Organic photovoltaic material performance prediction method based on machine learning

Abstract

The invention discloses a machine learning-based organic photovoltaic material performance prediction method, which comprises the following steps of: firstly, training a model on a labeled data set, namely a teacher model; label prediction is carried out on the unlabeled data set, and the last layer is predicted; next, training a new model called a student model, and training a labeled data set and an unlabeled data set based on the big data set; and finally, labeling the unlabeled data set again by using the learned student model, returning to the second step, and carrying out loop iteration to obtain a final machine learning prediction model. According to the method, the model accuracy can be improved by continuously and cyclically iterating the student model and the teacher model.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic materials, and in particular relates to a method for predicting the performance of organic photovoltaic materials based on machine learning. Background Art

[0002] The macroscopic properties of materials are not only related to their chemical composition and process, but also microstructure is one of the important influencing factors. Since the mid-20th century, with the continuous improvement of relevant theories of materials science and the substantial improvement of computer computing power, the third paradigm trend of materials science based on numerical simulation has promoted the development of molecular dynamics, first principles and finite element simulation theories and methods. These theories have greatly promoted the study of the relationship between the microstructure of materials and their properties. However, this research method based on numerical simulation still has difficulty in determining the relationship between material structure, properties and performance. In the face of complex problems, the amount of calculation increases geometrically, which not only consumes a lot of time, but also has high requirements on the theoretical knowledge and experimental ability of scientific researchers.

[0003] Since the 21st century, the computing power of computers has increased rapidly in the manner of Moore's Law, and the fourth paradigm of materials science driven by data has become the mainstream of research. In the traditional material research and development process, material metallographic images and numerical simulation images are also studied to explore the influence of material microstructure on macroscopic physical properties, and a large amount of valuable material image data has been accumulated. By using microscopic technology to collect material microstructures or using high-throughput preparation and characterization technology to generate a large amount of material microstructure data, the performance is manually annotated in detail to form a data set that can be used for convolutional neural network training, training recognition, classification or detection networks, greatly improving the efficiency and accuracy of image analysis, and reducing the workload of scientific researchers. At the same time, the principle of automatic update of feature extractors by convolutional neural networks avoids falling into experience blind spots.

[0004] In today's society, as the contradiction between the ever-expanding energy demand and the worsening environmental pollution continues to intensify, humans are paying more and more attention to the utilization and development of clean energy such as wind energy and solar energy. Organic photovoltaic semiconductor materials are currently a relatively cutting-edge research field. The reason is that traditional photovoltaic materials have many disadvantages such as expensive raw materials, complex production processes, and low energy conversion rates.

[0005] Organic photovoltaic semiconductor materials are energy collection devices that use organic materials to convert solar energy. Compared with traditional silicon-based materials, they have the advantages of flexibility, transparency and easy manufacturing, so they are widely used in flexible circuits, transistors and LED devices. The working principle of organic photovoltaic semiconductor materials is as follows: under the irradiation of sunlight, photons with energy are absorbed into the semiconductor, which excites the semiconductor electrons to migrate from the valence band to the conduction band, and a gap appears in the original position. This gap is called a "hole"; the hole and the excited electron move freely toward the counter electrode, and at the same time, the two are bound to each other to produce excitons; in the electric field, the electrons and holes separate and move to form charged carriers, and the mutual migration of charge carriers forms an electric current, which converts light energy into electrical energy.

[0006] Microstructure determines the photovoltaic performance of organic semiconductor films. The relationship between microstructure and performance is usually highly nonlinear and extremely difficult to evaluate, making the optimization of photovoltaic material microstructure challenging. Using convolutional neural networks as "feature extractors" to allow the network to automatically find features that fit the structure and performance prediction model through gradient reduction of the loss function has become a new research direction for studying the relationship between the microstructure and performance of organic photovoltaic semiconductor materials.

[0007] The dataset used comes from the research of Balaji Sesha Sarath Pokuri, Sambuddha Ghosal and others. This study used thermodynamically consistent binary phases to solve the Cahn-Hilliard equation to separate and simulate the microstructure dataset. The dataset contains nearly 65,000 image data with a resolution of 101×101. To evaluate the photovoltaic performance, the exciton drift diffusion equation is used to calculate the short-circuit current Jsc of each image. In the entire dataset, the minimum value of Jsc is 0.6mA / cm2 and the maximum value is 7.0mA / cm2. Therefore, the short-circuit current can be divided into 10 equal parts in the [0.6, 7] interval, and then the data samples are divided into 10 categories according to the short-circuit current performance of the material to set the data labels. Finally, a convolutional neural network is established, and the simulated images are used as the training dataset to complete the forward material performance prediction.

[0008] The algorithm is mainly divided into two parts:

[0009] Data preprocessing and model design. Obtain material microscopic image data from public material databases, open source codes or materials. Clean the data, remove data with incomplete or missing labels, normalize performance parameters, and use image enhancement methods such as translation, rotation, flipping, shearing, and scaling to expand the data. Finally, establish a data set that can be directly used by the model. The backbone network of the model is an existing framework such as DenseNet.

[0010] Forward material performance prediction. The constructed model is trained and optimized. After training, the model can achieve an ideal fitting effect. After the model is determined, the prediction of material microstructure and performance can be realized.

[0011] Among the methods introduced above, the machine learning-based method has the characteristics of short response time, fast speed, and generalization of extracted features. Compared with other methods, it has certain advantages in prediction effect.

[0012] The organic photovoltaic material performance prediction method based on machine learning has the ability to learn features at multiple levels. The features can be expressed in a diverse manner, with better discriminability and greater applicability. Summary of the invention

[0013] In order to overcome the shortcomings of the prior art, the present invention provides a method for predicting the performance of organic photovoltaic materials based on machine learning. First, a model is trained on a labeled data set, called a teacher model; then, label prediction is performed on an unlabeled data set to predict the last layer; next, a new model is trained, called a student model, with a labeled data set and an unlabeled data set, and then trained based on this large data set; finally, the unlabeled data set is re-labeled with the learned student model, and the result is returned to the second step, and the final machine learning prediction model can be obtained by cyclic iteration. The present invention can improve the model accuracy by continuously cyclically iterating the student model and the teacher model.

[0014] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0015] Step 1: Train a model on a labeled dataset, called a teacher model;

[0016] Step 2: Use the teacher model f obtained in step 1 to perform label prediction on the unlabeled data set, and predict the probability distribution given by the last layer of softmax, which is called soft label;

[0017] Step 3: Train a new model called the student model, combine the labeled data set and the unlabeled data set with pseudo labels classified by the teacher model into a large data set, and perform training based on the large data set;

[0018] Step 4: Re-label the unlabeled dataset using the learned student model, return to step 2, and iterate the loop to obtain the final machine learning prediction model.

[0019] Furthermore, the step 1 is specifically as follows:

[0020] The labeled data in the label dataset is represented as {(x1,y1),(x2,y2),...,(x n ,y n)}, the unlabeled data generated by image enhancement is represented as Train a convolutional neural network model on a labeled dataset, called the teacher model. The parameters of the teacher model are θ, and the teacher model uses cross entropy loss:

[0021]

[0022] Among them, n represents the total number of categories, y i represents the label, l(.) represents the cross entropy, f(.) represents the convolutional neural network, x i Represents data.

[0023] Furthermore, the step 2 is specifically as follows:

[0024] Step 2-1: The output of the neural network is:

[0025]

[0026] in, represents the neural network output, represents unlabeled data, and m represents the total number of categories;

[0027] Step 2-2: Let the convolutional network be called N, and its i-th convolutional layer be expressed as:

[0028] Y i =F i (X i ) (3)

[0029] Where X i is the input data, Y i is the output of the convolution layer. If the teacher model has k convolution layers, the model is represented as:

[0030] N=F k ⊙...⊙F2⊙F1(X1)=⊙ j＝1,...,k F j (X1) (4)

[0031] Multiple convolutional layers with the same structure are called a stage, and the convolutional network N is expressed in stages as follows:

[0032]

[0033] in, <H i ,W i ,C i > represents the dimension of the input tensor of the i-th layer, and the subscript i represents the sequence number of the stage. represents the i-th stage, which consists of the convolutional layer F i Repeat L i Secondary composition;

[0034] Step 2-3: The internal structure of the network model is implemented by multiple MBConv convolution blocks, which are composed of depthwise convolution and SENet.

[0035] The MBConv convolution block uses the Swish activation function, and the formula is as follows:

[0036] f(x)=x·sigmoid(βx) (9)

[0037] SENet first performs a Squeeze operation to compress features along the spatial dimension, converting each two-dimensional feature channel into a real number, and the output dimension matches the number of input feature channels; secondly, an Excitation operation is performed to generate weights for each feature channel through the parameter w, where the parameter w is learned to explicitly model the correlation between feature channels; finally, a Reweight operation is performed to use the weights output by Excitation as the importance of each feature channel after feature selection, and then weighted to the previous features channel by channel through multiplication, completing the recalibration of the original features in the channel dimension;

[0038] Step 2-4: Adjust model parameters based on EfficientNet-B0 to increase network depth, network width, and input image resolution.

[0039] Furthermore, the step 3 is specifically as follows:

[0040] Add noise to the student model and use cross entropy loss:

[0041]

[0042] For image datasets consisting of labeled data, divide them into training sets, validation sets, and test sets;

[0043] For unlabeled data, the EfficientNet-B0 network is used to predict the label of each image data.

[0044] Furthermore, Batch Normalization is used in the model, which is expressed as follows:

[0045]

[0046] Among them, μ and σ are the mean and variance of z respectively, ε is a number tending to 0, and the distribution of the node input value z is re-normalized to a normal distribution with a mean of 0 and a variance of 1;

[0047] Batch Normalization performs a scale operation on the input values ​​that satisfy the standard distribution, as shown in formula (8):

[0048] z′=α·z_norm+β (8)

[0049] In the formula, α and β are two additional parameters for each neuron that can be obtained through learning.

[0050] Furthermore, the EfficientNet-B0 consists of 16 MBConvs, two Convs, a global average pooling layer, and a fully connected layer.

[0051] The beneficial effects of the present invention are as follows:

[0052] In order to study the effect of different models on the prediction of simulated images of organic photovoltaic semiconductor materials and improve the prediction accuracy, this paper sets up three model comparison experiments: ResNet-50, DenseNet-169, and EfficientNet-B0. ResNet-50 consists of 4 modules, each of which consists of 3, 4, 6, 3 small modules plus a fully connected layer, with a total of 50 layers and a parameter volume of 25.6M. DenseNet-169 is mainly composed of 4 dense blocks and 3 transition layers, with a parameter volume of 14.31M. The model structure of EfficientNet-B0 is shown in Table 4-1, with a parameter volume of 5.3M. The training does not use a pre-trained model, because most training models are trained based on natural images, and the characteristics of simulated images of organic photovoltaic semiconductors are quite different from natural images. Therefore, the entire model is trained from scratch. There are 10 image categories in total, so the last fully connected layer is set with 10 neurons plus a softmax activation function. The performance of each model after training 70 epochs is shown in Table 1.

[0053] Table 1 Performance comparison of each model

[0054]

[0055] From the above chart, we can see that the classification accuracy of EfficientNet-B0 is as high as 92.3%, and the F1-score is 84.5, slightly higher than ResNet-50 and DenseNet-169. During training, the time required for one round of training is only 180 seconds, which is much lower than ResNet-50 and DenseNet-169. This fully proves that the network model has many advantages such as good prediction performance, low computing resource consumption, strong positive sample recognition ability and strong anti-interference ability. Therefore, this paper chooses it as the backbone network for prediction.

[0056] The performance of the model is directly related to the depth, width, and image resolution. Many convolutional neural networks adopt the strategy of scaling depth. The deeper the network, the stronger its ability to capture complex features and the better its generalization ability in the test set. However, the increasing depth of the model will lead to the problem of gradient vanishing. Although optimization methods such as cross-layer connections can reduce this training problem, the benefits of deep networks will be small. Increasing the width of the model can enhance the model's ability to capture small features, but wide and shallow networks have difficulty capturing high-level features. Increasing the image resolution can also enhance the model's ability to capture small features. In recent years, the resolution of many network image inputs has continued to increase, and target detection even uses 600*600 images, but the higher the image resolution, the smaller the model benefit.

[0057] The network uses a hybrid dimensionality magnification method to debug the three dimensions of depth, width, and image resolution compared with the baseline model EfficientNet-B0.

[0058] Five models from EfficientNet-B0 to EfficientNet-B4 were built for parameter adjustment. The scaling ratios of each model are shown in Table 2.

[0059] Table 2 Parameter scaling ratios of each model

[0060]

[0061] From Table 4-3, we can see that as the model scaling ratio continues to increase, the number of model parameters continues to increase, and the Dropout ratio continues to increase. This is to prevent overfitting caused by the increase in model parameters. After training for 70 epochs, the performance of the five models from EfficientNet-B0 to EfficientNet-B4 is shown in Table 3.

[0062] Table 3 Performance comparison of EfficientNet-B0 to EfficientNet-B4 models

[0063]

[0064] It can be seen that as the number of EfficientNet parameters continues to increase, the accuracy of model prediction will continue to increase, but the increase will gradually decrease. The accuracy of EfficientNet-B4 is only 0.02 higher than that of EfficientNet-B3, but the computing resources are much higher than EfficientNet-B3. Considering the model performance and computing resource consumption, EfficientNet-B3 was finally selected as the best EfficientNet backbone network.

[0065] Before the iterative training starts, the EfficientNet-B3 trained in the previous section is used as the teacher model to predict the unlabeled data, and the image augmentation method is used to increase the data noise. The data with soft labels and the original data are merged into a new training set, and a new EfficientNet-B3 is constructed as the student model. The new training set is used for training, and the model noise is increased using random depth to prevent the model from overfitting. After the student model is trained, it is set as the teacher model to re-predict the unlabeled data, and the cycle is repeated. When the accuracy increase is no more than 0.1% after one iteration, the iteration is stopped. The changes in the number of iterations and the accuracy are shown in Table 4.

[0066] Table 4. The number of iterations improves model performance

[0067]

[0068] It can be seen that the model accuracy is greatly improved after the first iteration. This is because the training set adds unlabeled data in the first iteration. The larger the data set, the more features the model can learn. As the number of iterations increases, the model accuracy increases steadily, but when the fourth iteration is reached, the accuracy is only 0.04% higher than the third round, which is lower than 0.1%. The iteration is stopped and the final model accuracy is 96.55%. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a flow chart of an image classification algorithm based on machine learning in an embodiment of the present invention.

[0070] Figure 2 Schematic diagram of parameter adjustment of the EfficientNet model in an embodiment of the present invention, (a) baseline model, (b) width scaling, (c) depth scaling, (d) image size scaling, and (d) composite scaling.

[0071] Figure 3 This is a comparison chart of the accuracy and parameter quantity of each model of the present invention.

[0072] Figure 4 This is a graph showing how the performance of EfficientNet of the present invention changes as the number of parameters increases.

[0073] Figure 5 This is a comparison chart between the number of iterations and the model accuracy of the present invention. DETAILED DESCRIPTION

[0074] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0075] The purpose of the present invention is to provide a method for predicting the performance of organic photovoltaic materials based on machine learning. According to the short-circuit current performance of the material, the data samples are divided into 10 categories and data labels are set. Finally, a convolutional neural network is established, and the simulated image is used as the training data set to complete the forward material performance prediction. An organic photovoltaic semiconductor material classification algorithm based on semi-supervised learning is adopted. First, a teacher model is trained, and then the teacher model is used to generate pseudo labels on the unlabeled data, and the labeled and pseudo-labeled data are used to train the student network together. This cycle is repeated several times to generate a finally trained network.

[0076] Figure 1 This is a flow chart of an organic photovoltaic material performance prediction based on machine learning in this embodiment, and its specific implementation steps are as follows:

[0077] Step 1: Train a model on a labeled dataset, called the teacher model. The model parameters are θ and the model uses cross entropy loss:

[0078]

[0079] Convolutional neural networks are usually developed under a fixed resource budget. If more resources are available, they will be scaled up to obtain better accuracy, such as increasing the network depth, network width, and input image resolution. However, it is difficult to manually adjust the enlargement or reduction of depth, width, and resolution. When the amount of computation is limited, it is difficult to determine which one to enlarge and which one to reduce. In other words, the combination space is too large to be exhausted by human power. Based on the above background, the author of EfficientNet proposed a new model scaling method, which uses a simple and efficient compound coefficient to enlarge the network from the three dimensions of depth, width, and resolution. It does not arbitrarily scale the dimensions of the network like traditional methods, and can obtain the optimal set of parameters based on neural structure search technology.

[0080] Step 2: Use the model f obtained in step 1 to predict labels on the unlabeled dataset. The prediction of the last softmax layer gives a probability distribution, called a soft label, because it only gives the confidence of each category instead of specifying a specific category. The hard label is the result label after the one-hot max. Experiments have shown that soft labels are better.

[0081]

[0082] like Figure 2It is a parameter adjustment diagram of the EfficientNet model of an embodiment of the present invention, including a baseline model, which is adjusted comprehensively by adding convolution kernel channels to scale the width, changing the number of network layers to scale the depth, and scaling the original image size.

[0083] The mathematical abstract representation of the model adjustment process refers to the entire convolutional network as N, and its i-th convolutional layer can be expressed as:

[0084] Y i =F i (X i ) (3)

[0085] Where X i is the input data, Y i is the output of the convolution layer. If the model has k convolution layers, the model is represented as follows:

[0086] N=F k ⊙...⊙F2⊙F1(X1)=⊙ j＝1,...,k F j (X1) (4)

[0087] However, in practice, multiple convolutional layers with the same structure are usually called a stage. For example, ResNet has 5 stages, and the convolutional layer structure in each stage is the same (except that the first layer is a downsampling layer). The convolutional network N can be expressed in stages as follows:

[0088]

[0089] in, <H i ,W i ,C i > represents the dimension of the input tensor of the i-th layer, and the subscript i (from 1 to s) represents the sequence number of the stage. represents the i-th stage, which consists of the convolutional layer F i Repeat L i Second composition.

[0090] Different from the usual convolutional layer design, the design focuses on finding the best network layer F i , model scaling attempts to expand the network length, width, or image size without changing the predefined F in the baseline network i .

[0091] The internal structure of the EfficientNet model is implemented through multiple MBConv convolution blocks. MBConv stands for mobile inverted bottleneck convolution, which is composed of depthwise convolution and SENet. The purpose of Batch Normalization is to standardize the distribution type of input values ​​of each node and speed up the convergence and training process.

[0092]

[0093] Among them, μ and σ are the mean and variance of z, which are numbers tending to 0 to avoid the denominator being zero. The distribution of the node input value z is re-normalized to a normal distribution with a mean of 0 and a variance of 1, so that it falls in the area where the activation function is more sensitive. At this time, even if the input value changes slightly, the loss function will also change significantly. In other words, the above operation can increase the gradient, and increasing the gradient means accelerating the training convergence speed. However, only the above distribution normalization operation will lead to the linearization of the nonlinear activation function, and the deepening of the network becomes meaningless. To address this problem, Batch Normalization performs a scale operation on the input values ​​that meet the standard distribution, as shown in formula (8):

[0094] z′=α·z_norm+β (7)

[0095] In the formula, α and β are two new parameters for each neuron that can be learned. A large number of experiments have verified that the above Batch Normalization operation can not only speed up the convergence and training process of the network, but also avoid overfitting and thus improve the classification effect of the network.

[0096] MBConv uses the Swish activation function. Swish is a new type of activation function that has the characteristics of unsaturation, smoothness, and non-monotonicity. Multiple tests have shown that the performance of Swish and Swish-B activation functions is excellent, and they have shown better performance than the current best activation function on different data sets. The formula is as follows:

[0097] f(x)=x·sigmoid(βx) (8)

[0098] The saturation of the sigmoid function easily leads to the disappearance of the gradient. By referring to the effect of ReLU, when x is very large, f(x) approaches x, but when x approaches negative infinity, f(x) approaches 0. The general trend of the function is similar to ReLU, but it is more complicated than ReLU.

[0099] MBConv introduced SENet. SENet first performs a Squeeze operation, which compresses features along the spatial dimension and converts each two-dimensional feature channel into a real number. This real number has a global receptive field to some extent, and the output dimension matches the number of input feature channels. It represents the global distribution of responses on the feature channels, and enables the layers close to the input to obtain a global receptive field, which is very useful in many tasks. The second is the Excitation operation, which is a mechanism similar to the gate in a recurrent neural network. The weight is generated for each feature channel through the parameter w, where the parameter w is learned to explicitly model the correlation between feature channels.

[0100] The last step is a Reweight operation, which regards the weight of Excitation's output as the importance of each feature channel after feature selection, and then weights it to the previous features channel by channel through multiplication to complete the recalibration of the original features in the channel dimension. The importance of each feature channel is automatically obtained through learning, and then useful features are promoted according to this importance and features that are not very useful for the current task are suppressed.

[0101] EfficientNet-B0 network structure EfficientNet-B0 is the benchmark model of EfficientNet and also the smallest model. The model parameters are adjusted based on EfficientNet-B0 to improve the network depth, network width, and input image resolution. EfficientNet-B0 consists of 16 MBConvs, two Convs, a global average pooling layer, and a fully connected layer.

[0102] Step 3: Combine the labeled dataset and the unlabeled dataset (with pseudo labels after classification by the teacher model) into a large dataset to train a new model called the student model, and then train based on this large dataset. Add noise to the student model using the cross entropy loss:

[0103]

[0104] The organic photovoltaic semiconductor material simulation image dataset contains 65,000 labeled data. The dataset is divided into training set, validation set and test set in a ratio of 8:1:1. The training set has 52,000 data, and the validation set and test set each contain 6,500 data. In order to make full use of all the data, this part of the data was not screened or sampled.

[0105] Unlabeled data, using a generative adversarial network to obtain 20,000 organic photovoltaic semiconductor simulation images. These images are unlabeled, and the EfficientNet-B0 network is used to predict the label of each image data. Labeled data labels are hard labels and unique hot labels, while the labels predicted by the EfficientNet network are soft labels, and the output is the probability of each category. Soft labels contain more information and improve the model more than hard labels. Then this part of the data is filtered and balanced. Instead of using all the data, the data with a confidence level higher than 0.6 is selected. A total of 15,324 data that meet the requirements are counted according to category.

[0106] Step 4: Re-label the unlabeled dataset using the learned student model, return to the second step, and iterate the loop to obtain the final machine learning prediction model.

[0107] Example:

[0108] Multiple control experiments were set up and multiple different models were constructed. In the comparison experiment of organic photovoltaic semiconductor material image classification using different models, three models, ResNet-50, DenseNet-169, and EfficientNet-B0, were constructed. The ResNet-50 and DenseNet-169 models were built using the Keras framework, and the EfficientNet-B0 was built using pytorch. In the process of adjusting the parameters of the EfficientNet network, five models, EfficientNet-B0 to EfficientNet-B4, were constructed. The larger the model, the more parameters it has, and the stronger its ability to learn unlabeled data, but it also occupies more resources and takes more training time.

[0109] The training process uses a batch training method. The training data set is large, and the amount of computation required to load all the data at once is too large, so a strategy of reading data in batches is adopted. The batch training method can quickly converge to the global minimum, and the loss decline curve is very smooth. If the amount of data in a batch is too large, it will increase the computational consumption, so the amount of data in a batch is set to 256. The Adam optimization algorithm is used to reduce the training loss. The initial learning rate is 0.01, belt_1 is set to 0.9, and belt_2 is set to 0.999. The Adam optimization algorithm can quickly and smoothly converge the loss function to the global minimum. Measuring model performance standards: Accuracy and F1-score are used to measure model performance. Accuracy is the proportion of samples with correct labels predicted by the model to the total samples, and is the most commonly used indicator for measuring model classification performance.

[0110]

[0111] The recall rate is the ratio of positive samples correctly identified by the model to all positive samples, reflecting the model's ability to identify positive samples. The formula is as follows:

[0112]

[0113] Precision is the ratio of truly positive samples among samples predicted by the model, reflecting the model's ability to resist interference from negative samples. The formula is as follows:

[0114]

[0115] To comprehensively evaluate the performance of the model, recall and precision are usually calculated at the same time. However, precision and recall are negatively correlated, that is, the higher the precision, the more stringent the conditions for the model to predict a positive sample. As the number of samples predicted as positive samples decreases, the recall will naturally decrease. Therefore, F1-score is used to measure model performance. F1-score is the harmonic mean of recall and precision. The specific formula is as follows:

[0116]

[0117] In order to study the effects of different models on the prediction of simulated images of organic photovoltaic semiconductor materials and improve the prediction accuracy, the present invention sets up three model comparison experiments: ResNet-50, DenseNet-169, and EfficientNet-B0. ResNet-50 consists of 4 modules, each of which is composed of 3, 4, 6, 3 small modules plus a fully connected layer, with a total of 50 layers and a parameter volume of 25.6M. DenseNet-169 is mainly composed of 4 dense blocks and 3 transition layers, with a parameter volume of 14.31M. The parameter volume of the EfficientNet-B0 model is 5.3M. The training does not use a pre-trained model, because most of the training models are trained based on natural images, and the characteristics of the simulated images of organic photovoltaic semiconductors are quite different from those of natural images. Therefore, the entire model is trained from scratch. There are 10 image categories in total, so the last fully connected layer is set with 10 neurons plus a softmax activation function. Figure 3 It can be concluded that the classification accuracy of EfficientNet-B0 is as high as 92.3%, and the F1-score is 84.5, slightly higher than ResNet-50 and DenseNet-169. During training, the time required for one round of training is only 180 seconds, which is much lower than ResNet-50 and DenseNet-169. This fully proves that the EfficientNet network model has many advantages such as good prediction performance, low computing resource consumption, strong positive sample recognition ability and strong anti-interference ability. Therefore, the present invention selects EfficientNet as the backbone network for prediction.

[0118] The performance of the model is directly related to the depth, width, and image resolution. Many convolutional neural networks adopt the strategy of scaling depth. The deeper the network, the stronger its ability to capture complex features and the better its generalization ability in the test set. However, the increasing depth of the model will lead to the problem of gradient vanishing. Although optimization methods such as cross-layer connections can reduce this training problem, the benefits of deep networks will be small. Increasing the width of the model can enhance the model's ability to capture small features, but wide and shallow networks have difficulty capturing high-level features. Increasing the image resolution can also enhance the model's ability to capture small features. In recent years, the resolution of many network image inputs has continued to increase, and target detection even uses 600*600 images, but the higher the image resolution, the smaller the model benefit.

[0119] EfficientNet uses a hybrid dimension magnification method to debug the three dimensions of depth, width, and image resolution compared to the benchmark model EfficientNet-B0. The parameter adjustment formula is as follows:

[0120] depth:d=α φ

[0121] width:w=β φ

[0122] resolution:r=γ φ

[0123] stα·β 2 γ 2 ≈2

[0124] α≥1,β≥1,γ≥1 (14)

[0125] The mixing coefficient φ is determined by the computing resources. α, β, and γ are hyperparameters that need to be debugged. Different parameters construct different models.

[0126] Figure 4 It can be seen that as the number of EfficientNet parameters continues to increase, the accuracy of model prediction will continue to increase, but the increase will gradually decrease. The accuracy of EfficientNet-B4 is only 0.02 higher than that of EfficientNet-B3, but the computing resources are much higher than EfficientNet-B3. Considering the model performance and computing resource consumption, EfficientNet-B3 was finally selected as the best EfficientNet backbone network.

[0127] Figure 5It can be seen that the model accuracy is greatly improved after the first iteration. This is because the training set adds unlabeled data in the first iteration. The larger the data set, the more features the model can learn. As the number of iterations increases, the model accuracy increases steadily, but when the fourth iteration is reached, the accuracy is only 0.04% higher than the third round, which is lower than 0.1%. The iteration is stopped and the final model accuracy is 96.55%.

Claims

1. A method for predicting the performance of organic photovoltaic materials based on machine learning, characterized in that: The steps include: Step 1: Train a model on a labeled dataset, called a teacher model; Step 2: Use the teacher model f obtained in step 1 to perform label prediction on the unlabeled data set, and predict the probability distribution given by the last layer of softmax, which is called soft label; Step 3: Train a new model called the student model, combine the labeled data set and the unlabeled data set with pseudo labels classified by the teacher model into a large data set, and perform training based on the large data set; Step 4: Re-label the unlabeled dataset using the learned student model, return to step 2, and iterate the loop to obtain the final machine learning prediction model.

2. The method for predicting the performance of organic photovoltaic materials based on machine learning according to claim 1, characterized in that: The step 1 is specifically as follows: The labeled data in the label dataset is represented as {(x1,y1),(x2,y2),...,(x n ,y n )}, the unlabeled data generated by image enhancement is represented as Train a convolutional neural network model on a labeled dataset, called the teacher model. The parameters of the teacher model are θ, and the teacher model uses cross entropy loss: Among them, n represents the total number of categories, y i represents the label, l(.) represents the cross entropy, f(.) represents the convolutional neural network, x i Represents data.

3. The method for predicting the performance of organic photovoltaic materials based on machine learning according to claim 2, characterized in that: The step 2 is specifically as follows: Step 2-1: The output of the neural network is: in, represents the neural network output, represents unlabeled data, and m represents the total number of categories; Step 2-2: Let the convolutional network be called N, and its i-th convolutional layer be expressed as: Y i =F i (X i ) (3) Where X i is the input data, Y i is the output of the convolution layer. If the teacher model has k convolution layers, the model is represented as: N=F k ⊙...⊙F2⊙F1(X1)=⊙ j＝1,...,k F j (X1) (4) Multiple convolutional layers with the same structure are called a stage, and the convolutional network N is expressed in stages as follows: in, <H i ,W i ,C i > represents the dimension of the input tensor of the i-th layer, and the subscript i represents the sequence number of the stage. represents the i-th stage, which consists of the convolutional layer F i Repeat L i Secondary composition; Step 2-3: The internal structure of the network model is implemented by multiple MBConv convolution blocks, which are composed of depthwise convolution and SENet. The MBConv convolution block uses the Swish activation function, and the formula is as follows: f(x)=x·sigmoid(βx) (9) SENet first performs a Squeeze operation to compress features along the spatial dimension, converting each two-dimensional feature channel into a real number, and the output dimension matches the number of input feature channels; secondly, an Excitation operation is performed to generate weights for each feature channel through the parameter w, where the parameter w is learned to explicitly model the correlation between feature channels; finally, a Reweight operation is performed to use the weights output by Excitation as the importance of each feature channel after feature selection, and then weighted to the previous features channel by channel through multiplication, completing the recalibration of the original features in the channel dimension; Step 2-4: Adjust model parameters based on EfficientNet-B0 to increase network depth, network width, and input image resolution.

4. The method for predicting the performance of organic photovoltaic materials based on machine learning according to claim 3, characterized in that: The step 3 is specifically as follows: Add noise to the student model and use cross entropy loss: For image datasets consisting of labeled data, divide them into training sets, validation sets, and test sets; For unlabeled data, the EfficientNet-B0 network is used to predict the label of each image data.

5. The method for predicting the performance of organic photovoltaic materials based on machine learning according to claim 4, characterized in that: Batch Normalization is used in the model, which is expressed as follows: Among them, μ and σ are the mean and variance of z respectively, ε is a number tending to 0, and the distribution of the node input value z is re-normalized to a normal distribution with a mean of 0 and a variance of 1; Batch Normalization performs a scale operation on the input values ​​that satisfy the standard distribution, as shown in formula (8): z′=α·z_norm+β (8) In the formula, α and β are two additional parameters for each neuron that can be obtained through learning.

6. The method for predicting the performance of organic photovoltaic materials based on machine learning according to claim 5, characterized in that: The EfficientNet-B0 consists of 16 MBConvs, two Convs, a global average pooling layer, and a fully connected layer.

Images