Predicting Vp/Vs in North America with a Machine Learning Approach

Vp/Vs (or the Poisson's ratio) provides critical information to constrain the bulk crustal composition, stress state, and tectonic evolution. The crustal Vp/Vs beneath the seismic station can be effectively determined by the receiver function H-κ stacking technique. However, complex crustal structures often cause large uncertainty in Vp/Vs measurements. Additionally, Vp/Vs observations are sparse in many regions of the world due to the uneven distribution of seismic stations. While interpolation methods have been widely applied to obtain a continuous distribution of Vp/Vs, these results are not accurate enough and may be strongly biased by the interpolation artifacts. Therefore, reliable mapping of the Vp/Vs variation in the crust remains a major challenge in seismological study.

We present a machine learning approach to estimate the Vp/Vs using multiple geophysical datasets. This approach assumes that the Vp/Vs is related to the physical and chemical properties of the crust. Specifically, we implement a Gradient Boosted Regression Tree algorithm (XGBoost) to develop an optimal prediction model for North America. To train the model, we use Vp/Vs as the target values and a compilation of geophysical observations as the predictor variables. These measurements are composed of two types of data: (1) continuous data, such as crustal velocities and gravity anomalies; (2) categorical data (tectonic type). We use 80% of the measurements to train the model, and the remaining 20% to validate the model. We first examine the reliability of this method by predicting Vp/Vs for the United States, where extensive geophysical data are available. Overall, the model achieves a high R² value of 0.88 in all measured results versus prediction results, indicating robust prediction results at most locations. In the second, a more challenging test, we predict Vp/Vs for Canada where measurements are sparse and uneven, and the amount of data is only 14% of that in the United States. The results show an R² value of 0.87 between the measured and predicted values. Feature importance analysis indicates that crustal shear wave velocity and tectonic type contribute most significantly to reducing the loss function. The prediction results show that bulk Vp/Vs varies between 1.70 and 1.90 across Canada, with a mean value of 1.82.  The cratonic regions generally exhibit high (>1.80) Vp/Vs with a relatively small (<0.05) variation, whereas Proterozoic orogens are characterized by a large Vp/Vs variation from 1.75 and 1.90. The lowest Vp/Vs of 1.66 is observed in the Phanerozoic orogenic belts of Cordillera, contrasting sharply with the basement rocks (~1.90) beneath the Alberta foreland basin. Overall, our study highlights the capability of the machine learning in discovering complex relationships between multi-dimensional geophysical data sets and resolving crustal Vp/Vs in continents globally.