This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 31680, “Successful Development and Deployment of a Global ROP-Optimization Machine-Learning Model,” by Timothy S. Robinson, SPE, Exebenus, Peter Batruny, SPE, Petronas, and Dalila Gomes, SPE, Exebenus, et al. The paper has not been peer reviewed. Copyright 2022 Offshore Technology Conference. Reproduced by permission.
Drilling rate of penetration (ROP) is influenced by many factors, both controllable and uncontrollable, difficult to distinguish with the naked eye. Thus, machine-learning (ML) models such as neural networks have gained momentum in the drilling industry. Previous models were field-based or tool-based, which affected accuracy outside of the trained field. The authors of the complete paper aim to develop one generally applicable global ROP model, reducing the effort needed to redevelop models for every application.
The authors have identified a need for an ROP model that can recommend parameters in real time, which ideally requires a general ROP model that can be applied with low prediction errors.
Formation properties are available in real time from logging tools; however, incorporating logging data into a global ROP model is challenging. A distance usually exists between the bit and logging tools where the properties of the formation drilled in real time at the bit are not available. While strides have been made in ML interpretation of logs, human interpretation is still needed.
A deep neural network was used for modeling ROP. This model was queried as a black box within the scope of a constrained optimization algorithm tasked with identifying combinations of controllable drilling parameters expected to increase ROP. An overview of the ROP optimization system powered by the ROP estimation model is shown in Fig. 1. The ROP model was fitted using training and validation data sets composed of data from independent wells. Four additional wells were held in reserve for final testing purposes, most importantly for assessing the model’s ability to generalize to completely unseen data. For optimization, a three-stage conditional parameter sweep was used where the value expected to optimize ROP was found sequentially for each of the controllable parameters while fixing the values of the other variables. The optimizing value for each parameter was used as a fixed value for the subsequent optimization steps. This approach maintains the simplicity of a grid search and converges to a local optimum quickly because of the relatively small number of search iterations required under this formulation.
A variety of well types with global geographical diversity was present in the data, using different downhole drives and bottomhole assemblies. Twenty-nine wells were used for training the ROP model, with an additional five wells used for validation and a further four wells reserved for holdout testing. No overlap existed between the wells constituting the training, validation, and holdout data sets.
The drilling data sets were preprocessed by methods detailed in the complete paper. To ensure consistency across all operations, an average ROP was calculated independently based on depth measurements and their associated timestamps. This ROP calculation was used as the target variable for the ROP estimation model and for assessing its predictive performance.
Training and validating the neural network for ROP estimation using data from a variety of independent wells is very useful for assessing model generalization. This technique is superior to the common approach of randomly splitting data sets because it minimizes information leakage between training and validation data.