关键词: 随钻参数, 物理信息神经网络（PINN）, 机器学习, 岩石强度, 随钻探测

Abstract: Accurate determination of rock mechanical parameters is crucial for engineering design. While-drilling technology has become one of the primary approaches for assessing rock strength. Existing while-drilling strength prediction models typically rely on a single approach, such as theoretical physical analysis or machine learning. As a result, their predictive performance is highly sensitive to field operating conditions and data quality. This study develops a physics-informed neural network (PINN) model to predict the uniaxial compressive strength of rock from while-drilling parameters. The model is built on a CNN-LSTM-Attention fusion architecture. It achieves accurate prediction of rock uniaxial compressive strength by integrating physical knowledge with data-driven learning. Specifically, the physical model for strength prediction based on multi-wing cutting drilling is transformed into constraint conditions within the network. Based on parameter sample construction and comparative analysis, the modified physical model improves rock strength prediction accuracy by 15%. Its correlation coefficient with the measured rock strength reaches 0.97. Compared with the 6 traditional neural network models, the determination coefficient of the strength prediction results of the PINN model on the test set is increased by at least 2%, and the mean absolute error is reduced by at least 9%. Further validation using actual while-drilling data from composite formations shows that, when the lithological sequence is granite, conglomerate, and sandstone, the average prediction errors of the PINN model are 3.63, 4.44 and 1.19 MPa, respectively. In all cases, the percentage error is less than 8%, demonstrating the model’s strong robustness in predicting the strength of formations with different lithologies. These findings provide practical guidance for tunnel while-drilling technology and the real-time automated characterization of rock mass strength.

Key words: while-drilling parameters, physics-informed neural networks (PINN), machine learning, rock strength, while-drilling detection