A single particle prediction effect method based on densely connected networks

Abstract

The present invention discloses a single particle prediction effect method based on a densely connected network, comprising the following steps: step 1, GEANT4 modeling: step 2, using the obtained LET data under different particle conditions as a data set; step 3, randomly dividing the total sample data set into a training set, a cross-validation set and a test set, and fixing the random seed to ensure that the experimental results are reproducible; step 4, building a prediction model based on a densely connected network; step 5, inputting the training set into the prediction model to optimize the prediction model; step 6, simultaneously inputting the cross-validation set into the prediction model during step 5, and adjusting the number of training set repetition training and the learning rate in back propagation in the optimization process by comparing its loss function curve with the training set loss function curve; step 7, evaluating its prediction accuracy and generalization ability. The present invention can accurately and quickly predict the linear energy transfer data of microelectronic devices, thereby analyzing the influence of single particle effects on it.

Description

Technical Field

[0001] The present invention belongs to the technical field of predicting single particle effects, and in particular relates to a single particle effect prediction method based on a densely connected network. Background Art

[0002] At present, most of the research on single particle effects is based on computer software simulation. Although this method can reduce experimental costs and shorten the design cycle to a certain extent, it lacks accurate simulation of particle interactions. Among the existing computer simulation tools for studying single particle effects of semiconductor devices, most of them focus on the analysis of the electrical mechanism of semiconductor devices, the simulation speed is relatively slow, and the analysis of radiation environment, semiconductor materials, nuclear reaction mechanism of incident particles and device materials is not perfect.

[0003] When choosing simulation software to study microelectronic devices, once the simulation results do not meet expectations, researchers need to rebuild the model, and model reconstruction and performance simulation will consume a lot of time and energy. In addition, if researchers do not have sufficient solid theoretical knowledge in the field of microelectronics, it will become very difficult to use software to build models and simulate devices.

[0004] Existing single-particle effect research is mostly based on computer software simulation, which is a practical measure for radiation research. While performing mathematical statistics and processing on experimental data, it can also supplement situations that cannot be reached under experimental conditions to a certain extent. However, it is difficult to use such software to model devices. Once the simulation results do not meet expectations, researchers need to rebuild the model; slow simulation speeds also lead to a waste of time and energy; and the analysis of radiation environment, semiconductor materials, incident particles, and nuclear reaction mechanisms of device materials is not perfect. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a single particle effect prediction method based on densely connected networks, which can demonstrate the combined capabilities of densely connected networks and the GEANT4 toolkit to accurately and quickly predict the linear energy transfer data of microelectronic devices, thereby analyzing the impact of single particle effects thereon.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A single particle prediction effect method based on densely connected networks comprises the following steps;

[0008] Step 1, GEANT4 modeling:

[0009] In the GEANT4 software toolkit, a rectangular parallelepiped sensitive body composed of silicon material is constructed and used as the target material for each particle to be incident. A world is established and a vacuum environment is used as a filler. A physical list is constructed and alpha particles and four heavy ions of different atomic masses (B, N, Ne, Ar) are added to the physical list for subsequent calls. A particle gun is set up to specify the particle type, incident direction, incident angle, and doping concentration and thickness of the sensitive body, and the script file is written in the form of defined variables. Energy deposition is counted and LET data is collected. The calculation expression of LET is as follows:

[0010] LET=1 / ρ×dE / dx

[0011] Where ρ is the density of silicon material, E is the energy generated by radiation, dE / dx is the basic orbital stopping power in the material, and the effective LET value after the total increase caused by energy deposition and energy loss in the sensitive space is expressed as linear charge deposition (LCD, Linear Charge Deposition), which is interpreted as the transfer charge or deposited charge per unit length. The conversion formula is as follows:

[0012] dN / dx=dP / dx=|dE / dx| / W=(ρ / W)·LET

[0013] N is the number of electron-hole pairs, W is the average ionization energy of the material, for silicon material (W si ≈3.6MeV, E=1.6×10 -19 C);

[0014] Step 2, taking the LET data of different particle types, incident directions, incident angles, and different sensitive body doping concentrations and thicknesses obtained in GEANT4 as a data set;

[0015] Step 3: randomly divide the total sample data set into training set, cross-validation set and test set, and fix the random seed to ensure that the experimental results can be reproduced;

[0016] Step 4: Build a prediction model based on the densely connected network;

[0017] Step 5: Input the training set into the prediction model, measure the gap between the predicted value and the true value through the loss function, and perform back propagation to update the relevant weights and bias parameters in the prediction model, thereby continuously optimizing the prediction model.

[0018] Step 6: In step 5, the cross-validation set is input into the prediction model at the same time, and the number of training set repetitions and the learning rate in back propagation in the optimization process are adjusted by comparing the loss function curve with the training set loss function curve, so as to improve the accuracy of the prediction model and reduce the adverse effects caused by overfitting of the training;

[0019] Step 7: Input the test set into the trained model to evaluate its prediction accuracy and generalization ability.

[0020] In step 1, the initial energy range of the particles is set to 10 -1 ~10 4 MeV, the energy range interval changes by a factor of 10; LET is generally used in a form normalized by material density, and the unit is usually MeV cm 2 / mg.

[0021] The step 2 data set has a total of 3240 sets of data, and the step 3 randomly divides the total sample data set (3240) into a training set, a cross-validation set and a test set according to a ratio of 6:2:2.

[0022] The prediction model in step 4 is based on a densely connected network, and the model includes an input expansion part and a feature extraction part;

[0023] The five particle conditions described in step 2 (particle type, incident direction, incident angle, sensitive body doping concentration, thickness) are used as the input of the network model. Since fewer input factors are not conducive to the training and feature extraction of the network model, two fully connected layers are used to expand the input of the network. The first fully connected layer consists of 64 neurons, and the second fully connected layer consists of 128 neurons. A batch normalization unit and a LeakyReLu activation function are added after each fully connected layer. During model training, before applying the activation function, the output of a layer is normalized first, and all batch data are forced to be under a unified data distribution, and then input to the next layer, so that the values ​​of the intermediate outputs of the entire neural network in each layer are more stable, so that the deep neural network is easier to converge and the risk of model overfitting is reduced;

[0024] The feature extraction part consists of an independent convolutional layer and 4 groups of densely connected blocks. A transition layer is added after each densely connected block to control the number of channels and data volume of the training data, thereby reducing the amount of training operations while ensuring the training accuracy to improve the inference speed of the model. Finally, the features extracted from the densely connected part are reduced in dimension and input into a fully connected layer composed of 264 neurons to obtain the final network prediction value, which is the LET curve composed of 264 points.

[0025] The feature extraction part first extracts the one-dimensional vector after input expansion into a multi-dimensional feature through a convolution layer with an input channel of 1, an output channel of 64, a kernel of 7*7, a stride of 2, and a padding of 3, and then inputs 4 groups of densely connected blocks, each of which is composed of 4 convolution layers with a kernel of 3*3, a stride of 1, and a padding of 1. The input channel of the first densely connected block is 32, and the output channel is 64. After the dense block processing, 128 channels are added. Each transition layer consists of a layer with a kernel of 1*1, a stride of 1, It consists of a convolutional layer with a padding of 1 and a one-dimensional average pooling layer with a kernel size of 2*2 and a stride of 2. The number of output channels and columns of each group of dense blocks is halved, thereby effectively controlling the size of the model. Except for the transition layer, a batch normalization unit and a LeakyReLu activation function are added after each convolutional layer. Finally, the multi-dimensional vector extracted from the features is compressed into a one-dimensional vector, and the feature vector is converted into a predicted value through a fully connected layer composed of 496 neurons. A batch normalization unit and a LeakyReLu activation function are added and output through a fully connected layer composed of 264 neurons.

[0026] In summary, GEANT4 was selected to conduct a preliminary analysis of the radiation environment, semiconductor materials, and the reaction mechanism between incident particles and device materials, simulate the interaction between particles and silicon materials, and obtain the corresponding LET data.

[0027] The five particle conditions in step 2 are used as the input of the dense connection network, and the corresponding 264 LET data are used as the output of the dense connection network. The proposed dense connection network is trained to achieve the effect of fast and efficient prediction of LET data.

[0028] Beneficial effects of the present invention:

[0029] 1. The present invention uses the Monte Carlo toolkit-GEANT4 to construct a rectangular parallelepiped sensitive body composed of silicon material, and obtains LET data under different particle types, incident directions, incident angles, and sensitive body doping concentrations and thicknesses as a data set. In this process, the advantage of GEANT4 is that it can simulate the interaction between particles and between particles and target materials, and calculate nuclear physics processes such as energy absorption and collision. It uses C++ class inheritance to facilitate modeling by researchers, laying the foundation for subsequent research on the single particle effect of devices.

[0030] 2. The present invention selects a densely connected network to predict the LET data when particles enter the target material (silicon) under different conditions. The densely connected network has the following convincing advantages: alleviating the vanishing gradient, strengthening feature propagation, encouraging feature reuse, and requiring fewer parameters than the traditional convolutional network. Therefore, a densely connected network is established to train and verify the data obtained by GEANT4, which greatly reduces the simulation time and improves the simulation efficiency.

[0031] 3. In GEANT4 modeling, the simulation of particle random processes is realized by overloading the GEANT4 base class and abstract class customization, and the linear energy transfer (LET) is statistically analyzed by bombarding silicon materials with multiple particles. GEANT4 can be used to complete the simulation of particle random processes and build specific applications. It is difficult to add new physical models based on the existing simulation software in the field of microelectronics, and the GEANT4 implementation is written based on the C++ object-oriented concept, which greatly reduces the size, complexity and interdependence of the code. In addition, you can directly refer to the header file to understand the definition of the library and the functions it provides.

[0032] 4. In the process of obtaining the data set in step 2, based on the Monte Carlo method, the probabilistic phenomenon is taken as the research object, and repeated random sampling is relied on to obtain numerical results. Therefore, 1000 groups of particles are set to be incident on the silicon material at each energy value, and the average value of the final linear energy transfer is taken as a set of LET data in step 2, which will greatly improve the accuracy of the experiment, which is reflected in the script through " / run / beamOn 1000". In addition, the single-particle simulation can use the multi-threading capability provided by GEANT4 to accelerate the simulation and greatly shorten the simulation time.

[0033] 5. The densely connected network of the present application realizes the mapping of the input x to its expanded form by splicing the input and output of each layer in dimension, thereby ensuring that there is always a large gradient value, effectively preventing the disappearance of the gradient, increasing the network depth, and thus improving the network performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the particle incident target material of the present invention.

[0035] Figure 2 It is a schematic diagram of the dense network structure of the present invention.

[0036] Figure 3 is a comparison of four sets of predicted linear energy transfer curves randomly selected from 648 test sets and the corresponding simulation curves. As shown in the figure, the curve has a high degree of fit and the prediction effect is good. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0038] like Figure 1 As shown, step 1, GEANT4 modeling:

[0039] In the GEANT4 software toolkit, a rectangular parallelepiped sensitive body made of silicon material was constructed and used as the target material for each particle to be incident on. A world was created and a vacuum environment was used as a filler. A physical list was constructed and α particles and four heavy ions of different atomic masses (B, N, Ne, Ar) were added to the physical list for subsequent calls. The initial energy range of the particles was set to 10 -1 ~10 4 MeV, the energy range interval changes by 10 times; set the particle gun, specify the particle type, incident direction, incident angle, and sensitive body doping concentration and thickness, and write the above input into the script file in the form of defining variables; count the energy deposition and collect LET data. In one embodiment, set the Ne particle gun, and specify the atomic number Z (10) and relative atomic mass A (20) corresponding to Ne through the G4IonTable::GetIonTable()->GetIon() method; the Ne particle gun is incident at 60° to the z-axis direction, G4ThreeVector(0.,0.,1.), ParticleDegree=60.; the sensitive body doping concentration is 10 20 / cm 3 , thickness is 50nm, G4double Thickness=50.*nm. (This step is explained with a specific example) LET is generally used in a form normalized by material density, and the unit is usually MeV·cm 2 / mg. The calculation expression of LET is as follows:

[0040] LET=1 / ρ×dE / dx

[0041] Where ρ is the density of the material, E is the energy generated by radiation, and dE / dx is the basic orbital stopping power in the material. The concept of linear charge deposition (LCD) is introduced to represent the effective LET value after the total amount increase caused by energy deposition and the energy loss in the sensitive space. It is interpreted as the transfer charge or deposited charge per unit length. The conversion formula is as follows:

[0042] dN / dx=dP / dx=|dE / dx| / W=(ρ / W)·LET

[0043] N is the number of electron-hole pairs, W is the average ionization energy of the material, for silicon material (W si ≈3.6MeV, E=1.6×10 -19 C).

[0044] Step 2: The LET data under different particle conditions obtained in GEANT4, a total of 3240 sets of data, are used as the data set. The specific selected parameters are shown in Table 1.

[0045] Step 3: randomly divide the total sample data set (3240) into training set, cross-validation set and test set in a ratio of 6:2:2, and fix the random seed to ensure that the experimental results are reproducible.

[0046] Step 4: Build a prediction model based on a densely connected network. The network structure is shown in the figure below: Figure 2 shown.

[0047] Step 5: Input the training set into the prediction model, measure the gap between the predicted value and the true value through the loss function, and perform back propagation to update the relevant weights and bias parameters in the prediction model, thereby continuously optimizing the prediction model.

[0048] In one embodiment, a training set containing 1944 samples is randomly divided into batches of 32 samples each, and each batch is input into the prediction model to obtain the prediction value. The Huber Loss loss function is used, which fully combines the advantages of the MSE and MAE loss functions, is easy to perform gradient descent, and is not easily affected by outliers. For the obtained loss function value, the Adam gradient descent optimization algorithm is used to update the weight of each layer of the network, and the dynamic learning rate is adjusted to achieve better training results. The Huber Loss loss function is represented as:

[0049]

[0050] Step 6: During step 5, the cross-validation set is input into the prediction model at the same time. By comparing its loss function curve with the loss function curve of the training set, the number of repeated training of the training set in the optimization process and the learning rate in the back propagation are adjusted to improve the accuracy of the prediction model while reducing the adverse effects caused by overfitting of the training.

[0051] Step 7: Input the test set into the trained model to evaluate its prediction accuracy and generalization ability.

[0052] In one embodiment, the particle types in the sample set are α, B, N, Ne, Ar, the particle incident directions are x, y, z, the particle incident angles are 15°, 30°, 45°, 60°, 75°, 90°, and the sensitive body doping concentrations are 0, 10 14 / cm 3 , 10 17 / cm 3 , 10 20 / cm 3 , 10 21 / cm3 , 10 22 / cm 3 The thickness of the sensitive body is 10nm, 20nm, 30nm, 40nm, 50nm, and 60nm.

[0053] In one embodiment, the densely connected network includes an input expansion part and a feature extraction part. Two fully connected layers are used to expand the input. An independent convolutional layer and four groups of densely connected blocks are used in the feature extraction part. A transition layer is added after each layer of densely connected blocks to control the number of channels and data volume of the training data. Finally, the features extracted by the densely connected part are reduced in dimension and input into the fully connected layer to output the predicted value.

[0054] In the input expansion part, the first fully connected layer is composed of 64 neurons, and the second fully connected layer is composed of 128 neurons. A batch normalization unit and a LeakyReLu activation function are added after each fully connected layer. In the feature extraction part, the one-dimensional vector after input expansion is first extracted as a multi-dimensional feature through a convolutional layer with an input channel of 1, an output channel of 64, a kernel of 7*7, a stride of 2, and a padding of 3, and then 4 groups of densely connected blocks are input. Each group of densely connected blocks is composed of 4 convolutional layers with a kernel of 3*3, a stride of 1, and a padding of 1. The input channel of the first densely connected block is 32, and the output channel is 64. After dense block processing After that, 128 channels are added. Each transition layer consists of a convolutional layer with a kernel of 1*1, a stride of 1, and a padding of 1, and a one-dimensional average pooling layer with a kernel of 2*2 and a stride of 2. The number of output channels and columns of each group of dense blocks is halved, thereby effectively controlling the size of the model. Except for the transition layer, a batch normalization unit and a LeakyReLu activation function are added after each convolutional layer. Finally, the multi-dimensional vector extracted from the features is compressed into a one-dimensional vector, and the feature vector is converted into a predicted value through a fully connected layer composed of 496 neurons. A batch normalization unit and a LeakyReLu activation function are added and output through a fully connected layer of 264 neurons.

[0055] In one embodiment, the output of the network is 264 points of LET.

[0056] This prediction model is based on a densely connected network and uses two fully connected layers to expand the input. In the feature extraction part, an independent convolutional layer and four groups of densely connected blocks are used. A transition layer is added after each densely connected block to control the number of channels and data volume of the training data. While ensuring the training accuracy, the amount of training calculations is reduced to improve the reasoning speed of the model. Finally, the features extracted from the densely connected part are reduced in dimension and input into the fully connected layer to output the predicted value.

[0057] Among them, in the input expansion part, the first fully connected layer consists of 64 neurons, and the second fully connected layer consists of 128 neurons. A batch normalization unit and a LeakyReLu activation function are added after each fully connected layer. In the feature extraction part, the one-dimensional vector after input expansion is first extracted as a multi-dimensional feature through a convolutional layer with an input channel of 1, an output channel of 64, a kernel of 7*7, a stride of 2, and a padding of 3. Then 4 groups of densely connected blocks are input. Each group of densely connected blocks consists of 4 convolutional layers with a kernel of 3*3, a stride of 1, and a padding of 1. The input channel of the first densely connected block is 32, and the output channel is 64. After the dense block processing, 128 channels are added. Each transition layer consists of a convolutional layer with a kernel of 1*1, a stride of 1, and a padding of 1, and a one-dimensional average pooling layer with a kernel of 2*2 and a stride of 2. The number of output channels and columns of each dense block is halved, thereby effectively controlling the model size. A batch normalization unit and a LeakyReLu activation function are added after each convolutional layer except the transition layer. Finally, the multi-dimensional vector extracted from the features is compressed into a one-dimensional vector, and the feature vector is converted into a predicted value through a fully connected layer composed of 496 neurons. A batch normalization unit and a LeakyReLu activation function are added and output through a fully connected layer of 264 neurons.

[0058] The core of GEANT4 is a set of rich physical models, which make up for the shortcomings of semiconductor device simulation tools. It can not only realize the interaction between particles and matter in a wide energy range, accurately simulate the transportation process of particles, complete the specific calculation of physical processes such as energy absorption and collision, but also track and effectively output information such as particle energy, position, secondary particles and intermediate results.

[0059] GEANT4 is an open source computing tool library based on C++. Its code adopts object-oriented method, and there is strong independence between modules, so researchers do not need to have too much understanding of the overall structure. In addition, GEANT4 operation is relatively intuitive, and only programming basics are needed to realize the establishment of relevant models, without too much microelectronics knowledge.

[0060] When using GEANT4 to predict single-particle effects of microelectronic devices, simulation analysis will take up a large part of the time. Using deep learning algorithms to predict devices will greatly improve the accuracy and efficiency of the experiment. As a newer convolutional neural network architecture, densely connected networks have the advantages of more compact models and less prone to overfitting. Combining the GEANT4 toolkit with densely connected networks, the linear energy transfer (LET) data of microelectronic devices can be accurately and quickly predicted, thereby analyzing the impact of single-particle effects on them. The present invention solves the problems of modeling difficulties, long time consumption and low accuracy in the traditional research process of predicting single-particle effects of microelectronic devices.

[0061] Compared with traditional convolutional neural networks such as ResNet, VGGNet, etc., densely connected networks can effectively improve network performance.

[0062] Table 1 is the parameter selection of the dataset

[0063]

[0064] Table 1

[0065] Figure 3 is a comparison chart of four groups of test results randomly selected from the test set and the simulation values. Figure 3a -d are the incident results of Ar, Ne, N, and α particles, respectively. Within the energy range, LET increases from a small value to an extreme value, and LET gradually decreases with the increase of energy. As the atomic weight of heavy ions increases, the LET extreme value continues to increase, and the energy corresponding to the LET extreme value also increases step by step. At the same time, the LET in the low energy range is 0 in some cases, which means that some heavy ions have a certain probability of escaping directly through the device without depositing energy.

Claims

1. A single-particle prediction effect method based on densely connected networks, It is characterized in that The steps include: Step 1, GEANT4 modeling: In the GEANT4 software toolkit, a rectangular parallelepiped sensitive body composed of silicon material is constructed and used as the target material for each particle to be incident. A world is established and a vacuum environment is used as a filler. A physical list is constructed and alpha particles and four heavy ions of B, N, Ne, and Ar with different atomic masses are added to the physical list. A particle gun is set up to specify the particle type, incident direction, incident angle, and doping concentration and thickness of the sensitive body, and the script file is written in the form of defined variables. Energy deposition is counted and LET data is collected. The calculation expression of LET is as follows: LET=1 / ρ×dE / dx Where ρ is the density of silicon material, E is the energy generated by radiation, dE / dx is the basic orbital stopping power in the material, and the effective LET value after the total increase caused by energy deposition and energy loss in the sensitive space is expressed as linear charge deposition, which is interpreted as the transferred charge or deposited charge per unit length. The conversion formula is as follows: dN / dx=dP / dx=|dE / dx| / W=(ρ / W)·LET N is the number of electron-hole pairs, W is the average ionization energy of the material, and for silicon material W si ≈3.6MeV, E=1.6×10 -19 C; Step 2, taking the LET data of different particle types, incident directions, incident angles, and different sensitive body doping concentrations and thicknesses obtained in GEANT4 as a data set; Step 3: randomly divide the total sample data set into training set, cross-validation set and test set, and fix the random seed to ensure that the experimental results can be reproduced; Step 4: Build a prediction model based on the densely connected network; Step 5: Input the training set into the prediction model, measure the gap between the predicted value and the true value through the loss function, and perform back propagation to update the relevant weights and bias parameters in the prediction model, thereby continuously optimizing the prediction model. Step 6: In step 5, the cross-validation set is input into the prediction model at the same time, and the number of training set repetitions and the learning rate in back propagation in the optimization process are adjusted by comparing the loss function curve with the loss function curve of the training set, so as to improve the accuracy of the prediction model and weaken the adverse effects caused by fitting in the training process; Step 7: Input the test set into the trained model to evaluate its prediction accuracy and generalization ability.

2. A single particle prediction effect method based on densely connected networks according to claim 1, It is characterized in that In step 1, the initial energy range of the particles is set to 10 -1 ~10 4 MeV, the energy range interval changes by a factor of 10; LET uses the form normalized by the material density, the unit is usually MeV cm 2 / mg.

3. A single particle prediction effect method based on densely connected networks according to claim 1, It is characterized in that The prediction model in step 4 is based on a densely connected network, and the model includes an input expansion part and a feature extraction part; The five particle types, incident direction, incident angle, sensitive body doping concentration, and thickness described in step 2 are used as the input of the network model. Two fully connected layers are used to expand the input of the network. The first fully connected layer consists of 64 neurons, and the second fully connected layer consists of 128 neurons. A batch normalization unit and a LeakyReLu activation function are added after each fully connected layer. When training the model, before applying the activation function, the output of a layer is normalized first, and all batch data are forced to be under a unified data distribution, and then input into the next layer, so that the values ​​of the intermediate outputs of the entire neural network in each layer are more stable, so that the deep neural network is easier to converge and the risk of model overfitting is reduced; The feature extraction part consists of an independent convolutional layer and 4 groups of densely connected blocks. A transition layer is added after each densely connected block to control the number of channels and data volume of the training data, thereby reducing the amount of training operations while ensuring the training accuracy to improve the inference speed of the model. Finally, the features extracted from the densely connected part are reduced in dimension and input into a fully connected layer consisting of 264 neurons to obtain the final network prediction value.

4. A single particle prediction effect method based on densely connected networks according to claim 3, It is characterized in that The feature extraction part first extracts the one-dimensional vector after input expansion into a multi-dimensional feature through a convolution layer with an input channel of 1, an output channel of 64, a kernel of 7*7, a stride of 2, and a padding of 3, and then inputs 4 groups of densely connected blocks, each of which is composed of 4 convolution layers with a kernel of 3*3, a stride of 1, and a padding of 1. The input channel of the first densely connected block is 32, and the output channel is 64. After the dense block processing, 128 channels are added. Each transition layer consists of a layer with a kernel of 1*1, a stride of 1, It consists of a convolutional layer with a padding of 1 and a one-dimensional average pooling layer with a kernel size of 2*2 and a stride of 2. The number of output channels and columns of each group of dense blocks is halved, thereby effectively controlling the size of the model. Except for the transition layer, a batch normalization unit and a LeakyReLu activation function are added after each convolutional layer. Finally, the multi-dimensional vector extracted from the features is compressed into a one-dimensional vector, and the feature vector is converted into a predicted value through a fully connected layer composed of 496 neurons. A batch normalization unit and a LeakyReLu activation function are added and output through a fully connected layer composed of 264 neurons.

5. According to the single particle prediction effect method based on densely connected network in claim 1, the data set in step 2 has a total of 3240 sets of data, and the total sample data set in step 3 is randomly divided into a training set, a cross-validation set and a test set in a ratio of 6:2:2.

Images