compositions, formation of porphyry deposits, and regional lithospheric strength and stress regime, and its evolution may track tectonic paradigm changes. Although present crustal thickness can be detected by geophysical methods, it is more difficult to quantify crustal thickness in the geological past. We propose a machine learning model that is trained with major oxide and trace element compositions of magmatic rocks to quantify crustal thickness. The performance of our machine learning model on the testing data and several collisional zones shows that it provides a better crustal thickness estimate than the widely used Sr/Y and (La/Yb)N methods. As a case study, we apply this model to the Mesozoic magmatic belt in the Erguna Block southeast of the Mongol–Okhotsk suture in the eastern Central Asian Orogenic Belt, and suggest that the Mongol–Okhotsk Ocean closed at ∼180 Ma. Key Points: The machine learning model provides better estimations of crustal thickness than the commonly used methods based on Sr/Y and (La/Yb)NThe reconstructed crustal thicknesses of the Kohistan–Ladakh, Talkeetna and Gangdese arcs are consistent with previous studiesThe crustal thickness evolution of Erguna Block suggests Mongol–Okhotsk Ocean closed at ∼180 Ma [ABSTRACT FROM AUTHOR]