Microstructure-sensitive modeling has been developed over the past two decades to predict fatigue crack initiation. Despite the fundamental insights provided by the high-fidelity crystal plasticity models used within these microstructural-sensitive frameworks, they have seen limited transition to standard engineering practice. This is largely due to the fact that significant resources (computational and expertise) are required for crystal plasticity modeling. Machine learning methods can be used to develop surrogate models, but conventional approaches require impractical quantities of training data, limiting their practical application. To address this issue, we developed a method that combines mechanics-based constraints with a generalizable microstructure-sensitive strain energy localization feature. Bayesian symbolic regression (BSR) is then used to produce low-dimensional, interpretable, and explainable fatigue crack initiation models with sparse (low resource) training data. BSR is also used to treat grain-level features, e.g., crystallographic orientations that neighbor crack-initiating pores, as aleatoric uncertainties. This represents a more practical scenario of obtaining specific pore feature geometry from, for example, X-ray tomography or through process modeling, while statistical representations of neighboring crystallographic orientations are known. Finally, to maintain a practical level of required computational resources, we limit the training data to 75 polycrystal simulations. The presented results demonstrate that improved accuracy (as compared to previous literature) can be achieved with sparse training data in a manner that is practical for transition to standard engineering practice.
Current fatigue design regulations employ conservative safety factors and do not account for the influence of concrete composition. Experimental evidence indicates that concretes with similar compressive strength but different compositions can exhibit considerably different fatigue life. Integrating concrete composition into fatigue design therefore has considerable potential to improve the accuracy of fatigue life predictions while maintaining structural safety and facilitating more resource-efficient concrete structures. Against this background, systematic fatigue tests under cyclic compressive loading were conducted on 18 high-strength concretes with varying compositions but comparable compressive strengths, under identical test conditions. For each concrete, fatigue behavior was evaluated at a minimum of four stress levels, resulting in a total of 241 fatigue tests. First, correlation analyses were performed for 16 of the 18 concrete compositions to identify test- and composition-related variables governing fatigue life. While the cement and binder contents exhibited negative correlation coefficients of approximately -0.50, the use of limestone showed a positive correlation coefficient of 0.63. Subsequently, six machine learning models were implemented to predict the number of cycles to failure based on concrete composition. The results demonstrated that the investigated models were able to predict fatigue life with reasonable accuracy. Among the evaluated approaches, neural networks and ensemble-based methods achieved the highest prediction accuracy, with R2 values exceeding 0.95 for the training data and 0.90 for the test data. The predictions were validated using two additional concrete compositions withheld from model development, which showed good agreement between predicted and experimentally determined fatigue lives.
