Due to the high demand for materials with superior mechanical properties and diverse functions, designing composite materials is an integral part in materials development. However, due to the heterogeneity and complexity of composites, measurements of properties and designs of optimal structures are often experimentally or computationally intractable. Here we report complete strain and stress tensors predictions given input composite geometries, based on a deep learning approach. The model not only predicts all strain and stress tensor components accurately, but also observes rigorous continuum mechanical principles. Furthermore, through tuning data statistics, the model performance is enhanced even with a small amount of data in terms of handling material microstructures with varying constituent ratios and hierarchical geometries. The method vastly improves the efficiency of predicting comprehensive mechanical behaviors of composite materials and provides a powerful tool for design problems such as multifunctional composites design and hierarchical structures optimization.

由于对具有优异机械性能和多种功能的材料的高需求，设计复合材料是材料开发中不可或缺的一部分。然而，由于复合材料的异质性和复杂性，优化结构的性能和设计的测量通常在实验或计算上难以处理。在这里，我们基于深度学习方法报告给定输入复合几何形状的完整应变和应力张量预测。该模型不仅准确地预测了所有应变和应力张量分量，而且遵守了严格的连续力学原理。此外，通过调整数据统计，即使在处理具有不同成分比率和分层几何结构的材料微结构方面的数据量很少时，模型性能也得到了提高。