Interactive 3D fault prediction using a weighted 2D-CNN and multidirectional 3D-CNN
Jesse Lomask, Samuel Chambers · 2022
We present an interactive 2D Convolutional Neural Network (CNN) with a weighted loss function. The interpreter is required to pick a minimal number of “fault” and “no fault” picks. These picks are used to weight the loss function of a 2D-CNN. The 2D-CNN is updated, and quickly predicted to maintain tight interactive feedback with the interpreter. Once the result is deemed sufficiently accurate, predictions of the 2D-CNN are fed into a pretrained 3D-CNN to create a novel solution for 3D fault prediction. This new method allows the interpreter to influence the prediction of the CNN by integrating any number of manually drawn picks. This method provides a 3D fault interpretation solution with the crispness and quality of a synthetically pretrained 3D-CNN solution, combined with the added value of interactivity, flexibility, and continual training and updating of a 2D-CNN solution.