基于对抗神经网络的镁合金力学性能逆向设计

Reverse Design of Mechanical Properties of Magnesium Alloy Based on Generative Adversative Nets

为了特定的应用场景或特性要求寻找合适的材料,本研究提出一种基于循环生成对抗网络的逆向设计框架,集成了长短期记忆人工神经网络和变分自编码器,应用在镁合金从力学性能到成分和挤压参数的逆向设计。框架模型相较于传统的人工神经网络和支持向量机算法预测精度分别提高了27%和47%,在测试集中的均方误差和平均绝对误差分别为0.09和0.15,同时提供了参考范围,以缩小镁合金逆向设计空间。

In order to find suitable materials for specific application scenarios or feature requirements,a reverse design framework based on the Cycle Generative Adversarial Network(Cycle-GAN)was proposed,which integrates Long Short-Term Memory(LSTM)artificial neural networks and Variational Autoencoders(VAE)and applied to the reverse design of magnesium alloys,covering aspects from mechanical performance to composition and extrusion parameters.Compared to traditional artificial neural networks and support vector machine algorithms,the framework model demonstrates an improvement of 27%and 47%in prediction accuracy,respectively.The mean squared error and mean absolute error on the test dataset are 0.09 and 0.15.Additionally,the framework provides a reference range,narrowing down the everse design space for of magnesium alloys.