A machine learning based-approach to predict the water content of mid-ocean ridge basalts

Water is critical in the evolution of the mantle due to its strong influence on the physicochemical properties of mantle rocks. Mid-ocean ridge basalts (MORBs) are commonly used to study the compositional characteristics of the convecting upper mantle. However, there remains abundant samples in the global MORB data sets without direct measurements of water contents. The commonly observed good correlation between H₂O and other incompatible trace components, such as Ce, has been applied to quantify water contents of MORBs. However, this approach assumes constant H₂O/Ce in the target samples, which is not always true as the H₂O/Ce ratios of MORBs could be rather heterogeneous even in some short ridge segments. Utilizing the present compositional data of global MORB glasses with measured water contents (n = 1,467), we construct a Random Forest Regression model based on machine learning, which can predict water concentrations of samples based on selected major and trace element data, without assuming a ratio between H₂O and other trace elements. This model allows us to precisely recover water contents for MORBs with comparable accuracy with traditional analytical methods. The predicted results of MORB glasses from this model (n = 1,931) expand the water content database of global MORBs and indicate a broad existence of high-H₂O MORBs. This new approach allows us to investigate the water content of MORBs from some ridges lacking previous water content measurements (e.g., the Chile Ridge and the Pacific-Antarctic Ridge) and infer changes in the water content of MORB sources through applying the model to transform fault samples.