ZHANG Peng, ZHANG Yongran, TANG Jingqi, TANG Keke
Phase-field fracture models can describe complex crack evolution without explicitly tracking crack surfaces, but the fine-mesh requirement induced by small length scales, the nonlinear coupling between the displacement and phase fields, and incremental-iterative solution procedures lead to high computational costs. To alleviate these bottlenecks, deep learning has been progressively introduced into full-field response prediction, local subproblem solution, and parametric model construction, providing multiple routes for accelerating and reformulating phase-field fracture computation. After outlining the variational basis and numerical solution procedure of phase-field fracture, this paper classifies existing methods into four categories according to where neural networks intervene in the computational workflow and which components they primarily replace: fully data-driven surrogate models, finite element-neural network hybrid solvers, physics-driven neural solvers, and neural operators. Drawing on representative studies, public benchmarks, and computational implementations, these methods are compared in terms of data and physical constraints, offline investment and online benefits, generalization, and engineering reliability. The analysis indicates that deep learning does not uniformly replace conventional numerical methods but reallocates the costs of data generation, model training, physical modeling, and online solution for different tasks; the applicability of a specific method therefore depends on data conditions,query scale, and reliability requirements.Future research should advance standardized benchmarks, unified full-cost evaluation, and reliability validation based on key mechanical quantities, while promoting the integration of neural networks with conventional numerical methods to progressively establish a comparable, verifiable, reproducible, and engineering-oriented intelligent fracture computation framework.
