Front-resolved mode-II–III nominal driving-force evolution and group-validated machine-learning prediction for high-speed railway wheel-tread cracks

Closed wheel-tread cracks can remain compression-dominated during rolling contact while sustaining substantial in-plane and out-of-plane shear. A sequential ABAQUS–FRANC3D–Python framework resolves this response over ten crack-geometry updates. A closure-aware nominal driving force combines event-paired differences in the mode-II and mode-III stress-intensity factors. Four arc-length descriptors summarise its front distribution: mean, peak, peak coordinate and coefficient of variation. Controlled comparisons evaluate initial crack size and orientation, axle load, speed and wheel–rail friction. Thirty Latin-hypercube finite-element conditions provide 4,272 front points for condition-grouped machine-learning validation. ExtraTrees achieves a case-macro cross-validation RMSE of 16.9587 MPa·mm1/2. Four held-out numerical conditions, comprising 579 front points, give a pointwise RMSE of 8.5794 MPa·mm1/2. Front coordinate and initial orientation have the largest mean absolute SHAP contributions. The predicted terminal profiles quantify differences in driving-force magnitude and localisation across conditions within the sampled domain.