Deep Neural Networks for SS Wave Database Construction: Refining the Iceland Mantle Plume Location in the Deep Mantle

Abstract SS precursors are crucial for probing mantle discontinuities and thermal structure but constructing high‐quality SS datasets—essential for precursor analysis—remains labor‐intensive due to limited automated solutions. While deep learning has advanced P‐ and S‐wave analysis, its application to systematic SS phase selection remains underexplored. We present PhaseSelectNet (PSNet), a deep neural network integrating Time‐Net and Polarity‐Net to facilitate data set construction. Polarity‐Net—the first automated tool for determining SS polarities—ensures reliable stacked SS precursors by preventing polarity misjudgments, which can distort waveforms. Trained on a global data set composed of high‐quality and low‐quality SS waveforms, Time‐Net and Polarity‐Net both demonstrate the capability to accurately identify high‐quality SS seismograms, while respectively provide precise polarities and arrival times. Our application of PSNet to the Iceland region—where a mantle plume interacts with the Mid‐Atlantic Ridge (MAR)—has revealed mantle discontinuity topography extending beyond previous observational limits constrained by seismic station coverage. The results identify the thinnest portion of the mantle transition zone (MTZ) southeast of Iceland, suggesting this region as the probable upwelling center at MTZ depths. This finding provides new insights into Icelandic key tectonic characteristic: the systematic southeastward migration of its rift system over the past 24 Myr, which has produced a distinct curvature in the otherwise linear trajectory of the North Atlantic Ridge. The continuous flow of high‐temperature magma, originating from a southeastern mantle upwelling along the base of the lithosphere, triggered thermomechanical weakening preferentially southeast of existing rift structures, systematically shifting the rift axis over time.