Deep learning driven prediction of dynamic stress-strain response in limestone: insights into transient mechanical behavior under complex loadings for shield tunneling

The complex geological conditions pose a serious challenge to improving the rock-breaking efficiency of shield cutter machines in subway tunnel constructions. This study develops a three-axis thermal-hydraulic-mechanical (THM) coupled dynamic impact system, which is integrated with a scaled shield cutter model to generate a unique dataset of stress wave propagation under multi-field coupling. A bidirectional long short-term memory (LSTM) neural network with an attention mechanism is proposed to establish a nonlinear time mapping relationship between stress waves. The determination coefficient (R²) exceeds 0.97, the symmetric mean absolute percentage error (sMAPE) is less than 10 %, and the average relative uncertainty (ARU) is less than 4 %, confirming its high accuracy and reliability. Based on predictions and test results, the new quantitative laws for the transient dynamics of limestone are further revealed. The results reveal that loading conditions do not alter the correlation trend between loading rate and dynamic parameters but significantly influence the degree of the loading rate's effect. Axial pressure dominates energy absorption with an energy contribution rate of 60 %. Confining pressure amplified the sensitivity of loading rate by 170 %, while THM coupling suppressed the dynamic deformation modulus by over 30 %. The accurate prediction of the nonlinear response of stress waves in limestone by the LSTM neural network establishes the connection between transient dynamic observation and rock fragmentation physics mechanism, providing support for quantifying energy absorption and conversion in the rock fragmentation process, and providing key strategies for optimizing shield machine performance in extreme environments.