Machine learning assisted seismic assessment of rock crack propagation and ground instability around underground passageways

Seismic instability around underground passageways is governed by rock cracking, ground deformation, groundwater condition, support response, and earthquake demand. This study presents a finite element (FE) and machine learning (ML) framework for seismic screening of rock and soil transition zones. Abaqus 2023 was used with a 200 m by 120 m plane strain domain, CPE4R elements, B21 lining elements, T2D2 support elements, and COH2D4 crack interfaces. Mesh and time increment studies limited changes in convergence, bending moment, crack extension, and plastic zone radius to less than 1.0% between adopted and finer analyses. The internal database contains 1536 numerical cases from 16 geotechnical condition groups, 24 earthquake records, and four intensity levels. Motion separated fitting, validation, and locked test blocks contain 1152, 192, and 192 cases. A separate 128 case numerical holdout uses eight geotechnical configurations and four earthquake records with no shared condition or motion identifiers. Laboratory material records support parameter assignment, while centrifuge tunnel response observations are retained separately for qualitative physical response checking and are excluded from ML development. Nine continuous FE responses and one categorical stability target are evaluated. Linear regression was selected from the fitting and validation stage before locked test evaluation and subsequently gave a mean locked test coefficient of determination (R 2 ) of 0.979 and a four fold geotechnical group mean R 2 of 0.951. Across the numerical holdout, R 2 values range from 0.921 to 0.960, with a mean of 0.946. The selected support vector classifier gives locked test accuracy of 0.969, geotechnical group F1 score (F1) of 0.909, and holdout accuracy of 0.805 with receiver operating characteristic area under the curve of 0.896. Sigmoid calibration gives a multiclass Brier score of 0.319 and expected calibration error of 0.118. Target specific grouped permutation analysis identifies the seismic intensity block as producing the largest predictive performance loss within the investigated holdout, followed by groundwater depth and response dependent contributions from support stiffness, burial depth, and crack density. The framework provides a documented mechanics informed screening approach within the investigated numerical parameter domain.