欢迎来到天天文库
浏览记录
ID:52388158
大小:367.81 KB
页数:5页
时间:2020-03-27
《结合神经网络方法和扩大训练基组构建新B3LYP泛函.pdf》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、物理化学学报(WuliHuaxueXuebao)ActaPhys一.Sin.,2010,26(1):188—192January[Article】www.whxb.pku.edu.cn结合神经网络方法和扩大训练基组构建新B3LYP泛函张家虎王秀军(华南理工大学化学与化工学院应用化学系,广州510640)摘要:神经网络方法成功地应用于修正密度泛函理论B3LYP方法中的三个参数(口。、a和ac)以构建新B3LYP交换相关泛函.本文采用包含输入层、隐藏层和输出层的三层式神经网络结构.总电子数、多重度、偶极矩
2、、动能、四极矩和零点能被选为物理描述符.296个能量数据被随机地分成两组,246个能量数据作为训练集以确定神经网络的最优结构和最优突触权重,50个能量数据作为测试集以测试神经网络的预测能力.修正后的三个参数舀、、从输出层处得到,并用于计算体系的热化学性质如原子化能(AE)、电离势0P)、质子亲合能(PA)、总原子能(TAE)和标准生成热(△e).修正后的计算结果优于传统B3LYP/6—311+G(3d~2p)方法的计算结果.经过神经网络修正后,296个物种的总体均方根偏差从41.0kJ·mol‘。减少
3、到14.2kJ·mol~.关键词:B3LYP泛函;神经网络;描述符;训练基组中图分类号:O641NeuralNetworkApproachforaNewB3LYPFunctionalwithanEnlargedTrainingSetZHANGJia~HuWANGXiu—Jun(DepartmentofAppliedChemistry,CollegeofChemistryandChemicalEngineering,SouthChinaUniversityofTechnology,Guangzhou51
4、0640,P.RChina、Abstract:Aneuralnetworkapproachwasusedtocorrectthreeparameters(aoaandac)intheB3LYPmethodforconstructinganewB3LYPexchangecorrelationfunctiona1.Athree—layerarchitecturewhichconsistedofaninputlayer,ahiddenlayer,andanoutputlayer,wasadoptedinth
5、eneuralnetwork.Thetotalnumberofelectrons,spinmultiplicity,dipolemoment,kineticenergy,quadrupolemoment,andzeropointenergywerechosenasthemostimportantphysicaldescriptors.Inthiswork,296energyvalueswererandomlydividedintotwosubsets:246energyvalueswereusedas
6、thetrainingsettodeterminetheoptimizedstructureoftheneuralnetworkandtheoptimizedsynapticweights;50energyvalueswereusedasatestingsettotestthepredictioncapabilityofourneuralnetwork.Threemodifiedparametersdo,dx,andthatwereobtainedfromtheontputlayerwereusedt
7、ocalculatethermochemicaldatasuchastheatomicenergy(AE),ionizationpotential(IP),protonaffinity(PA),totalatomicenergy(TAE),andstandardheatofformation(A.Thenewresultsobtained,basedontheneuralnetworkapproach,aremuchbetterthantheresultscalculatedusingtheconve
8、ntionalB3LYP/6—311+G(3df,2p)method.Upontheneuralnetworkcorrection,theoverallroot—mean—square(RMS)errorforthe296speciesdecreasedfrom41.0to14.2kJ·mol一.KeyWords:B3LYPfunctional;Neuralnetwork;Descriptor;Trainingset密度泛函理论(DFT)以其良好的计算性
此文档下载收益归作者所有