Machine Learning–Based Imaging of the Upper Mantle and Transition Zone Using Fundamental and Overtone Surface Wave Dispersion within an Integrated Geophysical–Petrological Framework

Surface wave dispersion data, encompassing both fundamental modes and higher overtones, provide powerful constraints on the thermochemical structure of the upper mantle and mantle transition zone. Fundamental modes are predominantly sensitive to shallow structures within the lithosphere and upper asthenosphere, whereas higher overtones sample progressively greater depths, offering enhanced sensitivity to the mantle transition zone and the uppermost lower mantle. The joint utilization of fundamental and overtone dispersion therefore enables improved resolution of key mantle features, including the 410 km and 660 km discontinuities, and variations in thermal and compositional structure across depth.

We carried out an extensive sensitivity analysis. The results demonstrate that both fundamental and overtone dispersion curves exhibit strong sensitivity to upper mantle structure, with particularly pronounced responses at the 660 km discontinuity, where our thermodynamic models predict sharp contrasts in seismic velocity and density. In the uppermost lower mantle, extending to depths of approximately 1500 km, fundamental modes are significantly affected only at long periods (>200 s), whereas higher overtones show substantial sensitivity across a broad period range (20–150 s) with different behaviour for Rayleigh and Love waves.

A new machine learning strategy embedded within a thermodynamically self-consistent geophysical–petrological framework allows us to efficiently link thermochemical crustal and mantle structure and surface wave dispersion data (fundamental mode and overtones) preserving physically consistent relationships among temperature, composition, seismic velocities, and density. The machine learning algorithm is incorporated into an inversion strategy to image lithospheric, asthenospheric, transition zone and uppermost lower mantle thermochemical structure accounting for the topography associated with the 410 km and 660 km mineral phase transitions in a consistent manner.

These results highlight the critical role of overtone data in complementing fundamental mode observations and demonstrate that machine learning–based imaging substantially enhances the resolution of mantle transition zone models, particularly when the uppermost lower mantle is incorporated consistently within thermochemical frameworks. The machine learning framework also facilitates the incorporation of complex, non-linear relationships between seismic data and thermochemical properties.