A novel method to predict the mechanical responses of arbitrary microstructures from the deep learning of microstructures and their stress-strain response is presented in this work. Two-phase microstructural images that consist of different grain sizes and compositions are generated and quantified using the two-point statistical homogenisation scheme. Finite element (FE) simulations are used to predict the in-plane elastoplastic response of the generated microstructures. To minimize the computational efforts, microstructures and the stress-strain data are projected into the lower order orthogonal spaces by using the principal component analysis (PCA). Effective methods to visualise and understand the distribution of microstructure-response data in the transformed dimensional space are presented in detail. The reduced order statistically homogeneous microstructures along with the reduced stress-strain data are learned by using the convolutional neural networks (CNN). A new set of randomly generated microstructures are fed into the trained convolutional network to predict the stress-strain response. The derived failure strength and modulus of the predicted response curves are showing a scatter index of 1.74% and 10.53% against the true FE predicted values. The mechanical responses of randomly generated two-phase fibre reinforced plastic (FRP) composite microstructures are predicted using the developed deep learning model. Thus, the proposed strategy can predict the mechanical properties of arbitrary microstructural design with better accuracy and minimal computational effort.

这项工作提出了一种从微观结构的深度学习及其应力应变响应中预测任意微观结构力学响应的新方法。使用两点统计均质化方案生成和量化由不同晶粒尺寸和成分组成的两相微观结构图像。有限元 (FE) 模拟用于预测生成的微结构的面内弹塑性响应。为了最大限度地减少计算工作，微结构和应力应变数据通过使用主成分分析 (PCA) 投影到低阶正交空间。详细介绍了可视化和理解微结构响应数据在变换维度空间中分布的有效方法。使用卷积神经网络 (CNN) 学习降阶统计均匀微观结构以及减小的应力应变数据。一组新的随机生成的微结构被输入到训练好的卷积网络中以预测应力-应变响应。预测响应曲线的衍生失效强度和模量显示相对于真实 FE 预测值的分散指数为 1.74% 和 10.53%。使用开发的深度学习模型预测随机生成的两相纤维增强塑料 (FRP) 复合微结构的机械响应。因此，所提出的策略可以以更好的精度和最少的计算工作量预测任意微结构设计的机械性能。