Comparing 28 global P- and S- wave tomography models by Machine Learning analysis for the interpretation of the Earth’s mantle structures

Seismic velocity anomalies observed in the mantle can have several origins, the main contributions being anomalies of temperature and composition. The difference between P- and S-wave models has been used to separate thermal and compositional contributions in imaged seismic structures and identify large-scale compositional heterogeneity in the Earth's mantle. According to our two-step Machine Learning (ML) analysis of 28 P- and S-wave global tomographic models, P- and S-models differences are not intrinsic and can be reduced by changing the models in their respective null spaces. Because we find, P- and S-wave images of mantle structure are not necessarily distinct from each other, a purely thermal explanation for seismic structure is sufficient at present; significant mantle compositional heterogeneities do not need to be invoked. In this study, 28 commonly used tomographic models are examined, ranging from ray theory (e.g., UU-P07, MIT-P08) to Born scattering (e.g., DETOX) and full-waveform techniques (e.g., CSEM, GLAD). Combined Varimax Principal Component Analysis is used to reduce the dataset's dimensionality (by 82%) while preserving the relevant information of each tomographic model (94% of the original variance). Reduced-sized models are followed by a hierarchical clustering analysis (HC) using Ward’s method to categorize all the models into a hierarchy of groups based on their similarities. HC divided the set of tomographies into two main clusters: the first cluster, which we named "Pure P-wave", is composed of six P-wave models that only use longitudinal body wave phases (e.g., P, PP, Pdiff); the second cluster "Mixed" includes both P- and S-wave models; P-wave models in this cluster use inversion methods that include inputs from other geophysical and geological data sources, that cause them to be more similar to S-wave models than to pure P-wave models without a significant loss of fitness to P-wave data. Results suggest that the differences between some individual P-wave and S-wave models are smaller than the differences between grouping of models that are only P-wave or S-wave. These variable differences clearly convey that no consistent separation exists between the P- and S-wave models. We have also calculated the Distance Matrices along the Principal Components. Comparing clustering results with Distance Matrices shows that the differences between the "Pure P-wave" and "Mixed" clusters are mainly in the upper mantle. Accordingly, our results indicate that P-wave structures do not need to be very distinct from a thermal interpretation of S-wave structures and support a relatively “Homogenous” mantle.