First Break, Volume 42, Issue 2, Feb 2024, p. 49 - 56
Summary
We describe a series of experiments designed to improve fault likelihood predictions of a pre-trained machine learning model on an unseen dataset with real fault interpretations. The goal is to establish a best practice workflow for tuning fault prediction models. The model is a 3D model that is tuned by training with a Masked Dice loss function on 2D interpretations. In our experiments we vary the number of interpreted sections in the training set and vary the number of epochs to train the model. We also compare continuous training versus transfer training. To optimise transfer-training we conduct experiments with freezing different parts of the model.