1.College of Mathematics, Physics and Electronic Information Engineering, Wenzhou University, Wenzhou 325035, China 2.Department of Information, Wenzhou Vocational and Technical College, Wenzhou 325035, China 3.Department of Physics, Zhejiang University, Hangzhou 310027, China
Fund Project:Project supported by the National Natural Science Foundation of China (Grant Nos. 11775161, 61874078), the Natural Science Foundation of Zhejiang Province, China (Grant No. LY17A040007), and the Research Foundation of Education Bureau of Zhejiang Province, China (Grant No. Y201738867).
Received Date:29 April 2018
Accepted Date:25 July 2018
Available Online:01 October 2019
Published Online:20 October 2019
Abstract:Traditional Monte Carlo simulation requires a large number of samples to be employed for calculating various physical parameters, which needs much time and computer resources due to inefficient statistical cases rather than mining data features for each example. Here, we introduce a technique for digging information characteristics to study the phase transition of polymer generated by Monte Carlo method. Convolutional neural network (CNN) and fully connected neural network (FCN) are performed to study the critical adsorption phase transition of polymer adsorbed on the homogeneous cover and stripe surface. The data set (conformations of the polymer) is generated by the Monte Carlo method, the annealing algorithm (including 48 temperatures ranging from T = 8.0 to T = 0.05) and the Metropolis sampling method, which is marked by the state labeling method and the temperature labeling method and used for training and testing of the CNN and the FCN. The CNN and the FCN network can not only recognize the desorption state and adsorption state of the polymer on the homogeneous surface (the critical phase transition temperature TC = 1.5, which is close to the critical phase transition temperature TC = 1.625 of the infinite chain length of polymer adsorbed on the homogeneous surface regardless of the size effect), but also recognize the desorption state, the single-stripe adsorption state and the multi-stripe adsorption state of polymer on the stripe surface(the critical phase transition temperature T1 = 0.55 and T2 = 1.1, which are consistent respectively with T1 = 0.58 and T2 = 1.05 of polymer adsorbed on the stripe-patterned surface derived from existing research results). We obtain almost the same critical adsorption temperature by two different labeling methods. Through the study of the relationship between the size of the training set and the recognition rate of the neural network, it is found that the deep neural network can well recognize the conformational state of polymer on homogeneous surface and stripe surface of a small set of training samples (when the number of samples at each temperature is greater than 24, the recognition rate of the polymer is larger than 95.5%). Therefore, the deep neural network provides a new calculation method for polymer simulation research with the Monte Carlo method. Keywords:deep neural network/ adsorption phase transition/ polymer/ Monte Carlo method
这里的$\Delta E$表示每一次运动前后的能量变化, ${k_{\rm{B}}}$为玻尔兹曼常数, T为温度. 本文采用卷积神经网络和全连接神经网络研究高分子链的吸附相变, 在每个温度下抽取了9600个样本用于神经网络的训练和测试. 神经网络的梯度更新采用累积更新算法, 并使用滑动平均以及正则化来防止过拟合. 本文所采用的神经网络模型如图1所示. 图 1 神经网络结构示意图 (a) 卷积神经网络, INPUT表示输入层, Convolution表示卷积层, MAXPOOL表示池化层, Full connection表示全连接层, OUTPUT表示输出层, PADDING方式均为SAME; (b) 全连接网络的一般结构, 其中hidden layer表示隐藏层, 使用正则化和dropout来防止过拟合, DIM表示输入张量的维度 Figure1. Schematic diagram of the neural network structure: (a) Convolutional neural network, INPUT is the data entry, OUTPUT is the learning result, and the padding way is SAME; (b) the general structure of a full-connected network, where regularization and dropout are used to prevent overfitting, and DIM represents the dimension of the tensor.
3.结果与讨论首先采用机器学习研究了均质表面上高分子链吸附相变的问题. 由前人的研究可知高分子链在均质表面存在着高温时的脱附状态(desorption state, DE态)、低温时吸附状态(adsorption state, AD态)以及在这两个状态之间存在的一个临界相变点. 高分子链在表面上的吸附率以及构象如图2所示. 图 2 吸附率与温度之间的关系. 其中链长N = 160, 插图(a) 是温度T = 1.0时的吸附态构象, 插图(b) 是温度T = 2.0时的脱附态构象 Figure2. Relationship between adsorption rate and temperature. Wherein the chain length N = 160, inset (a) is the conformation of polymer adsorbed on the surface at temperature T = 1.0, and inset (b) is the conformation of polymer desorbed from surface at temperature T = 2.0.
从图2可以看出当吸附率等于0 (即没有单体接触表面)时即为脱附状态, 其典型的构象如图2(b)所示. 当吸附率非零时即发生了吸附现象, 称为吸附状态, 典型构象如图2(a)所示. 临界相变点大约在$T = 1.6$附近, 因此对于温度标记法, 我们选取$T \in \left[ {1.1,\;1.4} \right]$的样本作为吸附态标记, $T \in \left[ {1.8,\;3.0} \right]$的样本作为脱附态标记. 然后采用卷积网络和全连接网络来对标记好的样本进行训练以及识别, 其中全连接网络采用不同数量的隐藏层来进行实验, 结果如图3所示. 图 3 识别率与训练样本的Epoch, 神经网络层数以及每个温度采用的训练样本数之间的关系 (a) 识别率与Epoch的关系图, SPT (sample per temperature)表示在每个温度下抽取的用于训练神经网络的样本数目, 采用状态标记法标记样本. 其中nh = 1表示隐藏层数目为1, 其他的类似, nh = 1至 nh = 3均采用SPT = 192的样本用于训练, 剩余的样本用于测试与验证, 插图描述的是识别率与隐藏层数目nh的关系, 该识别率为每个学习器最终稳定的识别结果; (b) 识别率与每个温度采用训练样本数目的关系图, 采用状态标记法标记样本, 隐藏层数等于3, 纵坐标为不同训练样本在足够Epoch下达到稳定时的识别率, 测试集均为SPT = 7680, 且与训练集不重复 Figure3. The relationship between the recognition rate and the Epochs of training case, the number of neural network layers and the number of training samples obtained from each temperature: (a) the plot of recognition rate versus Epochs. SPT (sample per temperature) represents the number of samples extracted at each temperature for training the neural network. The sample is labeled by status. Where nh = 1 indicates that the number of hidden layers is equal to 1, and the others are similar. All of nh = 1 to 3 uses a sample of SPT = 192 for training, and the remaining samples are used for verification. The illustration depicts the relationship between the recognition rate and the number of hidden layers, which is the final stable recognition result for each classifier; (b) the plot of the recognition rate versus the number of training samples selected at each temperature. The sample is marked by status and the number of hidden layers is equal to 3. The y-axis is the stable recognition rate of different number of training samples under a sufficiently large Epoch. The validation set is SPT = 7680 and is not repeated with the training set.
从图3(a)中可以看出, 随着隐藏层数目的增加, 识别率先提高然后趋于稳定, 当隐藏层数目大于等于3时识别率基本稳定在97.1%, 因此本文缺省所采用的隐藏层数均为3. 当Epoch > 30时训练基本趋于稳定. 从图3(b)中可以看出, 即使每个温度下抽取用于训练神经网络的样本数目(sample per temperature, SPT)足够小, 例如SPT=1时神经网络仍然拥有91.88%的识别率, 也足以识别大部分的样本. 当${\rm{SPT}} \geqslant 24$时识别率达到95.5%以上, 这说明本文采用较小的样本数就可以达到较高的高分子构象的识别率. 然后我们采用卷积神经网络进行研究, 识别结果如图4所示. 图 4 神经网络训练的识别结果图. 横坐标为温度, State表示每个温度下的样本被识别为某个状态的概率, S表示状态标记法, T表示温度标记法, AD表示吸附态, DE表示脱附态. 图为两种标记方法的识别结果, 卷积网络的识别率为98.3%, AUC值为0.9989, 全连接网络为97.6%, AUC值为0.9982, 两种标记方法的临界相变温度${T_{\rm{C}}} = 1.5$ Figure4. A plot of the result based on the neural network. The x-axis is the temperature, State represents the probability that the sample at each temperature is recognized as a certain state, The letter S represents the state labeling method, the letter T represents the temperature labeling method, AD represents the adsorption state, and DE represents the desorption state. The figure shows the learning results of the two labeling methods. The recognition rate of the convolutional network is 98.3%, the AUC value is 0.9989, the fully connected network is 97.6%, the AUC value is 0.9982, and the critical phase transition temperature is 1.5 of the two labeling methods.
从图4可以看出, 卷积神经网络和全连接神经网络都得到了较高的识别率以及较大的AUC值, 因此神经网络可以较好地识别高分子链在均质表面的两个状态, 且卷积神经网络的识别率稍高一些. 两种方法都可以用来确定其吸附相变点, 且得到的临界相变温度相同, 均为${T_{\rm{C}}} = 1.5$, 稍小于无限链长的临界相变温度${T_{\rm{C}}} = 1.625$[19], 存在这个差值是因为存在有限尺寸效应, 本实验的高分子链长为N = 160, 随着链长的增加, 其临界吸附温度会趋于无限链长的临界吸附温度. 接下来采用神经网络对高分子链在条纹表面的吸附状态识别进行研究. 高分子链在条纹表面的吸附率如图5所示, 插图分别是三种状态下的典型构象, 其中表面上的两种不同条纹对高分子单体有不同的吸附作用, 颜色深的条纹对高分子单体有吸附作用, 白色条纹对高分子单体只有体积排斥作用. 高分子链在条纹表面的构象涉及三种状态, 单条纹吸附态(single-stripe adsorption state)、多条纹吸附态(multi-stripe adsorption state)以及脱附态(desorption state), 因此三种状态之间的转变伴随着两个临界相变点. 图 5 高分子链在条纹表面的吸附率随温度的变化以及典型的三态构象示意图 (a) 单条纹吸附状态, 温度T = 0.3; (b) 多条纹吸附状态, T = 0.9; (c) 脱附状态, T = 3.0. 其中链长N = 160, 条纹宽度L = 4, 条纹方向垂直于x轴, 沿着y轴方向延伸, 选取的空间尺寸为$25 \times 120 \times 20$, 在条纹表面上, 深色部分为吸附条纹, 白色部分为作用力排斥条纹 Figure5. The schematic diagram of the adsorption rate of polymer adsorbed on the stripe surface changes with temperature and typical tri-state conformations: (a) the single-strip adsorption state, where the temperature is 0.3; (b) the multi-strip adsorption state, where the temperature is 0.9; (c) the desorption state, where the temperature is 3.0. Wherein the chain length N is 160, and the stripe width L of the adsorption surface is 4. The stripe direction is perpendicular to the x axis and extends along the y axis, and the selected space size is $25 \times 120 \times 20$. For the adsorption surface, the dark part is the adsorption surface and the white part is the non-force surface.
从图5可以看出, 在高温时高分子链的吸附率几乎为0, 即为脱附态, 该状态与均质表面的脱附态相一致; 在低温时高分子链吸附率非常高, 而且高分子链被单条纹所吸附, 我们把这个吸附状态称为单条纹吸附态, 如图5(a)所示; 而在中间温度存在高分子链吸附在多条纹上, 其吸附率也介于脱附态和单条纹吸附之间, 我们把这个吸附态称为多条纹吸附态. 在多条纹吸附态, 高分子链分布在不同的吸附条纹上, 如图5(b)所示. 对于温度标记法, 我们选取$T \in \left[ {1.35,\;1.5} \right]$的样本作为脱附态标记, $T \in \left[ {0.75,\;0.9} \right]$的样本作为多条纹吸附态标记, $T \in \left[ {0.25,\;0.4} \right]$的样本作为单条纹吸附态标记. 然后我们对在条纹表面上的高分子链构象样本进行了训练与识别, 结果如图6所示. 图 6 神经网络训练的识别结果图 横坐标为温度, 纵坐标State表示每个温度下的样本被识别为某个状态的概率, 图标中S表示状态标记法, T表示温度标记法, SS表示单条纹吸附态, MS表示多条纹吸附态, DE表示脱附态. 其中卷积网络的识别率为94.78%, AUC值为0.9930, 全连接网络为93.85%, AUC值为0.9918, 状态标记法的临界相变温度${T_1} = 0.55$, ${T_2} = 1.1,$温度标记法的临界相变温度${T_1} = 0.55$, ${T_2} = 1.05$ Figure6. A plot of the result of the neural network training. The x-axis is the temperature, the State indicates the pro-bability that the sample at each temperature is recognized as a certain state, S indicates the state labeling method, T indicates the temperature labeling method, SS indicates the single-striped adsorption state, MS indicates the multi-striped adsorption state, and DE indicates desorbed state. The figure shows the learning results of two kinds of labeling methods. The recognition rate of convolutional network is 94.78%, where the AUC value is 0.9930. the fully connected network is 93.85%, where the AUC value is 0.9918, and the critical phase transition temperature of state labeling method is 0.55 and 1.1. The critical phase transition temperature of the temperature labeling method is 0.55 and 1.05.
从图6可以看出, 卷积网络和全连接网络同样具有较高的高分子构象状态的识别率, 且AUC值非常接近1, 这说明神经网络可以识别高分子链在条纹表面的三种状态, 且卷积神经网络识别率略高于全连接网络. 两种样本标记方法所得到的临界相变温度基本相同, 其中多条纹吸附向单条纹吸附的相变点为${T_1} = 0.55$, 脱附态向高分子多条纹吸附的相变点为${T_2} = 1.1$, 与文献[30]计算得到的临界相变温度${T_1} = 0.58$和${T_2} = 1.05$相一致. 完成了上述实验之后, 我们发现条纹的表面的高分子识别率要低于均质表面, 因此我们对识别过程中的误判进行了统计, 如图7所示. 图 7 神经网络学习结果的分布图 (a) 均质表面下的学习结果分布, 绿色表示识别正确的样本, 其他的表示识别错误的样本; (b) 条纹表面下的学习结果分布, 蓝色表示识别正确的样本, 其他的表示识别错误的样本 Figure7. The distribution of neural network learning results: (a) the distribution of learning outcomes on the homogeneous surface, green indicates that the correct sample, and other samples that identify the error; (b) the distribution of learning results on the pattern-stripe surface, blue indicates that the correct sample, and other samples that identify the error.