Physics-guided deep learning for seismic inversion

An important part of the geophysical analysis is seismic inversion, which provides insight into the Earth's structure. Conventional seismic inversion methods require enormous computational power, human intervention, long processing time, and uncertainties in predictions. With exponential growth in the Artificial Intelligence field, particularly in the Deep Learning field, there is a possibility to apply various Deep Learning and Machine Learning algorithms on multiple seismic inversion problems. Here Convolution Neural Networks were used to create a model to directly estimate porosities from seismic data. By teaching the model exact physics, the need for large labeled data was removed. The purpose of the thesis is to give a reader all the information to create the model without strong prior knowledge of exploration geophysics and deep learning algorithms. In the methodology chapter, first, basic information on seismic exploration, seismic inversion, and problems related to it is written. It is followed by an introduction to artificial intelligence, machine learning, and deep learning methods. How deep learning models are created and what their parameters mean. In the end, the model is introduced and all its elements are explained. The model was updated multiple times to find the best predictions and the models’ accuracies were shown for each of them. The advantages and disadvantages of this and other models were discussed.