(福州大學(xué) 材料科學(xué)與工程學(xué)院，福州350002)
摘 要: 利用嵌入原子模型， 采用分子動(dòng)力學(xué)方法計(jì)算了貴金屬Au低指數(shù)晶面及部分簡(jiǎn)單高指數(shù)晶面的表面能。 同時(shí)， 采用Levenberg-Marquardt 算法， 建立了Au表面能的BP神經(jīng)網(wǎng)絡(luò)模型； 結(jié)合分子動(dòng)力學(xué)模型的計(jì)算數(shù)據(jù)， 通過大量數(shù)據(jù)的自學(xué)習(xí)訓(xùn)練， 完成神經(jīng)網(wǎng)絡(luò)模型對(duì)Au高指數(shù)晶面表面能的預(yù)測(cè)。 計(jì)算結(jié)果表明： 該方法具有較高的預(yù)測(cè)精度， 能正確預(yù)言低指數(shù)晶面表面能的排序， Au各晶面的表面能隨其晶面與(111)密排面夾角的增大呈現(xiàn)先增大后減小的特點(diǎn)。
關(guān)鍵字: 表面能； 嵌入原子勢(shì)； 人工神經(jīng)網(wǎng)絡(luò)；Levenberg-Marquardt算法
molecular dynamics combined with neural networks
（ School of Materials Science and Engineering, Fuzhou University, Fuzhou 350002, China）
Abstract: Via embedded-atom model and molecular dynamics simulation, the surface energies of three low-index and some high-index planes were calculated for precious metal Au, and the error back-propagation network (BP) developed by Levenberg-Marquardt algorithm was adopted. Combining the data calculated with the molecular dynamics model, a great deal of data were trained many times and compared with the calculated data, and the prediction of high-index surface energy was performed. The results show that the method has high predicting accuracy. The order of the three low-index planes was predicted exactly. The surface energies on the other planes show a tendency that first increasing and then decreasing with angle between the planes and (111) plane increasing.
Key words: surface energy; embedded-atom potential; artificial neural network; Levenberg-Marquardt algorithm
