Uncertainty quantification for deep learning in geoscience applications

Interpretability and uncertainty quantification is crucial in order to build machine learning models that are robust. One could argue there are at least two aspects where quantifying the uncertainty would help the practitioners. Firstly, it is useful to highlight where the machine learning model needs help, e.g. during the training process, so it has a better prediction accuracy. Secondly, it helps the practitioners to make an informed decision using the final predictions and the accompanying uncertainty measures. In this paper, we will show two geoscience applications, automatic fault prediction and 3D reservoir property prediction, where deep learning has been deployed and the corresponding uncertainty has been captured.