A new approach is proposed for rapid and accurate prediction for composite failure in combination of the phase field and machine-learning methods. First, using experimentally-fitted tangent modulus instead of elastic modulus as constitutive relationship, a modified phase field method (MPFM) is established for the crack propagation and mechanical response, which can be effectively applied for composites with a nonlinear constitutive relationship. Interestingly, both the crack propagation path and mechanical responses of two typical examples of composites using MPFM are well consistent with previously available experimental and calculated ones. Furthermore, the data-driven back propagation neural network (BPNN) is constructed to greatly accelerate the prediction on a database generated by MPFM, emphasizing several critical parameters, for example, fiber orientations, external load, maximum failure strain, and critical strain energy release rate. Of much importance, the well-trained BPNN builds a bridge between the traditional computational fracture mechanics and machine learning algorithms, enabling non-specialists to accurately calculate the mechanical response of composites, moreover, saving over 99% of computing time in comparison with MPFM.

中文翻译：

---

用于快速准确预测复合材料失效的相场和加工学习方法

提出了一种结合相场和机器学习方法来快速准确预测复合材料失效的新方法。首先，使用实验拟合的切线模量代替弹性模量作为本构关系，建立了裂纹扩展和力学响应的修正相场法（MPFM），该方法可有效应用于具有非线性本构关系的复合材料。有趣的是，使用 MPFM 的两个典型复合材料示例的裂纹扩展路径和机械响应与之前的实验和计算结果非常一致。此外，构建了数据驱动的反向传播神经网络（BPNN），以大大加速对 MPFM 生成的数据库的预测，强调几个关键参数，例如纤维取向、外部载荷、最大失效应变和临界应变能释放速度。更重要的是，训练有素的BPNN在传统计算断裂力学和机器学习算法之间架起了一座桥梁，使非专业人士能够准确计算复合材料的力学响应，而且与MPFM相比节省了99%以上的计算时间。