Silicon-rich ONO structure

Abstract

The invention discloses a silicon-rich ONO structure which comprises a charge tunneling layer, a charge trapping layer and a charge blocking layer. The charge trapping layer comprises a first charge trapping layer, a second charge trapping layer and a third charge trapping layer; wherein the second charge trapping layer is silicon-rich silicon nitride; the deposition thickness of the first charge trapping layer, the second charge trapping layer and the third charge trapping layer is predicted according to the convolutional neural network, and then optimal deposition process parameters are reversely deduced; according to the invention, accurate control of the thickness of the silicon nitride layer in the non-volatile memory device can be realized, so that the performance and the reliability of the device are improved. Meanwhile, the invention aims to overcome the problems of high cost, low efficiency, complicated parameter adjustment and the like in the traditional manufacturing process, and provides a more advanced, efficient and reliable process optimization scheme for the semiconductor manufacturing industry, so that the production efficiency is improved, the manufacturing process cost is reduced, and further development and application of the semiconductor manufacturing technology are promoted.

Description

Technical Field

[0001] The invention belongs to the technical field of semiconductor manufacturing, and in particular relates to a silicon-rich ONO structure. Background Art

[0002] Nowadays, flash devices usually use non-volatile memory technology, and the ONO structure is a common stacking structure used to store charge in flash devices. It is where data is stored in flash devices, and the role of the oxide layer is to ensure the isolation of charges, which enables flash devices to achieve reliable charge storage or erase and write operations.

[0003] In the modern semiconductor industry, technologies based on convolutional neural networks (CNNs) have been widely used in image processing, pattern recognition, and data prediction. The ONO structure, or oxide-nitride-oxide structure, is a common stacking structure in non-volatile memory. Among them, the silicon nitride layer, as a nitride layer, plays a key role in storing charges in Flash devices.

[0004] At present, the thickness control of the composite silicon nitride layer of the ONO structure is crucial for the semiconductor manufacturing process. Accurate thickness control directly affects the performance and reliability of the device. Traditional methods usually rely on complex process parameter adjustments and offline measurements, which brings problems such as low production efficiency, high cost and unstable process.

[0005] Traditional methods rely on offline measurement and complex process adjustments, resulting in high production costs and low efficiency in improving process levels. The present invention introduces advanced technology based on convolutional neural networks, which improves the prediction accuracy of the thickness of the composite silicon nitride layer and reduces production costs through real-time data analysis and learning.

[0006] In the traditional method, the thickness of the composite silicon nitride layer needs to be adjusted by adjusting multiple parameters, which increases the complexity of the process. The present invention utilizes the automatic feature learning capability of the convolutional neural network, reduces the manual adjustment of multiple parameters, and improves the simplicity and consistency of the process. Summary of the invention

[0007] The present invention aims to overcome the problems of high cost, low efficiency and complex parameter adjustment in traditional processes, and provide a more advanced, efficient and reliable process optimization solution for the semiconductor manufacturing industry. A silicon-rich ONO structure is proposed to improve device performance and reliability by achieving precise control of the thickness of the silicon nitride layer in non-volatile memory devices.

[0008] The present invention provides a silicon-rich ONO structure, comprising a charge tunneling layer, a charge trapping layer, and a charge blocking layer arranged in sequence from bottom to top; the charge trapping layer comprises a first charge trapping layer, a second charge trapping layer, and a third charge trapping layer arranged in sequence from bottom to top; The material of the first charge trapping layer is silicon nitride Si3N4, the material of the second charge trapping layer is silicon-rich silicon nitride SiN, and the material of the third charge trapping layer is silicon nitride Si3N4; the thickness ratio of the first charge trapping layer, the second charge trapping layer, and the third charge trapping layer is (20-30): (40-50): (40-50); The materials of the charge tunneling layer and the charge blocking layer may include but are not limited to silicon dioxide, silicon nitride and silicon oxynitride.

[0009] Preferably, the thickness ratio of the charge tunneling layer, the charge trapping layer and the charge blocking layer is (20-40): (100-130): (50-60).

[0010] Preferably, the charge tunneling layer has a thickness of 20 nm-40 nm, the first charge trapping layer has a thickness of 20 nm-30 nm, the second charge trapping layer has a thickness of 40 nm-50 nm, the third charge trapping layer has a thickness of 40 nm-50 nm, and the charge blocking layer has a thickness of 50 nm-60 nm.

[0011] Preferably, the charge blocking layer is formed by depositing silicon oxide by high temperature oxidation (HTO) or low pressure chemical vapor deposition (LPCVD), and the charge tunneling layer is formed by depositing silicon oxide by low temperature transient vapor deposition (ISSG) or plasma enhanced chemical vapor deposition (PECVD).

[0012] Preferably, the charge trapping layer is deposited on the charge tunneling layer in sequence by using LPCVD method or atomic layer deposition (ALD) method to form a first charge trapping layer, a second charge trapping layer and a third charge trapping layer; and finally heat treatment is performed.

[0013] Preferably, in the deposition process of the first charge trapping layer, the gas flow ratio of NH3 and SiH2Cl2 is (900-1000): (100-200), in sccm, more preferably 900 sccm: 100 sccm, in the deposition process of the second charge trapping layer, the gas flow ratio of NH3 and SiH2Cl2 is (50-100): (100-200), in sccm, more preferably 50 sccm: 100 sccm, and in the deposition process of the third charge trapping layer, the gas flow ratio of NH3 and SiH2Cl2 is (900-1000): (100-200), in sccm, more preferably 900 sccm: 100 sccm.

[0014] Preferably, the first charge trapping layer, the second charge trapping layer, and the third charge trapping layer are deposited by predicting thicknesses according to a convolutional neural network, thereby inversely inferring optimal deposition process parameters of the first charge trapping layer, the second charge trapping layer, and the third charge trapping layer; Preferably, the convolutional neural network adopts a ResNet18 model based on a residual neural network.

[0015] Preferably, the deposition process parameters of the first charge trapping layer, the second charge trapping layer and the third charge trapping layer include deposition temperature, deposition time, pressure, NH3 gas flow rate, SiH2Cl2 gas flow rate, Wafer ID of the silicon wafer in the furnace tube machine, and point coordinates of the silicon wafer measurement points (i.e., (x, y) point coordinates).

[0016] The beneficial effects of the present invention are: The present invention proposes an ONO structure composed of a charge tunneling layer, a multi-layer charge trapping layer and a charge blocking layer, wherein the second charge trapping layer is a silicon-rich silicon nitride layer, which optimizes the charge storage capacity. In addition, the present invention also uses a convolutional neural network to predict and control the thickness of each of the multi-layer charge trapping layers, predicts the deposition thickness through the convolutional neural network, reversely infers and determines the optimal deposition process parameters, simplifies the production process, improves production efficiency, and of course improves the accuracy of thickness control, reducing the need for manual adjustment of multiple process parameters. Therefore, the present invention realizes the precise control of the thickness of the composite silicon nitride layer in the ONO structure, significantly improves the performance and reliability of the Flash storage device, while reducing the production cost and improving the production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The silicon-rich ONO structure provided in this embodiment; Figure 2 This is the deposition process of the ONO structure in this embodiment; Figure 3This is the training process of the ResNet18 convolutional neural network model of the present invention; Figure 4 This is a graph showing the relationship between the loss and the number of iterations epoch in ResNet18 training in this example. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0019] The present invention provides a silicon-rich ONO structure. In this embodiment, the ONO structure is as follows: Figure 1 The invention comprises a charge tunneling layer, a charge trapping layer and a charge blocking layer; wherein the charge trapping layer is composed of a first charge trapping layer, a second charge trapping layer and a third charge trapping layer, namely silicon nitride-silicon-rich silicon nitride-silicon nitride, and the charge tunneling layer and the charge blocking layer are composed of silicon oxide. The thickness of the charge tunneling layer is about 20 nm-40 nm, which is conducive to the charge passing through the charge tunneling layer to reach the charge trapping layer.

[0020] In this example, the charge blocking layer uses a high temperature oxidation (HTO) method to deposit silicon oxide, and the charge tunneling layer uses a low temperature transient vapor deposition (ISSG) method to deposit silicon oxide. The charge tunneling layer deposited by ISSG can achieve a higher number of cycles and improve the durability of the device.

[0021] In this example, the silicon nitride is prepared from NH3 and SiH2Cl2, the structural formula of the silicon nitride layer is Si3N4, and the structural formula of the silicon-rich silicon nitride is SiN. On the silicon oxide layer, the silicon nitride layer, the silicon-rich silicon nitride layer and the silicon nitride layer are sequentially deposited by the LPCVD method to form a charge trapping layer. Figure 2 The specific process is:

[0022] First, the gas flow ratio of NH3 and SiH2Cl2 is set to 900-1000 sccm: 100-200 sccm, and a silicon nitride layer is deposited under specific deposition process conditions with a thickness of 20 nm to 30 nm to obtain a first charge trapping layer.

[0023] Then, the flow ratio of NH3 and SiH2Cl2 gases is adjusted to 50 sccm:100 sccm, and a silicon-rich silicon nitride layer with a thickness of 40 nm to 50 nm is deposited under specific deposition process conditions to obtain a second charge trapping layer.

[0024] Finally, the gas flow ratio of NH3 and SiH2Cl2 is restored to 900-1000 sccm: 100-200 sccm, and the silicon nitride layer is continuously deposited under specific deposition process conditions with a thickness of 40 nm to 50 nm to obtain the third charge trapping layer.

[0025] The ratio of nitrogen to silicon elements is achieved by changing the gas flow rate.

[0026] Finally, in order to improve the performance of the overall structure, the prepared ONO structure is heat treated to reduce defects in the film, improve the quality between layers, and enhance the performance of the charge trapping layer.

[0027] In this example, the silicon-rich ONO structure and the non-uniform composition of these charge trapping layers are superior to the traditional composition of SONOS charge trapping layers in many aspects. The additional silicon atoms can form more charge traps, increase the tunneling rate of charges, help the efficiency of charge capture and release, and improve the erase efficiency.

[0028] In the above steps, the first charge trapping layer and the second charge trapping layer deposited on the charge tracking layer are both prepared by LPCVD, which has the advantages of being able to prepare high-density thin films, having good chemical stability, and being able to be deposited over a large area.

[0029] The respective deposition thicknesses of the first charge trapping layer, the second charge trapping layer, and the third charge trapping layer are reversed by using a convolutional neural network to infer the respective optimal deposition process parameters of the first charge trapping layer, the second charge trapping layer, and the third charge trapping layer; The convolutional neural network adopts the ResNet18 model based on the residual neural network, such as Figure 3 Its construction and training process are as follows: Step 1: Model construction and physical model introduction Build a ResNet18 model based on residual neural network: Based on deep learning, a ResNet18 convolutional neural network model is established. Deposition process parameters such as NH3 gas flow rate, SiH2Cl2 gas flow rate, deposition temperature and other deposition process parameters are used as input to predict the thickness of the ONO composite silicon nitride layer.

[0030] In the ResNet18 model, the physical model of chemical vapor deposition is combined with factors such as reactant concentration, temperature, and pressure to more accurately capture the key influencing factors in the CVD process. The model improves the prediction accuracy of the ONO layer thickness by learning the complex associations in the CVD process.

[0031] In step 1, the process parameters include temperature, time, pressure, NH3 gas flow, SiH2Cl2 gas flow, Wafer ID, (x, y) point coordinates, and a machine learning mapping relationship model is established between the process parameters of ONO thin film chemical vapor deposition and the film thickness: In the formula, x represents the process parameters at different temperatures and pressures; represents a nonlinear mapping function that maps x to a high-dimensional feature space; Represents the corresponding physical properties and thickness; w represents the weight vector of the feature space; b represents the deviation variable.

[0032] The above-mentioned Wafer ID refers to the position of the silicon wafer in the furnace tube machine, which is divided into top, center, and bottom in the machine; the point coordinates are during the measurement process, for example, a wafer is measured at 49 points, and the ID coordinates are the positions of the 49 points.

[0033] Step 2: Data collection and dataset construction Process parameter collection: Through the actual conditions and process steps set for the 55nm production line, ONO composite silicon nitride process parameters such as NH3 gas flow rate and SiH2Cl2 gas flow rate are collected.

[0034] Construct training data set: obtain experimental data through experiments, including physical parameters and corresponding ONO layer thickness. Make sure that the data set covers most experimental conditions to improve the prediction ability of the model. In the process of constructing the data set, through the association between lotID and process parameters, the data is automatically captured by the script and added to the existing data set.

[0035] The existing process parameters and the corresponding silicon nitride thickness data sets are processed, normalized, and divided into training and test sets, roughly in the ratio of 80% and 20%.

[0036] Step 3-6: Model training and performance evaluation Step 3: Build a ResNet18 model based on residual neural network, set appropriate number of layers and parameters, introduce convolutional layer, pooling layer, fully connected layer and other structures to better extract features.

[0037] Construct a convolutional neural network and build a Resnet18 convolutional neural network based on a residual neural network. The first layer of the network structure is a 7x7 convolutional layer. It uses 64 7x7 convolution kernels to perform convolution operations on the input data, and adds a batch normalization layer and ReLU activation function. The role of this layer is to perform preliminary feature extraction and downsampling, reducing the input data from its original size to size. Then there are 4 residual blocks, each of which contains two convolutional layers and a cross-layer connection. Among them, the first convolutional layer uses 64 3x3 convolutional kernels, followed by a batch normalization layer and ReLU activation function; the second convolutional layer uses 64 3x3 convolutional kernels, followed by a batch normalization layer. If the number of channels in the middle of the residual block is not equal to the number of channels of the input and output, a 1x1 convolutional layer will be used to complete the dimension transformation. Through cross-layer connections, residual blocks can make the model easier to train and converge. Next are 4 similar residual blocks, each of which also contains two convolutional layers and a cross-layer connection, but the number of convolutional kernels used increases successively. The first residual block uses 128 convolutional kernels, the second uses 256 convolutional kernels, and the third uses 512 convolutional kernels. Through this gradual deepening and widening method, the model can learn more complex and abstract features, thereby improving classification accuracy. Finally, there is a global average pooling layer that averages the feature maps of each channel and compresses them into a feature vector. Then, a fully connected layer maps the feature vector to the probability distribution of the output category. In addition, in order to prevent overfitting, some regularization layers (such as Dropout) can be added to randomly discard some neurons.

[0038] Step 4: Use the training set to train the ResNet18 model based on the residual neural network and optimize the model parameters by minimizing the loss function. Consider using pre-trained weights or transfer learning to speed up the training process. During the optimization process, in order to speed up the training process, change the number of residual blocks or the number of neurons in ResNet to simplify the model while ensuring accuracy.

[0039] Step 5: Connect the ResNet18 model based on the residual neural network to the CVD physical model in series. This helps to comprehensively consider the advantages of deep learning and physical models and improve the overall performance of the model.

[0040] Step 6: Use the test set to evaluate the model performance, including the model's accuracy, generalization ability, and robustness. By adjusting the hyperparameters and model structure, ensure that the model performs well under different experimental conditions. MSELoss (mean square error loss) is a commonly used loss function that evaluates the performance of the model by calculating the square error between the predicted value and the true value. The calculation formula is as follows:

[0041] in and Represent the estimated value and true value of the model respectively, and N represents the total number of samples. After completing the above work, the training data is also normalized to the maximum and minimum values, and the data is scaled to the specified range. This can enhance the stability of the values, avoid accuracy problems caused by too large or too small values, and reduce the dimensional differences of different features. After normalization, the loss obtained from training tends to be stable, and the test loss can reach 0.4157, such as Figure 4 shown.

[0042] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A silicon-rich ONO structure, characterized in that It includes a charge tunneling layer, a charge trapping layer, and a charge blocking layer arranged in sequence from bottom to top; the charge trapping layer includes a first charge trapping layer, a second charge trapping layer, and a third charge trapping layer arranged in sequence from bottom to top; Among them, the material of the first charge trapping layer is silicon nitride Si3N4, the material of the second charge trapping layer is silicon-rich silicon nitride SiN, and the material of the third charge trapping layer is silicon nitride Si3N4; the thickness ratio of the first charge trapping layer, the second charge trapping layer, and the third charge trapping layer is (20-30): (40-50): (40-50).

2. A silicon-rich ONO structure according to claim 1, characterized in that: The charge tunneling layer and the charge blocking layer are made of silicon dioxide, silicon nitride or silicon oxynitride.

3. A silicon-rich ONO structure according to claim 1, characterized in that: The thickness ratio of the charge tunneling layer, the charge trapping layer and the charge blocking layer is (20-40): (100-130): (50-60).

4. A silicon-rich ONO structure according to claim 1, characterized in that: The thickness of the charge tunneling layer is 20 nm-40 nm, the thickness of the first charge trapping layer is 20 nm-30 nm, the thickness of the second charge trapping layer is 40 nm-50 nm, the thickness of the third charge trapping layer is 40 nm-50 nm, and the thickness of the charge blocking layer is 50 nm-60 nm.

5. A silicon-rich ONO structure according to claim 1, characterized in that: The charge blocking layer is deposited with silicon oxide by a high temperature oxidation method or a low pressure chemical vapor deposition method, and the charge tunneling layer is deposited with silicon oxide by a low temperature instantaneous steam deposition method or a plasma enhanced chemical vapor deposition method.

6. A silicon-rich ONO structure according to claim 1, characterized in that: The charge trapping layer is deposited in sequence on the charge tunneling layer by using a low pressure chemical vapor deposition method or an atomic layer deposition method; and finally a heat treatment is performed.

7. A silicon-rich ONO structure according to claim 1, characterized in that: In the deposition process of the first charge trapping layer, the gas flow ratio of NH3 and SiH2Cl2 is (900-1000): (100-200), in the deposition process of the second charge trapping layer, the gas flow ratio of NH3 and SiH2Cl2 is (50-100): (100-200), and in the deposition process of the third charge trapping layer, the gas flow ratio of NH3 and SiH2Cl2 is (900-1000): (100-200).

8. The silicon-rich ONO structure according to claim 1, characterized in that: The deposition thickness of the first charge trapping layer, the second charge trapping layer and the third charge trapping layer is predicted according to a convolutional neural network, and then the optimal deposition process parameters of the first charge trapping layer, the second charge trapping layer and the third charge trapping layer are reversed.

9. A silicon-rich ONO structure according to claim 8, characterized in that: The convolutional neural network adopts the ResNet18 model based on the residual neural network.

10. The silicon-rich ONO structure according to claim 1, characterized in that: The deposition process parameters of the first charge trapping layer, the second charge trapping layer and the third charge trapping layer include deposition temperature, deposition time, pressure, NH3 gas flow rate, SiH2Cl2 gas flow rate, the position of the silicon wafer in the furnace tube machine, and the point coordinates of the silicon wafer measurement points.

Images