Predicting crack initiation in highly deformed rail steel; Plasticity modeling, machine-learning-accelerated plasticity simulations, and fatigue criteria

Doctoral thesis, 2026

Rolling Contact Fatigue (RCF) crack initiation, causing major problems for the railway industry, is often associated with large plastic deformations accumulated in the surface layer of rails and wheels. These deformations are known to affect the mechanical and fatigue behavior of the material and to induce anisotropy. Even so, many RCF crack initiation studies do not account for the influence of the deformed near-surface material. Moreover, numerical simulations of the long-term accumulation of plastic deformations under traffic loading are computationally demanding. To address these challenges, the goal of this thesis is to develop methodologies for both more accurate predictions of RCF crack initiation in the highly deformed near-surface rail material and computationally efficient modeling of plasticity.

In the first part of the thesis, a framework is developed for predicting RCF crack initiation in railheads subjected to mixed traffic loading, with a focus on the effects of the highly deformed rail surface layer and profile geometry changes. In particular, an anisotropic material model is calibrated against experiments with railway-like loading at different material states, corresponding to different depths in the railhead. The identified material parameters are then used to consider spatially varying properties in the railhead, governed by an accumulated shear strain distribution reconstructed from measurements in field samples. The framework further incorporates an improved crack initiation criterion that accounts for the accumulated plastic deformations. The findings underline the significant importance of accounting for the deformed near-surface material and rail profile geometry changes when predicting RCF crack initiation.

The second part of the thesis presents two approaches to reducing the computational cost of cyclic plasticity simulations. First, a cycle-domain material model is used to predict peak stress and plastic strain in each loading cycle. Machine Learning (ML) models, Neural Networks (NN) and symbolic regression, are adopted to formulate the per-cycle evolution of plastic strains. Each provides, respectively, high expressiveness and explicit evolution equations. Both ML-based cycle-domain models achieve accurate predictions within and beyond the training regime. As a second approach, an NN-based explicit time integrator is formulated that is capable of taking large strain increments while maintaining accuracy and numerical robustness. The framework is physics-encoded, as it fulfills fundamental plasticity constraints by construction. Material point and finite element simulations demonstrate high accuracy, generalizability, and a considerable reduction in computational time.