• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Machine learning potential aided structure search for low-lying candidates of Au clusters

    2022-08-01 06:02:50TongheYing應(yīng)通和JianbaoZhu朱健保andWenguangZhu朱文光
    Chinese Physics B 2022年7期
    關(guān)鍵詞:朱文

    Tonghe Ying(應(yīng)通和), Jianbao Zhu(朱健保),?, and Wenguang Zhu(朱文光)

    1Department of Physics,University of Science and Technology of China,and Key Laboratory of Strongly-Coupled Quantum Matter Physics,Chinese Academy of Sciences,Hefei 230026,China

    2International Center for Quantum Design of Functional Materials(ICQD),Hefei National Laboratory for Physical Sciences at the Microscale,and Synergetic Innovation Center of Quantum Information and Quantum Physics,University of Science and Technology of China,Hefei 230026,China

    Keywords: machine learning potential,gold cluster,first-principles calculation

    1. Introduction

    For solid material, such as binary compounds and supported nanoparticles,[1,2]determination of its global stable structure is the cornerstone for a detailed study of its mechanical,electronic,magnetic,or chemical properties. In principle,the structure can be determined computationally by searching the potential energy surface (PES) of the system. To obtain the ground-state structure, unbiased search algorithms, such as genetic algorithm(GA),[3]basin-hopping(BH),[4]particleswarm optimization(PSO)[5]have been developed. However,all of these methods require a large number of energy and force calculations in the framework of density functional theory(DFT).Since the computational cost of DFT grows cubically with the size of the system and the configuration space also grows exponentially with the number of atoms,[6]the global structure search for large systems may be unpractical.In this context,alternative methods like empirical potentials[7]are often employed, but the transferability and accuracy are not always satisfactory.

    During lasts decades, ML techniques have been proved useful in various aspects of computational material physics.[8–16]However, most of them are used to assist the simulation of molecular dynamics (MD), where ML potential only needs to cover a part of configuration space.[17]In order to do the global optimization, we have to obtain an ML potential covering almost the whole configuration space,which can be very challenging. Nevertheless, once such an ML potential[18–20]is constructed,e.g.,neural-network-based approaches,[21–23]Gaussian approximation potentials,[24]it can largely accelerate the global optimization benefiting from the speed of ML predictions and still preserve the accuracy of first-principles calculations at the same time. Because of the characteristic of atomic systems, they can be naturally regarded as a graph,with vertices representing individual atoms and edges representing bonds. Therefore, high dimensional neural networks,especially convolutional graph neural networks,[21,25–29]which operate on graph structured data,are suitable to speed up the global optimization. These neural network potentials are based on atomic energy contributions and can be applied to systems of different sizes,which go beyond the limitations of previous neural networks for fixed-size systems.[30]

    Gold is one of the most inert metals in the periodic table, but at the nanometer scale it exhibits outstanding properties for nanoscience such as nanoelectronics, nanobiology,and nanocatalysis.[31–34]Since many intriguing structures and properties are found for nanometer-sized gold clusters, they are attracting interest as the building blocks of novel nanostructured materials and devices.[35,36]Among these applications,cluster structures play an important role in determining their properties, and some of them are highly stable at magic number sizes.[37]In this work,we utilized the combination of PSO[5]and a DFT trained SchNet potential[38]to accelerate the global optimization of Au clusters whose size is below 20 atoms. Trained with our newly developed dataset,the SchNet model shows a great performance on predicting the energy and forces of Au cluster structures in the process of optimizations. With this ML potential accelerated global optimization,we have successfully reproduced the well-known tetrahedral structure of Au20[35]and revealed the 2D–3D structural transition of Au clusters.

    2. Methods

    The workflow starts with PSO, a branch of evolutionary methodology for crystal structure prediction. Within the PSO scheme, each particle is influenced by both the best local and the best global particle,i.e., a particle can learn from its past experiences to adjust its flying speed and direction.Here,the population size is set to be 150 and other parameters are maintained as suggested by Wang.[5]In order to explore the search space of clusters as widely as possible to obtain low energy structures, we do not impose the symmetry constraint on the structure generation step. The most expensive part of the global optimization algorithm is the local relaxation performed at the DFT level. However, in our method,these calculations are replaced by the ML potential, which is obtained by training on our developed dataset. Since the best global particle of each generation largely controls the evolution of PSO, it is reevaluated by calculating the energies of the best local particles obtained from the ML potential at the DFT level. It is worth noting that our method can be classified into an evolution-type algorithm, which puts emphasis on structure search by generational change,while a selectiontype algorithm such as Bayesian optimization[39]aims to select structures preferentially from a large number of candidates by ML. Compared to the selection-type algorithm, generated initial structures by our method show a trend to more stable structures as generations increase. Moreover, in our method,we utilize a well-trained ML potential to replace the timeconsuming local optimization by DFT, promoting the acceleration of structure search.

    2.1. Graph neural network potential

    In order to reduce the total computational cost during the structure exploration,we exploit a graph convolutional neural network called SchNet[38]to describe interatomic potentials in the Au cluster system,which is a combination of a graph neural network and a convolutional neural network. The graph neural network regards atoms and pairwise interactions between them as nodes and edges, respectively. These nodes and edges are represented by feature vectors and they can be iteratively updated through convolution layers. The convolution layer is actually a message-passing function,which gathers information from neighboring atoms within a cutoff radiusRcand updates the feature vector of the central atom. To have an intuitive understanding of SchNet,Behler–Parrinello neural network(BPNN)[40]is often utilized for comparison,which is widely used last decades.In BPNN,the chemical environment around an atom is represented by a set of symmetry functions,which are rotationally,translationally and permutationally invariant, satisfying the requirement of descriptors. In order to distinguish between different chemical elements,each chemical element has its own neural network architecture with fitted parameters. However, in SchNet, descriptors are replaced by iteratively updated feature vectors, whose parameters can be learnt through training. Furthermore,all chemical elements in SchNet share the same architecture, the uniqueness of them guaranteed by respective randomly initialized feature vectors.Figure 1 shows an overview of the SchNet architecture. At each layer, the cluster is represented atom-wise analogous to pixels in an image and interactions between atoms are modeled by interaction blocks where radial basis functions

    play a key role in expanding the interatomic distances,riandrjrepresenting the position coordinate of atomiand atomjrespectively. The centersμkare chosen uniformly every 0.2 ?A between zero and the distance cutoff 5 ?A while the scaling parameterγ=1 ?A.Interactions of atomias the convolution with surrounding atoms can be formulated as

    where°represents the element-wise multiplication andlmeans the layer. The filter-generating networkWl:R3→RFmaps the atom positions to the filter values ofFdimensions.In the pooling layer,the total energy of a system is described as the summation of atomic contributions

    where the atomic energyεiis a function of its chemical environment.

    For the training of the model, the forces are included in the training loss function.The interatomic forces are related to the energy,so we can obtain an energy-conserving force model by differentiating the energy model w.r.t. the atom positions

    Then a combined loss

    can be defined,whereρis the trade-off parameter between the energy and forces loss. To minimize the loss function, after having tested many different hyperparameter settings, we finally use the ADAM optimizer[41]with 32 mini-batches and the learning rate is decayed with a ratio 0.96 from 10-3to 10-5once learning stagnates. The selection of hyperparameters,such as the trade-offρ,feature dimensionsF,interaction numbersT, will be discussed later, and the others follow the setting proposed previously.[38]

    Fig. 1. Illustrations of the SchNet architecture overview. The shifted softplus is defined as ssp(x)=ln(0.5ex+0.5)and the radial basis function is abbreviated to rbf. Wl(ri-rj)is the filter-generating network(orange color)and the box in gray represents the interaction block,which can be repeated for n times.

    2.2. Data generation

    The dataset used to train the ML model is a key factor.To ensure its diversity, data of 14 different sized Au clusters(Au10~Au19, Au21, Au26, Au33, Au35), with each size containing 100 different configurations, are used as initial structures for the simulated annealing (SA). Thus, an initial training dataset of 1.4×106structures is generated by SA from 1000 K to 0 K, which is kept 20 ps and the time step equals 20 fs. However,there are some unreasonable or similar structures in these generated data, which need further refinement.First, structures with the absolute value of any component of forces being above 10 eV/?A,which are considered as unphysical configurations, are removed from the dataset. Second,if some configurations are considered to be similar, only one of them can be kept. The two steps of refinements make the initially generated dataset decreased from 1.4×106to 38276 configurations.

    In the present work, we ensure that two structures are considered different by employing an interatomic and energy comparison criterion.[3]If the difference of energies of two configurations is more than 0.1 eV,they are considered as different configurations.Otherwise,the structure comparison criterion needs to be employed. A sorted list,Di, of all interatomic distances is calculated for all candidatesi. Two configurationsiandjare regarded equal,provided

    kindexes each entry in the listDi. The first criterion is the relative accumulated difference between two candidates and the second criterion is the maximum difference between two distances of the two candidates. Typical values areδrel=0.03 anddmax=0.7 ?A.

    The first-principles DFT calculations were performed using the ViennaAb InitioSimulation Package(VASP).[42]Valence electrons were described using the projector-augmented wave[43](PAW)method. The plane wave expansions were determined by an energy cutoff of 250 eV. The exchange and correlation functional was treated using the Perdew–Burke–Ernzerhof (PBE)[44]parametrization of generalized gradient approximation (GGA) for structural relaxations and total energy calculations. The isolated cluster was represented by a large cubic supercell of edge length 20 ?A and obviously only the Gamma point was used for itsk-point sampling.Optimized atomic structures were achieved when forces on all the atoms were<0.03 eV/?A and the energy was<1×10-4eV.Because the spin-orbit coupling(SOC)is positively correlated with the atomic number while Au has a large atomic number,to obtain the energies of found structures more accurately,we have also carried out the calculations including SOC, which is implemented in the noncollinear version[45]of the PAW method.

    3. Results and discussion

    In order to obtain accurate energy and force predictions,we first perform a model selection on the given reference data.Table 1 lists the results for various settings of numbers of interaction blocksT, numbers of feature dimensionsFof the atomic representations,and the energy-force trade-offρof the combined loss function. It is obvious that the model’s ability to predicting the energy and forces becomes better asTandFincrease.Finally,the model architecture ofT=6 andF=128 works best for the energy as well as forces. For the energy predictions, we obtain the best result when usingρ=0.1 as this puts more emphases on the energy loss while we achieve the best force predictions whenρ=0.01. Here,we select theρ=0.1 model as this achieves the most balanced performance.

    Table 1. Mean absolute errors for energy and forces predictions. We compare SchNet models with varying numbers of interaction blocks T,feature dimensions F,and energy-force tradeoff ρ. The selected model is in bold.

    We train a series of neural network (NN) potentials by gradually enlarging the size of the training dataset from 3827 to 30620.Figure 2(a)shows the evolution of the mean absolute error(MAE)of the energy and forces for these NN potentials,respectively. It is clearly seen that the value of the MAE of both the energy and force significantly decreases as the size of training data increases, indicating a gradual improvement of the potentials. The NN potential with 30620 training data gives the best performance, and will be used in the following work. The training process is done iteratively and after every iteration, testing of the trained weights is done by calculating the MAE of the energy and forces. After the 787th iteration the learning rate is below 10-5,and the training process is stopped. The difference between the MAE of training and testing is 0.043 eV for energy and 0.012 eV/ ?A for forces,as shown in Fig.2(b). The correlation plot between DFT energy (forces) and NN energy (forces) of the testing dataset is shown in Fig.2(c). It is evident from the plot that the accuracy of our NN potential is comparable to that of DFT.

    Since Au20is a tetrahedral structure that was confirmed by experiment and DFT calculation,[35]we take a look at the PSO evolution of its energy distributions, which is shown in Fig. 3. At the beginning of the search, found structures are in relative high-energy region, while their energies decrease quickly as the evolution continues. The structure that is found in the 3rd generation is an intermediate product, a local minimal one shown in the inset. The minima has been kept for several generations,until the global ground structure is found in the 21th generation. We can see that Au20is the famous tetrahedral pyramid,which is proved to be highly stable and is a unique molecule with atomic packing similar to that of bulk gold. It is important to notice that we do not include Au20into our training set,making the found Au20ground structure more convincing.

    Fig. 2. (a) The evolution of the MAE of the predicted energy (left panel)and forces(right panel)as a function of the training set size. (b)The MAE of the energy and forces of training and testing dataset decays as a function of iterations. (c)The comparison of the energy and forces calculated by the ML model and DFT.The black dashed line is the result of the least square fitting.

    Fig. 3. Evolution of energy distributions during PSO structure search for Au20 clusters,where local optimization is conducted by NN potential.Structures in the inset are a local minimal configuration and the ground one.

    Fig. 4. Structures of ground state candidates, found using the SchNet model. The final binding energies are all reoptimized by DFT. There are three items below every structure picture.The first one indicates the number of atoms,while the second and the third items mean the binding energy calculated by PBE and PBE+SOC respectively. The unit of values is eV.

    In addition, we perform structure searches for Au clusters between 4 and 20 atoms, with final configurations reoptimized by DFT, as shown in Fig. 4. To get more accurate energy results,we calculate the binding energy of Au clusters both with PBE and PBE+SOC respectively.The binding energies obtained including the SOC are higher than those without including the SOC for both 2D clusters and 3D clusters. It is obviously seen that small sized Au clusters tend to form planar structures while the larger ones are more likely to have stereo structures,with the critical size for the 2D–3D structural transition being Au14,which is generally consistent with previous research.[46]Furthermore, it can be noticed that many structures evolve from the prior structure with one atom added in a certain position. All 2D Au clusters whose size is below 14 atoms can be described by this rule and the tetrahedral pyramid of Au20can also be regarded as the evolution from Au19with one atom added in the top.More interestingly,some structures found by PSO even have lower binding energies than those reported by former literatures.[46,47]For example, the binding energies of Au13, Au15, Au17, Au18are lower by 0.214 eV,0.191 eV,0.326 eV,0.370 eV than that in the previous literatures,after we recalculate their binding energies with the same parameter of PBE+SOC in this paper. We also plot the energy per atom of our predicted clusters as a function of the cluster size,as displayed in Fig.5,showing an overall downward trend. This is due to the increase of the average number of nearest-neighbors with increasing size. Moreover, the binding energy per atom of even-sized clusters shows a little lower than the ones of their odd-sized neighbors, which can be explained by the effect of electron pairing. It is also remarkable that Au7is one local maximum point in Fig. 5, whose lower symmetry compared to neighbors Au6and Au8results in an increase in energy.

    Fig.5.Energy per atom of ground structures,obtained by the SchNet model.The energy values are given by DFT.

    4. Conclusion

    In this work,we developed an ML atomic potential serving the task of structure search for Au clusters,which has two major advantages. Firstly, SchNet considers the structure as a graph, with atoms representing nodes and interactions between atoms representing edges. This kind of representation bypasses tedious descriptors and makes it possible to be applied in the field where lacking of expert knowledge for the design of descriptors. Secondly, our ML potential replaces the time-consuming local optimization performed by VASP,which allows more structure optimization trials for large systems. By an extensive search in the size range of 4–20, we exhibited their low-lying structures and revealed that the 2D–3D structural transition takes place at the cluster sizeN=14.Such a successful search for low-lying structures of Au clusters not only proves that the current SchNet model is robustly and efficiently valid in seeking low-lying candidates of Au clusters but also indicates that ML is indeed powerful in interatomic potential modeling.

    The code and supporting data for this article are available from Ref.[48].

    Acknowledgements

    Computational support was provided by Supercomputing Center in USTC and National Supercomputing Center in Tianjin.

    Project supported by the National Key Research and Development Program of China (Grant Nos. 2017YFA0204904 and 2019YFA0210004).

    猜你喜歡
    朱文
    Prediction of quantum anomalous Hall effect in CrI3/ScCl2 bilayer heterostructure
    Modeling the heterogeneous traffic flow considering the effect of self-stabilizing and autonomous vehicles
    Metal substrates-induced phase transformation of monolayer transition metal dichalcogenides for hydrogen evolution catalysis*
    走三邊
    秦川好
    唱起號(hào)子走漢江
    Interaction Solutions for Kadomtsev-Petviashvili Equation with Variable Coefficients?
    熱鬧的大山
    Teacher:Teacher—dominant or Student—centered
    西部論叢(2017年3期)2017-09-11 06:21:44
    朱文韜 平凡之中展現(xiàn)別樣風(fēng)采
    北方人(2017年12期)2017-07-25 09:17:06
    亚洲国产精品久久男人天堂| 午夜福利视频1000在线观看| 亚洲国产高清在线一区二区三| 亚洲性久久影院| 人人妻人人看人人澡| 国产大屁股一区二区在线视频| 亚洲av福利一区| 国产亚洲午夜精品一区二区久久 | 国产成人一区二区在线| 最近2019中文字幕mv第一页| 日韩亚洲欧美综合| 精品久久久久久久久av| 2022亚洲国产成人精品| 欧美日韩综合久久久久久| 一夜夜www| 国产精品精品国产色婷婷| 2022亚洲国产成人精品| 精品久久久久久久人妻蜜臀av| 午夜视频国产福利| 看黄色毛片网站| 亚洲欧美成人精品一区二区| 国产午夜精品论理片| 久久99热6这里只有精品| 亚洲图色成人| 国产亚洲一区二区精品| 国产精品久久久久久久久免| 禁无遮挡网站| av在线天堂中文字幕| 欧美zozozo另类| 91aial.com中文字幕在线观看| 91av网一区二区| 热99在线观看视频| www.色视频.com| 亚洲精品乱码久久久久久按摩| 亚洲自拍偷在线| 国产高清不卡午夜福利| 九九在线视频观看精品| 蜜桃亚洲精品一区二区三区| 久久精品国产99精品国产亚洲性色| 性色avwww在线观看| 在线免费十八禁| 亚洲欧美清纯卡通| 久久久久久久久久黄片| 亚洲一级一片aⅴ在线观看| 26uuu在线亚洲综合色| 日韩三级伦理在线观看| 亚洲av男天堂| 丝袜喷水一区| 国产伦理片在线播放av一区| 成人美女网站在线观看视频| av黄色大香蕉| 午夜福利成人在线免费观看| 精品无人区乱码1区二区| 久久久色成人| 91精品国产九色| 亚洲人与动物交配视频| 男人的好看免费观看在线视频| 亚洲精品国产成人久久av| 日韩av在线大香蕉| 性插视频无遮挡在线免费观看| 日韩 亚洲 欧美在线| 少妇裸体淫交视频免费看高清| 国语对白做爰xxxⅹ性视频网站| 乱人视频在线观看| 99国产精品一区二区蜜桃av| 亚洲国产精品成人综合色| 1024手机看黄色片| 嫩草影院精品99| 欧美一区二区国产精品久久精品| 卡戴珊不雅视频在线播放| 亚洲av熟女| 亚洲成人中文字幕在线播放| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 欧美一级a爱片免费观看看| 国产精品野战在线观看| 国产乱人偷精品视频| 国产在线一区二区三区精 | 人体艺术视频欧美日本| 国产黄片美女视频| 亚洲三级黄色毛片| 国产精品.久久久| 联通29元200g的流量卡| 国产探花在线观看一区二区| 日韩一区二区视频免费看| 又爽又黄无遮挡网站| 欧美另类亚洲清纯唯美| 亚洲精品乱码久久久v下载方式| 可以在线观看毛片的网站| 高清在线视频一区二区三区 | 一个人看视频在线观看www免费| 99视频精品全部免费 在线| 两个人的视频大全免费| 成年版毛片免费区| 国产成人精品一,二区| 国产精品人妻久久久久久| 欧美日本视频| 亚洲天堂国产精品一区在线| 亚洲性久久影院| 亚洲成人av在线免费| 欧美区成人在线视频| 久久久久国产网址| 亚洲欧美精品自产自拍| 国产女主播在线喷水免费视频网站 | 少妇裸体淫交视频免费看高清| 色播亚洲综合网| 午夜老司机福利剧场| 97热精品久久久久久| 亚洲最大成人av| 大香蕉97超碰在线| 99热这里只有精品一区| 国产精品伦人一区二区| 男女那种视频在线观看| 国产精品久久电影中文字幕| 99热这里只有是精品50| 在线播放无遮挡| 欧美激情久久久久久爽电影| 小蜜桃在线观看免费完整版高清| 亚洲av二区三区四区| 色网站视频免费| 91久久精品电影网| 成人毛片60女人毛片免费| 日韩欧美 国产精品| 中文亚洲av片在线观看爽| 三级经典国产精品| 搞女人的毛片| 岛国毛片在线播放| 精品熟女少妇av免费看| 国产毛片a区久久久久| 欧美日韩国产亚洲二区| 久久久久国产网址| 日韩,欧美,国产一区二区三区 | 亚洲国产精品成人综合色| 国产黄a三级三级三级人| 日韩成人av中文字幕在线观看| 国产精品av视频在线免费观看| 高清av免费在线| 欧美成人一区二区免费高清观看| 男人的好看免费观看在线视频| 如何舔出高潮| 99热这里只有精品一区| 精品久久久久久久人妻蜜臀av| 欧美日韩国产亚洲二区| 精品一区二区三区人妻视频| 色吧在线观看| 国产精品综合久久久久久久免费| 亚洲成人av在线免费| 日本五十路高清| 久久精品国产自在天天线| 国产精品精品国产色婷婷| 在线免费观看不下载黄p国产| 欧美成人a在线观看| 国模一区二区三区四区视频| 国产精品蜜桃在线观看| 亚洲精品乱码久久久v下载方式| 啦啦啦韩国在线观看视频| 亚洲国产日韩欧美精品在线观看| 能在线免费观看的黄片| 色尼玛亚洲综合影院| 特级一级黄色大片| 亚洲av日韩在线播放| 女人久久www免费人成看片 | 观看美女的网站| 国产精品国产三级国产专区5o | 成人毛片60女人毛片免费| 亚洲精品日韩在线中文字幕| 一边亲一边摸免费视频| 日本欧美国产在线视频| 亚洲av免费高清在线观看| 日日干狠狠操夜夜爽| 麻豆成人午夜福利视频| 成人午夜高清在线视频| 成年免费大片在线观看| 久久草成人影院| 亚洲成av人片在线播放无| 伦理电影大哥的女人| 男人舔女人下体高潮全视频| 亚洲精品,欧美精品| 亚洲av免费在线观看| 日本爱情动作片www.在线观看| av视频在线观看入口| 亚洲欧美一区二区三区国产| 国产v大片淫在线免费观看| 熟妇人妻久久中文字幕3abv| 黄色一级大片看看| 国产真实伦视频高清在线观看| 久久久亚洲精品成人影院| 高清av免费在线| 成人午夜高清在线视频| 建设人人有责人人尽责人人享有的 | 1000部很黄的大片| 欧美bdsm另类| 亚洲av.av天堂| 成年av动漫网址| 老师上课跳d突然被开到最大视频| 在线播放国产精品三级| 成人高潮视频无遮挡免费网站| 九草在线视频观看| 黄片无遮挡物在线观看| eeuss影院久久| 视频中文字幕在线观看| 日本爱情动作片www.在线观看| 国产v大片淫在线免费观看| 亚洲精华国产精华液的使用体验| 免费无遮挡裸体视频| 身体一侧抽搐| 男女下面进入的视频免费午夜| 91久久精品国产一区二区成人| 国产淫片久久久久久久久| 能在线免费看毛片的网站| 午夜福利在线观看吧| 91精品国产九色| 啦啦啦观看免费观看视频高清| 人妻制服诱惑在线中文字幕| 纵有疾风起免费观看全集完整版 | 又爽又黄无遮挡网站| 久久久久网色| 日韩一区二区视频免费看| 国产白丝娇喘喷水9色精品| 久久精品人妻少妇| 久久久午夜欧美精品| 一级二级三级毛片免费看| 毛片女人毛片| 免费看日本二区| 久久精品国产亚洲av涩爱| 久久久亚洲精品成人影院| 长腿黑丝高跟| 精品免费久久久久久久清纯| 中文字幕制服av| 国产伦在线观看视频一区| 日本与韩国留学比较| 国产精品综合久久久久久久免费| 亚洲av免费在线观看| 丰满少妇做爰视频| 久久草成人影院| 在线播放国产精品三级| 国产一级毛片七仙女欲春2| 亚洲av.av天堂| 99热网站在线观看| 六月丁香七月| 91精品伊人久久大香线蕉| 亚洲精品乱码久久久v下载方式| a级一级毛片免费在线观看| 大话2 男鬼变身卡| 国产免费一级a男人的天堂| 成年女人永久免费观看视频| 久久韩国三级中文字幕| 亚洲欧美精品综合久久99| 亚洲,欧美,日韩| 国产亚洲5aaaaa淫片| 91久久精品国产一区二区三区| 99九九线精品视频在线观看视频| 99久国产av精品| 男女那种视频在线观看| 我的老师免费观看完整版| 国产精品一区二区性色av| 亚洲av免费在线观看| 淫秽高清视频在线观看| 亚洲精品456在线播放app| 尤物成人国产欧美一区二区三区| 人妻制服诱惑在线中文字幕| 久久精品国产亚洲av涩爱| 搡女人真爽免费视频火全软件| 日本免费一区二区三区高清不卡| 九九热线精品视视频播放| 日韩高清综合在线| 极品教师在线视频| 日韩国内少妇激情av| 亚洲中文字幕日韩| 免费电影在线观看免费观看| av卡一久久| 欧美高清性xxxxhd video| 欧美激情国产日韩精品一区| 午夜福利高清视频| 国产av不卡久久| 欧美性猛交╳xxx乱大交人| 99热6这里只有精品| 99热6这里只有精品| 日本熟妇午夜| 波多野结衣高清无吗| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 国产综合懂色| 欧美97在线视频| 午夜精品一区二区三区免费看| 久久久亚洲精品成人影院| 免费av毛片视频| 久久久国产成人免费| 午夜精品国产一区二区电影 | 亚洲欧美精品专区久久| 亚洲av男天堂| 乱系列少妇在线播放| 欧美人与善性xxx| 九草在线视频观看| 日产精品乱码卡一卡2卡三| 少妇裸体淫交视频免费看高清| 日本黄色片子视频| 国产午夜精品一二区理论片| 中文字幕av成人在线电影| 丰满乱子伦码专区| 亚洲欧美清纯卡通| 99久久精品一区二区三区| 蜜臀久久99精品久久宅男| 亚洲av一区综合| 一本一本综合久久| 久久亚洲国产成人精品v| 国产成人91sexporn| 日本av手机在线免费观看| 又爽又黄a免费视频| 国产伦精品一区二区三区四那| 美女黄网站色视频| 免费大片18禁| 一卡2卡三卡四卡精品乱码亚洲| 精品熟女少妇av免费看| 亚洲av成人精品一二三区| 99在线视频只有这里精品首页| 国产亚洲av片在线观看秒播厂 | 老司机影院成人| 尾随美女入室| 男人舔女人下体高潮全视频| 又粗又硬又长又爽又黄的视频| 国产亚洲一区二区精品| 日本一本二区三区精品| 国产精品久久久久久久电影| 伦精品一区二区三区| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 性插视频无遮挡在线免费观看| 国产视频内射| 国产黄色视频一区二区在线观看 | 亚洲av福利一区| 又爽又黄无遮挡网站| 国产av码专区亚洲av| 亚洲第一区二区三区不卡| 我要搜黄色片| 毛片女人毛片| av在线亚洲专区| 九九久久精品国产亚洲av麻豆| 亚洲激情五月婷婷啪啪| 六月丁香七月| 亚洲图色成人| 日本一本二区三区精品| 亚洲成人精品中文字幕电影| 国国产精品蜜臀av免费| 日韩亚洲欧美综合| 日韩国内少妇激情av| www.色视频.com| 免费在线观看成人毛片| 国产精品一区二区三区四区免费观看| 日韩人妻高清精品专区| 亚洲精品影视一区二区三区av| 国产精品国产三级国产专区5o | 91在线精品国自产拍蜜月| 亚洲国产精品国产精品| 91午夜精品亚洲一区二区三区| 一级毛片久久久久久久久女| 少妇的逼好多水| 欧美成人一区二区免费高清观看| 亚洲国产成人一精品久久久| 国产一级毛片七仙女欲春2| 97人妻精品一区二区三区麻豆| 97热精品久久久久久| 日韩视频在线欧美| 中文天堂在线官网| 男女视频在线观看网站免费| 99久久成人亚洲精品观看| 免费一级毛片在线播放高清视频| 国内揄拍国产精品人妻在线| 日本一二三区视频观看| 国产在线男女| 女人久久www免费人成看片 | 午夜精品国产一区二区电影 | 欧美zozozo另类| 国产精品久久久久久精品电影小说 | 一个人免费在线观看电影| 极品教师在线视频| 国产精品一区二区三区四区免费观看| 国产精品三级大全| 99久国产av精品| 亚洲人成网站在线播| 久久欧美精品欧美久久欧美| 嫩草影院入口| 久久亚洲国产成人精品v| 七月丁香在线播放| 性色avwww在线观看| 一个人看视频在线观看www免费| 26uuu在线亚洲综合色| av视频在线观看入口| 看黄色毛片网站| 国产精品一区二区性色av| 99久久无色码亚洲精品果冻| 91精品一卡2卡3卡4卡| 亚洲乱码一区二区免费版| 亚洲国产精品成人久久小说| 午夜精品一区二区三区免费看| 亚洲怡红院男人天堂| 亚洲国产欧美在线一区| 日韩一区二区三区影片| av福利片在线观看| 深爱激情五月婷婷| 国产美女午夜福利| 国产成人freesex在线| 亚洲av免费在线观看| 69人妻影院| 可以在线观看毛片的网站| 欧美区成人在线视频| 欧美性猛交╳xxx乱大交人| 亚洲av成人精品一区久久| 日本五十路高清| 免费观看的影片在线观看| 国产又黄又爽又无遮挡在线| 婷婷色麻豆天堂久久 | 久久久久久久久久成人| 久久久亚洲精品成人影院| 一二三四中文在线观看免费高清| 亚洲人成网站在线观看播放| 欧美日本亚洲视频在线播放| 99在线视频只有这里精品首页| 中文字幕免费在线视频6| 久久99蜜桃精品久久| 午夜福利视频1000在线观看| 日韩欧美国产在线观看| 久久精品国产自在天天线| 精品一区二区三区视频在线| 国产69精品久久久久777片| 午夜福利高清视频| 精品久久国产蜜桃| 国产精品久久电影中文字幕| 欧美日本亚洲视频在线播放| 深爱激情五月婷婷| 国产精品野战在线观看| 日韩精品有码人妻一区| 99热这里只有是精品在线观看| 国产毛片a区久久久久| 黑人高潮一二区| 欧美3d第一页| 波野结衣二区三区在线| 日日干狠狠操夜夜爽| 亚洲欧美精品综合久久99| 美女xxoo啪啪120秒动态图| 人妻系列 视频| 一级av片app| 人体艺术视频欧美日本| 亚洲国产欧美在线一区| 欧美成人午夜免费资源| 欧美+日韩+精品| 变态另类丝袜制服| 嘟嘟电影网在线观看| 国产极品精品免费视频能看的| 欧美成人一区二区免费高清观看| 少妇的逼水好多| av免费在线看不卡| 能在线免费观看的黄片| 色网站视频免费| 国产高清国产精品国产三级 | 天美传媒精品一区二区| 国产精品电影一区二区三区| 国产一区二区在线观看日韩| 色吧在线观看| 国产av一区在线观看免费| 中文亚洲av片在线观看爽| 好男人视频免费观看在线| 国产精品野战在线观看| 午夜福利在线在线| 最近2019中文字幕mv第一页| 人体艺术视频欧美日本| 午夜福利在线观看吧| 人人妻人人澡人人爽人人夜夜 | 变态另类丝袜制服| 啦啦啦观看免费观看视频高清| 欧美区成人在线视频| videossex国产| 成年版毛片免费区| 哪个播放器可以免费观看大片| 国产69精品久久久久777片| 日本熟妇午夜| 精品免费久久久久久久清纯| 少妇熟女欧美另类| 国产极品天堂在线| 午夜精品在线福利| 亚洲性久久影院| 天堂√8在线中文| 中文字幕亚洲精品专区| 少妇丰满av| 麻豆成人午夜福利视频| 国产精品一区二区三区四区免费观看| 国产黄色小视频在线观看| 国产色爽女视频免费观看| 久久婷婷人人爽人人干人人爱| 亚洲精品456在线播放app| 夜夜看夜夜爽夜夜摸| 黄片wwwwww| 国产精品国产三级国产专区5o | 国产精品熟女久久久久浪| 波多野结衣巨乳人妻| 久久精品综合一区二区三区| 国产午夜精品一二区理论片| 中文字幕av成人在线电影| 最近2019中文字幕mv第一页| 亚洲中文字幕日韩| 97在线视频观看| 国产色婷婷99| 亚洲熟妇中文字幕五十中出| 国产91av在线免费观看| 女人久久www免费人成看片 | 九草在线视频观看| 国产成人91sexporn| 国产精品伦人一区二区| 久久国产乱子免费精品| 亚洲高清免费不卡视频| 人体艺术视频欧美日本| 国产亚洲精品久久久com| 久久久久久久久中文| 久久久久久大精品| 亚洲在线自拍视频| 亚洲中文字幕日韩| 亚洲av免费高清在线观看| 亚洲精品aⅴ在线观看| 国产免费视频播放在线视频 | 亚洲国产精品合色在线| 性插视频无遮挡在线免费观看| 国产精品一区二区三区四区免费观看| av在线播放精品| 亚洲,欧美,日韩| 免费av毛片视频| 日韩欧美国产在线观看| 日韩三级伦理在线观看| 精品久久久久久久久久久久久| 99热网站在线观看| 99久国产av精品| 亚洲欧洲国产日韩| 亚洲国产精品国产精品| 国产成年人精品一区二区| 干丝袜人妻中文字幕| 男的添女的下面高潮视频| 精品无人区乱码1区二区| 汤姆久久久久久久影院中文字幕 | 欧美激情在线99| av专区在线播放| 我要看日韩黄色一级片| 尾随美女入室| 免费观看性生交大片5| 亚洲综合精品二区| 1000部很黄的大片| 国产精品美女特级片免费视频播放器| 午夜福利高清视频| 91久久精品国产一区二区成人| .国产精品久久| av卡一久久| www.av在线官网国产| videossex国产| 国产精品一区二区性色av| 亚洲综合精品二区| 欧美三级亚洲精品| 亚洲国产精品成人久久小说| 汤姆久久久久久久影院中文字幕 | 久久国内精品自在自线图片| 人体艺术视频欧美日本| 国产亚洲91精品色在线| 欧美变态另类bdsm刘玥| 我要搜黄色片| 成人综合一区亚洲| 亚洲精品国产成人久久av| 欧美潮喷喷水| 偷拍熟女少妇极品色| 国产91av在线免费观看| 丰满人妻一区二区三区视频av| 可以在线观看毛片的网站| 中文精品一卡2卡3卡4更新| 老女人水多毛片| 我的女老师完整版在线观看| 欧美潮喷喷水| 欧美日本亚洲视频在线播放| 99热这里只有精品一区| 亚洲国产欧美在线一区| 成人美女网站在线观看视频| 综合色av麻豆| 国产在线一区二区三区精 | 国产精品一二三区在线看| 精华霜和精华液先用哪个| 日韩一区二区三区影片| 欧美bdsm另类| 日本一本二区三区精品| 欧美性猛交黑人性爽| 干丝袜人妻中文字幕| 国产精品一区二区性色av| 韩国av在线不卡| 男女下面进入的视频免费午夜| 晚上一个人看的免费电影| 国产av码专区亚洲av| 国产精品美女特级片免费视频播放器| 日韩av不卡免费在线播放| 在线免费观看不下载黄p国产| 婷婷色综合大香蕉| 最近最新中文字幕大全电影3| 99久久中文字幕三级久久日本| 亚洲av成人精品一区久久| 身体一侧抽搐| 国产69精品久久久久777片| 自拍偷自拍亚洲精品老妇| 高清毛片免费看| 亚洲精品日韩av片在线观看| 极品教师在线视频| 久久久久久久久久黄片| 亚洲人成网站在线观看播放| 免费观看人在逋| 国产精品人妻久久久久久| 久久精品综合一区二区三区| 黄色一级大片看看| 老司机影院成人| 欧美三级亚洲精品| 三级国产精品片| 亚洲伊人久久精品综合 | 三级男女做爰猛烈吃奶摸视频| 午夜精品在线福利| 亚洲欧美日韩无卡精品| 欧美极品一区二区三区四区| 日韩欧美三级三区|