Seismic signatures: mixing with a tomographic filter and identifying with cluster analysis

During the past 2 decades, data coverage and methodological developments have considerably improved the resolution of seismic tomography maps, refining our mapping of the deep Earth’s mantle structure. Nevertheless, the uneven distributions in sources (the earthquakes) and receptors (the seismic stations) leads to non-uniqueness of the solution and requires the prescription of a priori information (mostly damping and smoothing), the effect of which is to smear out seismic images and degrade their effective resolution. Alternatively, statistical quantities have been used to investigate the nature, purely thermal or thermo-chemical, of the structures observed by seismic tomography. In particular, it has long been recognized that the statistical distribution of shear-velocity anomalies (dlnV_S) in the lowermost mantle shows some degree of asymmetry in the form of a slow velocity tail, and that this slow tail is associated with the large low shear-wave velocity provinces (LLSVPs), the prominent feature on lowermost mantle tomographic maps. This bimodal distribution appears from around 2200 km and persists towards the deeper mantle. Yet, the phase transition to post-perovskite (PPv) at depth ~2700 km, if not happens globally, implies a trimodal distribution for dlnV_S. Here, we bring new insights on these questions. First, we investigate the effect of the seismic tomography ‘operator’ on seismic velocity anomalies triggered by different possible lowermost mantle thermo-chemical structures. For this, we first run simulations of thermal and thermo-chemical convection including or not the post-perovskite phase, and we calculate synthetic velocity anomalies predicted by these simulations. We then apply to these synthetic velocity anomalies a tomographic filter built for the tomographic model HMSL-SP06. We show that seismic signatures corresponding to different materials (regular mantle, thermo-chemical piles and PPv) are clearly distinct on statistical distribution of unfiltered shear-and compressional velocity anomalies, dlnV_S and dlnV_P, but get mixed or partially mixed after applying the filter. Interestingly, for synthetic velocity anomalies built from thermo-chemical simulations, a low velocity tail clearly appears on dlnV_S histograms, but not on dlnV_P histograms, similar to what is observed in real seismic tomography maps. For synthetic velocity anomalies built from purely thermal simulations, dlnV_S histograms do not feature any low velocity tail, and distribution histograms for both dlnV_S and dlnV_P are fairly Gaussian. Overall, our results therefore support the hypothesis that the LLSVPs observed at the bottom of the mantle are composed of hot, chemically differentiated material. They further show that the mixing of seismic signatures due to tomographic filter, implying the statistical distribution of dlnV_S and dlnV_P may be richer and more complex than it appears to be from seismic tomography models. Acknowledging the mixing of seismic signatures inherent to tomography models, we then apply cluster analysis with trimodal distribution to four recent tomographic models: GLAD-M35, REVEAL, SPiRaL-1.4, and TX2019slab.  We identify three velocity clusters, slow, neutral, and fast, which we associate with thermo-chemical piles, regular mantle, and PPv. Based on this analysis, we provide a probability map of the three clusters, which may be used to better understand the lowermost mantle structure and facilitate future geodynamic studies.