Applications of machine learning to mineral physics data and the inference of the thermochemical structure of the Earth's mantle.

The physical and chemical properties of the Earth’s mantle govern the cause of natural disasters, such as earthquakes and volcanoes. Since we do not have direct access to mantle materials, their properties are often inferred from laboratory measurements and surface observations (e.g. seismic data from earthquake recordings). This thesis addresses some key problems we face while utilising these data to constrain the thermal and chemical properties of the mantle. Firstly, we propose a data-driven approach based on machine learning to explain the laboratory measurements and quantify their uncertainties in the absence of an adequate physical model. Our results show that although conventional approaches based on fitting the measurements to an assumed model may appear better constrained, they could potentially provide biased results. Secondly, we use the data-driven approach to explore which thermochemical parameters can be constrained (and to what extent) with limited seismic observables- wave speeds and density. Our results show that these observables constrain temperature and major chemical parameters (silicon, magnesium, and iron), and they indicate the presence of thermochemical heterogeneities at the lowermost mantle. The dense and slow piles at the bottom of the lower mantle seen in seismic data can be explained by an enrichment in silica and iron content- characteristic feature of enstatite chondrites. The inferred heterogeneities have profound implications for the dynamics of the mantle and outer core. The methodology developed in this thesis is extremely efficient. It can easily incorporate additional observables and thus, has wide applications in the seismology and mineral physics community.