Search

Three-dimensional models of Earth’s seismic structure can be used to identify temperature-dependent phenomena, including mineralogical phase and spin transformations, that are obscured in 1-D spherical averages. Full-waveform tomography maps seismic wave-speeds inside the Earth in three dimensions, at a higher resolution than classical methods. By providing absolute wave speeds (rather than perturbations) and simultaneously constraining bulk and shear wave speeds over the same frequency range, it becomes feasible to distinguish variations in temperature from changes in composition or spin state. We present a quantitative joint interpretation of bulk and shear wave speeds in the lower mantle, using a recently published full-waveform tomography model. At all depths the diversity of wave speeds cannot be explained by an isochemical mantle. Between 1000 and 2500 km depth, hypothetical mantle models containing an electronic spin crossover in ferropericlase provide a significantly better fit to the wave-speed distributions, as well as more realistic temperatures and silica contents, than models without a spin crossover. Below 2500 km, wave speed distributions are explained by an enrichment in silica towards the core-mantle boundary. This silica enrichment may represent the fractionated remains of an ancient basal magma ocean.

Earthquake prediction has long been thought to be nearly impossible. However, machine learning gives us new possibilities to investigate. Machine learning has already been successfully applied to "lab-quakes", but these are much more periodic than real earthquakes. Real earthquake physics are often poorly understood, so to get a better understanding of the ml techniques it is important to also use data that is well understood. To test whether machine learning methods can also be used for non-periodic data, they are applied to simulated turbulence data. This data is used because the physics are fully understood and it gives a nice proxy to earthquakes. Machine learning techniques show promise on this dataset to predict when the next failure will occur. This prediction is most accurate when only medium-low frequencies of the acceleration data are used during prediction. This shows that it might be possible to apply these techniques to real earthquakes as well, as long as the data is properly preprocessed so that the machine learning models can extract the useful information.