利用神经网络预测阴离子共偶碱金属卤化物二元系的相图特征

Feature predictions of common anion binary alkali halide phase diagrams by neural networks

利用Kohonen自组织特征映射神经网络来判定二元碱金属卤化物体系是否可以形成共晶体 ,对三个体系(其中一个体系不形成共晶体 ,两个体系形成共晶体 )的预测结果完全正确。在预测共晶温度和共晶体的组成时 ,使用了 6 5 1拓扑结构的误差反向传播神经网络 ,训练算法为自适应逆传播算法 ,对四个二元体系中共晶体的组成预测表明 ,均方根误差为 0 .0 5 35 ,平均相对误差为 1.77% ,对四个体系共晶温度的预测显示 ,均方根误差为 13 .48℃ ,平均相对误差为 1.11%。

In this paper, the Kohonen self organizing feature mapping neural network was developed to determine the formation of eutectics in binary alkali halide systems. Three binary alkali halide systems of which one can not form eutectics and the other two can form eutectics were correctly predicted. In the composition and eutectic temperature predictions, a 6×5×1 topological structure was employed and trained with an adaptive error backpropagation algorithm. Of the 4 predicted binary systems, the composition prediction has a root mean square error of 0.0535 and a mean relative error of 1.77%. For the eutactic temperature predictions, the root mean square error is 13.48℃ and the mean relative error is 1.11%.