Probabilistic tomography: from seismic waveforms to thermo-chemical models using neural networks
04 / 2010 - 12 / 2015
Knowledge of the solid state convection in the Earth's mantle is key to understand plate tectonics and its thermo-chemical evolution over geologic times. Seismic tomography has provided important information, but an integrated understanding of seismic tomography, mantle thermodynamics and chemistry is still lacking, although many attempts have been made. I propose such an integrated study based on significant advances on several fronts, possible because of the explosion of high performance computing. I will take advantage of recent breakthroughs in computational seismology, which allows to solve the wave equation in arbitrary complex media, in non-linear inversion based on neural network modelling and in computational mineral physics which allows to calculate the thermodynamic properties of minerals at mantle temperatures and pressures. Models are not perfect, and any interpretation or integration needs to quantify the uncertainties. At each step full probability density functions will be evaluated characterizing the complete knowledge to be gained on a given parameter from the data. The novelty will be to make the first long wavelength tomographic model based on a fully non-linear inversion of complete seismograms. This will allow to infer absolute values for all model parameters, rather than perturbations from uncertain reference models. Absolute values of elastic models will then be converted into absolute thermo-chemical models using a recent self-consistent equation-of-state, a new mineral physics data base and neural networks. These thermo-chemical models represent the present-day Earth structure and will be crucial to identify all possible scenarios of the Earth's evolution over geologic times.