基于深度学习的短纤维增强聚氨酯复合材料性能预测基于深度学习的短纤维增强聚氨酯复合材料性能预测

结合深度学习在图像识别领域的优势，将卷积神经网络（CNN）应用于有限元代理模型，预测了平面随机分布短纤维增强聚氨酯复合材料的有效弹性参数，并针对训练过程出现的过拟合，提出了一种数据增强的方法。为验证该代理模型的有效性，比较了其与传统代理模型在预测有效杨氏模量和剪切模量上的精度差异。在此基础上结合蒙特卡洛法利用卷积神经网络代理模型研究了材料微几何参数不确定性的误差正向传递。结果表明:相对于传统代理模型，卷积神经网络模型能更好地学习图像样本的内部特征，得到更加精确的预测结果，并在训练样本空间外的一定范围内可以保持较好的鲁棒性；随着纤维长宽比的增大，微几何参数的不确定性对材料有效性能预测结果会传递较大的误差。 

Prediction of properties of short fiber reinforced urethane polymer composites based on deep learning

Taking advantages of deep learning in the field of image recognition, the convolutional neural network(CNN) was applied to construct a surrogate model to predict the macroscopic performance of the planar random short fiber reinforced urethane composites, and a data enhancement method was proposed to suppress overfitting occurred in the training process. The accuracy in tensile and shear properties of materials predicted by traditional and CNN surrogate models were compared. Results show that compared with the traditional method, CNN model is much better in learning the internal features of the image samples and obtains more accurate prediction results. Meanwhile, robustness is well maintained in a certain range outside the training sample space. Based on this, the proposed CNN model was combined with Monte Carlo method to study the forward propagation of error in the uncertainty of microgeometric parameters. The simulation result demonstrates that as the fiber aspect ratio increases, the uncertainties of the microgeometric parameters will lead to a nonnegligible error in the prediction of the effective properties of the material.