Seismic inversion with gradient boosting

Marcelo Guarido, Luping Qu, Zhan Niu, Kai Zhuang, David J. Emery, Daniel O. Trad, Kristopher A. Innanen

Geophysics in the Cloud promoted a competition with the proposal to use machine learning algorithms to perform seismic inversion. The provided data showed to be challenging, due to the small number of well logs to train the model. During the competition, most of the competitors based their strategies using deep learning models (CNN, RNN, LSTM, etc). Our proposal was to evaluate the possibility of seismic inversion using more classical and simple approaches, hence the choice of the Gradient Boosting algorithm. Predictions of P and S impedances and density using an XGBoost model are stable and strongly rely on the trend features and use seismic traces to include low to mid-frequency content. Testing neural networks structures from Scikit-Learn and Tensorflow resulted in noisy and spiky inversions, pointing to overfitting and their instability drew attention to the issue of the small number of logs for training.