1. A method for identifying AFM image prediction material performance by a convolutional neural network is characterized in that a deep convolutional neural network is built and trained by a machine learning method, and the network predicts the mechanical property of a polymer material through an AFM phase diagram.

2. The method for identifying AFM image predicted material performance by convolutional neural network of claim 1, wherein the deep convolutional neural network is trained using the data set by taking the synthesized known polymer as the source of the data set, taking the result of the tensile property test by DMA as the label of the data set, and using the phase diagram acquired by AFM as the feature of the data set.

3. The convolutional neural network as claimed in claim 2, wherein the preprocessing of the images in the data set is performed by normalizing the pixel values to the [0, 1] range.

4. The method for identifying AFM image predicted material performance by convolutional neural network of claim 1, wherein said convolutional neural network uses Res-Net network model to build skip connection by residual learning to speed up training.

5. A method for convolutional neural network identification of AFM image predictive material performance as claimed in claim 4 wherein based on the Res-Net network model, the convolutional layer and max pooling layer are used first for input and local response normalization is used to ensure that the previous layer identifies the various modes. Subsequently adding a plurality of different residual error units, each consisting of two convolutional layers and having no pooling layer; with batch normalization and ReLU activation, an n × n kernel is used and the spatial dimensions are preserved.

Background

The mechanical properties of polymer materials are very important properties in material applications, such as tensile properties, fracture toughness, fatigue properties, impact toughness and the like, and are also factors which must be considered in the production and manufacturing process, and especially in the copolymer with a complex structure, the indexes of the mechanical properties are more variable. In the laboratory, after a new polymer product is synthesized, the surface structure is generally observed manually by an Atomic Force Microscope (AFM), and then a tensile property test is performed by dynamic thermomechanical analysis DMA to verify the mechanical properties.

The AFM image is observed manually, so that on one hand, the AFM is difficult to quantitatively give a specific tensile property value, and on the other hand, the DMA test result is difficult to explain the deep microscopic mechanism; on the other hand, observation AFM personnel need to be trained in advance, generally, more information can be read from the image by experiential researchers, different researchers can generate different human errors during reading, accuracy is greatly influenced by a main appearance, and efficiency is low.

Disclosure of Invention

The invention aims to provide a method for identifying AFM image prediction material performance by a convolutional neural network, aiming at the problems that the traditional AFM phase diagram is low in reading efficiency, poor in accuracy, large in limitation and difficult to popularize experience, and combining with a machine vision technology which is rapidly developed in the field of deep learning to carry out artificial intelligence identification. The invention can rapidly read the AFM image information through the deep neural network and accurately predict the relevant mechanical properties of the material.

The purpose of the invention is realized by the following technical scheme: a method for identifying AFM image prediction material performance by a convolutional neural network is characterized in that a deep convolutional neural network is built and trained by a machine learning method, and the network predicts the mechanical property of a polymer material through an AFM phase diagram.

Further, the synthesized known polymer is used as a data set source, the result of a tensile property test through DMA is used as a data set label, the phase diagram acquired by AFM is used as a characteristic of the data set, and the data set is used for training the deep convolutional neural network.

Further, the image in the data set is preprocessed by normalizing the pixel values to be in the range of [0, 1 ].

Furthermore, the convolutional neural network uses a Res-Net network model, skip connection is established through residual learning, and the training speed is increased.

Further, based on a Res-Net network model, firstly, a convolutional layer and a maximum pooling layer are used for inputting, local response normalization is used for ensuring that the previous layer identifies various modes, and then a plurality of different residual error units are added, wherein each residual error unit consists of two convolutional layers and has no pooling layer; with batch normalization and ReLU activation, an n × n kernel is used and the spatial dimensions are preserved.

The invention has the beneficial effects that:

(1) the method identifies the AFM image by using the deep convolutional neural network, associates the characteristics of the image with the mechanical properties of the corresponding material, belongs to the cross field of material science and deep learning computer vision, can help to identify rules which are difficult to observe by human eyes in the AFM image by virtue of the advantages of deep learning, and the mechanical properties of the material are probably dependent on the characteristics of the rules, thereby further widening the application occasions of the AFM image and helping to further research the relationship between the phase region distribution of the material and the mechanical properties of the material in the field of material science;

(2) by using the Res-Net residual convolution neural network, the convolution layer and the pooling layer can better identify key features of the image, skip links are added in the deep convolution process, signals fed into the layer are also added into the output of the layer positioned above the stack, and the network can better solve the problems that the model is deeper and the parameters are fewer and fewer, so that the features in the image are analyzed, and the rule is found and identified. In addition, a dropout regularization technology is added into the network to reduce data characteristics and reduce the degree of freedom of a model, so that the dependence of a training process on a data set is weakened, the contradiction between a large data driving technology and less time-consuming data of experiments in the field of materials is avoided, and the network can better analyze the relation between the AFM image and the mechanical property of the corresponding material.

Drawings

FIG. 1 is a schematic diagram of the research framework of the present invention;

FIG. 2 is a model schematic diagram of the Res-Net convolutional neural network of the present invention.

Detailed Description

As shown in FIG. 1, the method for identifying AFM image prediction material performance by the convolutional neural network comprises the following steps:

step 1, obtaining Atomic Force Microscope (AFM) images of different polymer materials, and marking the images based on the relative mechanical property results of the materials. The method comprises the steps of synthesizing block copolymers consisting of different monomers, gradient copolymers, homopolymers and other polymer materials by program control feeding, and testing the tensile property of the polymer materials by using DMA (direct memory access), so as to obtain the breaking elongation of the polymer materials with different monomers in different arrangement combinations under different synthesis conditions, wherein the breaking elongation is used as a label of an original data set. And obtaining a phase diagram of the polymer by using AFM (atomic force microscopy) as a characteristic of the data set, and obtaining distribution information of a polymer crystal region and an amorphous region.

Step 2, classifying and preprocessing the image, including centralization, standardization and normalization; and then, a training set, a verification set and a test set are divided, so that subsequent network training is facilitated. And training the deep convolutional neural network by using a training set and a verification set, and testing the result by using a test set. Preprocessing the image by normalizing the pixel values to be in the range of [0, 1] by numpy (all values are initially in the range of [0, 255 ]) avoids the disappearance of gradients that may be generated during training.

Step 3, building a Res-Net deep neural network framework, and selecting a proper optimizer, a loss function, a regularization method and iteration times; and the method is used for extracting the features of the AFM phase diagram to obtain the corresponding feature neural network. A training data generator (imagedata generator) was set up in Python using the TensorFlow framework; and (3) building by using a Sequential model, and avoiding the overfitting problem caused by too small data volume by using dropout regularization. Based on a Res-Net network model, skip connection is established through residual learning, so that the training speed can be greatly increased; as shown in fig. 2, in this network, we first use the convolutional layer and the max pooling layer for input, use the larger kernel to preserve data features as much as possible, and use local response normalization to ensure that the previous layer identifies the various modes, then add a number of different residual units, each consisting of two convolutional layers (without pooling layer), with batch normalization and ReLU activation, use 3 x 3 kernels and preserve spatial dimensions.

And 4, training a neural network based on the obtained data set to obtain a training model, and predicting the mechanical property (elongation at break) of the polymer material by identifying the relevant rule of an Atomic Force Microscope (AFM) phase diagram by using the network. The pictures obtained in the source AFM were read, converted to float32 multi-dimensional arrays, and the image data (along with their labels) were fed back to the neuron network.

Example 1:

step 1, controlling feeding through a program, synthesizing a gradient copolymer of styrene and n-butyl acrylate by RAFT emulsion polymerization, and sending the synthesized gradient copolymer to an Atomic Force Microscope (AFM) for slice inspection and observation of the surface morphology of the gradient copolymer to obtain an AFM phase diagram. Atomic Force Microscope (AFM) phase image preprocessing in the data set in step 1 before step 2, clipping dpi300 × 300 by using a training data generator (imagedata generator), and performing normalization processing to compress the values in RGB three channels [0, 255] to between [0, 1 ].

Step 2, building a deep convolutional neural network, wherein an input layer of the deep convolutional neural network is composed of 64 7 × 7 filters, then linking a maximum pooling layer containing 64 3 × 3 filters, each residual unit is composed of two convolutional layers, the 64 × 3 filters are contained in the residual unit, Batch Normalization (BN) and ReLU activation are achieved, a 3 × 3 kernel is used for preserving space dimensionality, then a Flatten layer is used for flattening, a full-connection network containing 1024 neurons is used as output, and a softmax activation function is adopted.

And 3, selecting a plurality of images from the AFM phase diagram data set for training, dividing the data set into a training set and a testing set by using a train _ test _ split function in sklern, and selecting the ratio of the training set to the testing set to be 0.9: 0.1. after the data set is divided, the training set is divided into a training set and a verification set so as to train the deep neural network. And further normalizing and centralizing the pictures, and ensuring the stability of the data by adopting the same picture processing operation on the training set, the verification set and the test set.

And 4, importing the training set and the verification set into the built deep neural network, adopting an SGD optimizer, training the data set by using an mse loss function, verifying the training batches by using the verification set in real time to finally obtain the trained deep convolutional neural network, testing the deep convolutional neural network on the test set, and completing prediction of the AFM image recognized by the convolutional neural network on the mechanical property of the copolymer.

Example 2:

step 1, controlling feeding through a program, synthesizing a block copolymer of styrene and n-butyl acrylate by RAFT emulsion polymerization, respectively synthesizing homopolymers of polystyrene and poly-n-butyl acrylate, and conveying the synthesized polymer material to an Atomic Force Microscope (AFM) for slice inspection and observation of the surface morphology of the polymer material to obtain an AFM phase diagram. Atomic Force Microscope (AFM) phase image preprocessing in the data set in step 1 before step 2, clipping dpi300 × 300 by using a training data generator (imagedata generator), and performing normalization processing to compress the values in RGB three channels [0, 255] to between [0, 1 ].

Step 2, building a deep convolutional neural network, wherein an input layer of the deep convolutional neural network is composed of 64 7 × 7 filters, then linking a maximum pooling layer containing 64 3 × 3 filters, each residual unit is composed of two convolutional layers, the 64 3 × 3 filters are contained, Batch Normalization (BN) and ReLU activation are achieved, a 3 × 3 kernel is used for preserving space dimensionality, a dropout is used for conducting regularization limiting model, then a Flatten layer is used for flattening, a full-connection network containing 1024 neurons is used as output, and a sigmoid activation function is adopted.

And 3, selecting a plurality of images from the AFM phase diagram data set for training, dividing the data set into a training set and a testing set by using a train _ test _ split function in sklern, and selecting the ratio of the training set to the testing set to be 0.8: 0.2, since the dropout regularization technology is used in the network, the same effect is obtained by training in consideration of reducing the data volume, and the rest of the image processing technology is the same as that in embodiment 1.

And 4, importing the training set and the verification set into the built deep neural network, adopting an RMSProp optimizer, setting the learning rate to be 0.001 and the attenuation rate to be 0.9, training the data set by using a mse loss function, reading the loss and the accuracy of the training set and the verification set, setting the epochs to be 100, and finally obtaining the trained deep convolutional neural network, wherein the operation of the training and testing process is consistent with that of the embodiment 1.

Example 3:

step 1, controlling feeding through a program, synthesizing a gradient copolymer of styrene and n-butyl acrylate by using RAFT emulsion polymerization, synthesizing a gradient copolymer of styrene and butadiene by using RAFT emulsion polymerization, respectively synthesizing homopolymers of three monomers, and conveying the synthesized polymer material to an Atomic Force Microscope (AFM) for slicing inspection and observing the surface morphology of the polymer material to obtain an AFM phase diagram. When the image is marked, besides the corresponding mechanical properties, molecular descriptors of monomer molecules are added for distinguishing, and the other means are the same as those in the embodiment 1.

And 2, building a deep convolution neural network, wherein the structure of the neural network is consistent with that of the embodiment 2.

And 3, dividing a data set, wherein the basic steps are consistent with those of the embodiment 2, and the method is consistent.

And 4, importing the training set and the verification set into the built deep neural network, adopting an Adam optimizer due to the fact that the data is more complex, training the data set by using a mse loss function, reading the loss and accuracy of the training set and the verification set at the same time, setting epochs to be 200, finally obtaining the trained deep convolutional neural network, wherein the operation of the training and testing process is consistent with that of the embodiment 1.