Towards drilling rate of penetration prediction: Bayesian neural networks for uncertainty quantification

Abstract

Drilling rate of penetration (ROP) prediction has long been a part of any drilling activity. By accurate prediction of ROP, optimization can be done that maximizes ROP and reduces drilling costs. ROP prediction relies on good quality data from nearby wells and is mostly performed through multiple-regression models. Recently, machine learning and artificial intelligence tools have been suggested as an alternative. On the other hand, the lack of knowledge about the uncertainty in these data-driven models can limit their applicability. A trained machine learning model can perform satisfactorily on the data on which it was trained, yet fail when applied to new data. In this work, we explore the application of Bayesian neural networks and the notion of uncertainty in ROP prediction. We compare the prediction of several deterministic and probabilistic models on a real drilling dataset. Despite accurately predicting the ROP in a 5000ft section of a horizontal well, the models are shown to have large uncertainties, potentially caused by the size of the training dataset. The results and analysis show how the adoption of a probabilistic framework can be leveraged to pinpoint the root cause of an ROP model’s lack of accuracy.