Machine learning for computational fracture and damage mechanics— Status and perspectives

Fracture and damage mechanics have evolved remarkably from simple, yet useful Linear Elastic Fracture Mechanics (LEFM) to relatively modern techniques such as Cohesive Zone Model (CZM) and phase-field approaches. The advent of computational power allowed researchers and engineers to conduct high-fidelity numerical simulations to model complex fracture mechanisms in advanced materials and structures. Nonetheless, large-scale fracture simulations remain computationally intensive, particularly under loading conditions such as impact and extreme environments. In this context, Machine Learning (ML) techniques have seen a surge in their use for mechanics and computational simulations. In this perspective article, we review the existing research landscape in the recent literature on the application of ML to fracture and damage modelling across different material and structural classes. Specific focus is placed on classifying the ML approaches adopted to model or predict fracture behaviour, followed by an extensive discussion on the challenges and limitations of such approaches. Future directions are proposed with an emphasis on the generality, interpretability and reliability of the ML models. We believe the article serves as a guidance document for engineers and scientists involved in the developmental process of Artificial Intelligence (AI)-driven fracture modelling tools. • A state-of-the-art review of machine learning techniques applied to fracture and damage modelling is presented. • Critically evaluates deep learning architectures, including GNNs, CNNs, PINNs, Neural Operators, and Generative Models. • Key challenges lie in high-fidelity dataset generation, generalisation, model scalability and predicting complex crack geometries. • Advocates a roadmap prioritizing physics-informed constraints, open-source reproducibility, and uncertainty quantification (UQ) for reliable simulation.