贝叶斯正则化神经网络预测金属晶体结合能的研究

采用贝叶斯正则化神经网络(BRNN)对61种金属晶体结合能进行了预测。对网络结构、训练集、预测集以及学习次数进行了优化，并用独立预测样本对贝叶斯正则化神经网络作了检验。预测结果表明，在推广能力方面，贝叶斯正则化神经网络优于熟知的反向传播(BP)神经网络和多元线性回归方法(MLR)。它可望成为元素和化合物构效关系研究的辅助手段。

The application of bayesian-regularization neural network in prediction of the cohesive energy of metallic crystalloid

The cohesive energy,of 61 metallic crystalloid is predicted by using Bayesian-Regularization neural networks(BRNN).Theeffect of structure of network,the size of learning set and predicting set,the learning epochs on predicted results was investigated.Bayesian-Regularization neural networks was verified with independent prediction samples.The suitable conditions are:input nodes:3;hidden nodes:6;output node:1;the learning epochs:300;the size of learning set:51.Predicted results indicated that the Bayes-ian-Regularization neural networks in extend ability were better than back-propagation(BP)neural network and multiple linear regres-sion(MLR)in quality.Therefore,we can expect that Bayesian-Regularization neural network might be used as an effective assistanttechnique for the investigation of quantitative structure-property relationship(QSPR)of the elements and compounds.