A Data-Driven Framework for Predicting Drilling Speed Based on the Kolmogorov–Arnold Network: A Case Study of Norway’s Volve Oilfield

Abstract This study applies the Kolmogorov–Arnold network (KAN) to mechanical drilling rate prediction. Based on publicly available drilling data from Norway’s Volve oil field, it establishes a predictive methodology integrating accuracy forecasting with symbolic analysis. Compared to traditional machine-learning and deep-learning approaches, the KAN network positions learnable activation functions on its connection edges. It achieves complex function approximation through linear combinations of B-spline functions, offering advantages such as mathematical interpretability and grid extension techniques. Results indicate the KAN model achieved a test R 2 of 0.9520 with a test average absolute percentage error of 4.53%. The optimized KAN model achieved an R 2 of 0.9626 with a test average absolute percentage error of 2.93%, while symbolic regression revealed coupling relationships among drilling parameters. This work pioneers new technical pathways for applying interpretable artificial intelligence in petroleum drilling engineering.