Stop Using Black-Box Models: Application of Explainable Artificial Intelligence for Rate of Penetration Prediction

Rate of penetration (ROP) prediction plays a crucial role in optimizing drilling efficiency and reducing overall costs in the petroleum industry. Although modern artificial intelligence (AI) models have shown promising performance in this task, their lack of interpretability hinders their practical applications. This study introduces the neural basis model (NBM), a self-explainable model, for ROP prediction. The adopted NBM is benchmarked against some well-known methods using a publicly available data set, demonstrating its promising performance. The key advantage of the NBM lies in its ability to provide clear explanations, where the influence of the input drilling parameters on the predicted ROP can be clearly visualized and analyzed. The study also compares the models’ performance under two scenarios—continuous learning and all-for-one. The results indicate that the models’ performance under the continuous learning scenario, in which the models are iteratively updated with new data from the same well, outperforms that under the all-for-one scenario, where models are trained on data from previously drilled wells and applied to new wells. The discrepancy is probably due to the absence of detailed formation characteristics from the data set. Therefore, the model trained in other wells could not generalize well on new wells. Currently, in petroleum engineering, AI is gradually playing an increasingly important role; however, the majority of AI-related works often directly employ black-box models, which lack interpretability and might cause serious risks if deployed in practice. This work introduces an approach to using neural networks to build self-explainable AI (XAI), with the aim of promoting the application of XAI in the petroleum industry.