The knowledge of the distribution of local micromechanical fields is crucial in the design of composite materials. Traditionally full-field methods (such as finite element methods) and fast Fourier transformation-based methods are used to obtain the local fields. However, full-field simulations are computationally expensive and time-consuming. Recently, there has been a push toward using the big-data-driven machine learning approaches to estimate the local fields and establish the structure–property linkages. In this work, we use one of the deep learning-based algorithms known as the UNet to predict the local strain fields in a two-phase composite material subjected to uniaxial tensile load. The model is trained and tested on 1200 two-phase microstructures comprising two-volume fraction categories and six different morphological classes. An R² score of 94% is achieved on the test dataset. A detailed statistical analysis is performed to understand the role of the volume fraction and the ratio of elastic moduli of the phases in the deep learning model’s trainability. The insights drawn in this work are then discussed in the context of generating artificial datasets and training a robust predictive deep learning model for localization.

局部微机械场分布的知识在复合材料的设计中至关重要。传统的全场方法（如有限元方法）和基于快速傅立叶变换的方法用于获取局部场。然而，全场模拟在计算上是昂贵且耗时的。最近，有人推动使用大数据驱动的机器学习方法来估计局部场并建立结构 - 属性联系。在这项工作中，我们使用一种称为 UNet 的基于深度学习的算法来预测受到单轴拉伸载荷的两相复合材料中的局部应变场。该模型在 1200 个两相微观结构上进行了训练和测试，包括两个体积分数类别和六个不同的形态类别。安在测试数据集上实现了 94% 的R ²分数。执行详细的统计分析以了解体积分数和相的弹性模量比在深度学习模型的可训练性中的作用。然后在生成人工数据集和训练用于本地化的强大预测深度学习模型的背景下讨论在这项工作中得出的见解。