Master thesis defense by Pernille Lundsgaard Rasmussen
Title: A Neural Network approach to investigate the thermo-chemical state of the deep interior of the Earth and terrestrial planets
Abstract: An artificial Neural Network (NN) is a machine learning tool able to approximate any given function. It is based on the so-called supervised learning, where the algorithm is presented with a large number of data examples to learn how to map the input to the output. This thesis studies a methodology to infer the thermo-chemical state of the Earth’s mantle from seismic observations, using NNs to approximate the physical relationships. The relationships are based on both thermodynamic computations of mineral physics to obtain the stable minerals constituting the bulk rock, and the corresponding bulk rock elastic properties, and seismological computations to predict the seismic data. However, the computational cost of this model is high. Thus, this thesis aims at developing a faster and sufficiently accurate approximation to the reference model, by training a NN model with synthetic data examples. To this end, two separate NNs approximating the thermodynamic and seismic part, respectively, are constructed in Python using TensorFlow with GPU support. An approach based on grid search is employed to optimize the hyperparameters of the NNs, which are crucial for their ability to learn. Combining the trained NNs into the full NN model results in an approximation which is 13 times faster than the reference model. This significantly reduces the computational cost, and opens the door for inverting data for a large number of seismic stations. As a proof of concept, the NN approximation is tested with a synthetic inversion, employing a probabilistic Markov Chain Monte Carlo (MCMC) method. The test shows that the NN model is sufficiently accurate. Finally, the NN approximation is used to infer the thermo-chemical state of the mantle beneath the seismic station Kongsberg (KONO), using three different inversion methods. One of these methods accounts for the modelling error introduced by the approximation. The inversion results are in accordance with results presented in other studies, and proves that the NN model works well within such an inversion scheme.