Automated Deep Learning (AutoDL) for Facies Prediction: Implementation and Strategy

Machine Learning (ML) and Deep Learning (DL) have shown promising results for classification and regression tasks. However, the current conventional implementation is heavily based on hyperparameter tuning and architecture design, which needs sufficient time and effort for trial-and-error and sophisticated expertise in ML/DL. An emerging framework that promises a high-quality ML/DL without human assistance is called Automated Machine Learning (AutoML) or Automated Deep Learning (AutoDL). A particular AutoDL framework, namely Auto-Keras, is chosen for the study. Auto-Keras is based on Bayesian optimization, which helps the network for effective network morphism, which leads to more efficient neural architecture search (NAS). This AutoDL approach is then implemented for facies prediction on the wells from the North Sea, where the results show that AutoDL results are superior to conventional ML results. These results can be used as an initial guess before the geologist studies the core samples or to predict facies in uncored wells. Additionally, geographical data distribution and the chosen scaler (Standard or MinMax) are crucial for producing the best possible prediction. The distribution of the facies, either relatively more homogeneous or heterogeneous, is also discussed within the study, where each of these cases has suitable strategies for AutoDL implementation.