基于深度学习的含能材料生成焓预测方法  

Enthalpy of Formation Prediction for Energetic Materials Based on Deep Learning

作　　者：徐雅斌[1,2,3] 孙胜杰 武装 XU Ya-bin;SUN Sheng-jie;WU Zhuang(Beijing Information Science and Technology University,School of Computer,Beijing 100101,China;Beijing Key Laboratory of Internet Culture and Digital Dissemination Research,Beijing 100101,China;Beijing Advanced Innovation Center for Materials Genome Engineering,Beijing Information Science and Technology University,Beijing 100101,China)

机构地区：[1]北京信息科技大学计算机学院,北京100101 [2]网络文化与数字传播北京市重点实验室,北京100101 [3]北京材料基因工程高精尖创新中心北京信息科技大学,北京100101

出　　处：《含能材料》2021年第1期20-28,I0002,共10页Chinese Journal of Energetic Materials

基　　金：北京材料基因工程高精尖创新中心北京信息科技大学资助;国家自然科学基金资助(61672101);网络文化与数字传播北京市重点实验室基金资助(ICDDXN004)。

摘　　要：为了加快新型含能材料研发的进度,减少因大量实验而带来的时间和资源的消耗问题,基于材料基因工程理论提出一种含能材料生成焓的预测方法。首先将搜集到的代表含能材料分子结构的原子坐标数据转换成表示分子内笛卡尔坐标系的库仑矩阵,以消除含能材料分子结构因平移、旋转、交换索引顺序等操作对生成焓预测造成的影响;然后,根据提出的基于Attention机制的卷积神经网络(Convolutional Neural Network,CNN)和双向长短期记忆网络(Bi-directional Long Short-term Memory Network,Bi-LSTM)的融合模型对含能材料的生成焓进行预测。这样,既可以有效提取数据的特征,又能充分考虑数据间的相关性,同时还能够突出重要特征对预测结果的影响。对比实验结果表明,提出的基于深度学习的方法在生成焓的预测上拥有最低的实验误差,其平均绝对误差(Mean Absolute Error,MAE)、平均绝对百分误差(Mean Absolute Percentage Error,MAPE)、均方根误差(Root Mean Square Error,RMSE)和均方根对数误差(Root Mean Squared Logarithmic Error,RMSLE)分别为0.0374、1.32%、0.0541和0.028,实现了“结构—性能”的预测目标,为含能材料生成焓的预测提供了一种新方法。In order to speed up the development of new energetic materials and reduce the time and resource consumption caused by a large number of experiments,a method for predicting enthalpy of formation of energetic materials is proposed based on the theory of material genetic engineering.Firstly,the collected atomic coordinate data representing the molecular structure of energetic materials were converted into a coulomb matrix representing the cartesian coordinate system in the molecule to elim inate the influence of translation,rotation,index order and other operations on the prediction of enthalpy of formation.Then,the enthalpy of formation of energetic materials was predicted according to the proposed fusion model of Convolutional Neural Network(CNN)and Bi-directional Long Short-term Memory Network(Bi-LSTM)based on Attention mechanism.In this way,not only can the characteristics of the data be extracted effectively,but also the correlation between the data and the lack of long-term dependence can be fully considered.Meanwhile,the influence of important characteristics on the prediction results can be highlighted.The comparison of experimental results shows that the proposed method based on deep learning has the low est experimental error in the prediction of enthalpy of formation.Its Mean Absolute Error(MAE),Mean Absolute Percentage Error(MAPE),Root Mean Square Error(RMSE)and Root Mean Squared Logarithmic Error(RMSLE)are 0.0374,1.32%,0.0541 and 0.028,respectively.The prediction goal of"structure-performance"is realized,and a new method is provided for the prediction of enthalpy of formation of energetic materials.

关 键 词：含能材料 生成焓 Attention机制 卷积神经网络 双向长短期记忆网络 

分 类 号：TJ55[兵器科学与技术—军事化学与烟火技术] TP399[自动化与计算机技术—计算机应用技术]