Training method of neural network model and simulation method of ion implantation process

Abstract

Embodiments of the present disclosure provide a training method of a neural network model and a simulation method of an ion implantation process. The training method comprises: obtaining first training data and second training data related to the ion implantation process, the implanting ions colliding with target atoms of a device to be implanted in the ion implantation process, the first training data being associated with states of the implanting ions and the target atoms before the collision, and the second training data being associated with states of the implanting ions and the target atoms after the collision; inputting the first training data to the neural network model; and updating the neural network model based on a difference between a result output by the neural network model and the second training data.

Description

Technical Field

[0001] Embodiments of this disclosure relate to semiconductor and integrated circuit design techniques, and more specifically, to methods for training neural network models, neural network models, methods for simulating ion implantation processes, electronic devices, and computer-readable storage media. Background Technology

[0002] Ion implantation is a crucial step in semiconductor and integrated circuit manufacturing, involving the implantation of charged ions with specific energy into semiconductor materials. By precisely implanting accelerated, high-energy charged ion beams into the material matrix, the spatial distribution of dopant elements can be controlled and adjusted. This allows for precise and controllable alteration of the electrical properties of semiconductor materials, thereby enabling integrated circuits to achieve complex functions. As semiconductor device dimensions continue to shrink, the precision required for dopant distribution becomes increasingly stringent.

[0003] Currently, to optimize semiconductor processes, technology-aided design (TCAD) can be used to simulate physical processes such as ion implantation, diffusion, and oxidation, thereby providing predictive results for real-world processes. For example, the Monte Carlo method can be used for ion implantation simulation. However, these methods suffer from computational complexity, high memory consumption, and long simulation times, making it difficult to meet the evolving needs of semiconductor processes. Summary of the Invention

[0004] In view of the above problems, according to the exemplary embodiments of the present disclosure, a method for training a neural network model, a neural network model, a method for simulating an ion implantation process, an electronic device, and a computer-readable storage medium are provided.

[0005] In a first aspect of this disclosure, a method for training a neural network model is provided, comprising: acquiring first training data and second training data related to an ion implantation process, wherein the implanted ion collides with a target atom of a device to be implanted during the ion implantation process; the first training data being associated with the states of the implanted ion and the target atom before the collision; and the second training data being associated with the states of the implanted ion and the target atom after the collision; inputting the first training data into a neural network model; and updating the neural network model based on the difference between the output of the neural network model and the second training data.

[0006] In a second aspect of this disclosure, a neural network model is provided, generated using the method according to the first aspect.

[0007] In a third aspect of this disclosure, a method for simulating an ion implantation process is provided, comprising: generating second data based on first data using a neural network model according to the second aspect, wherein implanted ions collide with target atoms of a device to be implanted during the ion implantation process, the first data being associated with the states of the implanted ions and target atoms before the collision, and the second data being associated with the states of the implanted ions and target atoms after the collision; and generating simulation results of the ion implantation process based on the second data.

[0008] In a fourth aspect of this disclosure, an electronic device is provided, comprising: a processor; and a memory coupled to the processor, the memory having instructions stored therein, the instructions causing the electronic device to perform a method according to the first or third aspect when executed by the processor.

[0009] In a fifth aspect of this disclosure, a computer-readable storage medium is provided, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, implements the method according to the first or third aspect.

[0010] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] The above and other features, advantages, and aspects of the various embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0012] Figure 1 A schematic diagram of a scenario of an ion implantation process according to an embodiment of the present disclosure is shown.

[0013] Figure 2 A schematic diagram of a scenario of an ion implantation process according to an embodiment of the present disclosure is shown.

[0014] Figure 3 A schematic diagram of a computing device according to an embodiment of the present disclosure is shown.

[0015] Figure 4 A schematic flowchart illustrating a method for training a neural network model according to an embodiment of the present disclosure is shown.

[0016] Figure 5 A schematic flowchart illustrating the process of acquiring first training data and second training data according to an embodiment of the present disclosure is shown.

[0017] Figure 6A schematic diagram of a neural network model according to an embodiment of the present disclosure is shown.

[0018] Figure 7 A schematic flowchart illustrating a method for simulating an ion implantation process according to an embodiment of the present disclosure is shown.

[0019] Figure 8 A schematic flowchart of an additional process for simulating an ion implantation process according to an embodiment of the present disclosure is shown.

[0020] Figure 9 A schematic block diagram of an example device that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0021] The embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0022] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The terms "an embodiment / example" or "this embodiment / example" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0023] Figure 1 A schematic diagram of scenario 100 of an ion implantation process according to an embodiment of the present disclosure is shown. Figure 1 As shown, in scenario 100, device 110 is implanted with ions 120. As an example, device 110 can be a semiconductor device, such as a transistor, memory, or other type of device in the process of manufacturing. By performing ion implantation, the electrical characteristics of device 110 can be altered to achieve the desired performance. For example, in the case where device 110 is a transistor in the process of manufacturing, ion implantation can form specific regions, such as well regions, source regions, or drain regions.

[0024] Figure 2 A schematic diagram of scenario 200 of an ion implantation process according to an embodiment of the present disclosure is shown. Figure 2 As shown, as an example, Figure 1The device 110 can have a three-dimensional crystal structure 210 (only a planar view of the structure is shown schematically in the figure), with atoms 211 at each lattice site. During ion implantation, ions 220 can be implanted into the crystal structure 210 of the device 110 and collide with one or more atoms 211. It is understood that... Figure 2 Only an example implementation of crystal structure 210 is shown, but the specific structure of crystal structure 210 is not limited to this and can be any other structure.

[0025] As mentioned earlier, techniques such as the Monte Carlo method can be used to simulate semiconductor process steps like ion implantation, thereby providing predictions for real-world processes. The Monte Carlo method uses physical models to determine ion trajectories and final distributions by simulating the random collisions of individual ions with atoms in the material. This method requires tracking detailed trajectories of a large number of ions and involves highly complex calculations and lengthy simulations. For example, some simulations may require tracking tens of thousands to millions of ion trajectories, with computation times lasting hours or even days. Therefore, traditional methods cannot meet the demands of rapid iteration, development, and optimization in modern semiconductor processes. Furthermore, traditional methods require extensive recalculation when dealing with new materials, new processes, or complex three-dimensional device structures, lacking flexibility and adaptability.

[0026] This disclosure provides an improved scheme for simulating ion implantation processes. In this improved scheme, a neural network model is trained using training data related to the ion implantation process to obtain a neural network model that can replace the Monte Carlo method for predicting ion implantation processes. The trained neural network model can complete the simulation of the ion implantation process in a shorter time, thereby significantly reducing the development time and cost of semiconductor devices. Furthermore, in some embodiments, this neural network model can support multiple ion species and material types with high accuracy, and can quickly adapt to new materials, new processes, and complex three-dimensional device structures.

[0027] Figure 3 A schematic diagram of a computing device 300 according to an embodiment of the present disclosure is shown. The computing device 300 can be any device with computing capabilities. In one example, the computing device can be any type of fixed computing device, mobile computing device, or portable computing device; for example, computing devices include, but are not limited to, desktop computers, laptop computers, notebook computers, netbook computers, tablet computers, multimedia computers, mobile phones, etc. In another example, all or some components of the computing device can be distributed in the cloud. Figure 3As shown, computing device 300 may deploy a neural network model 310. This neural network model 310 can be trained based on training data input to or stored in computing device 300. In some embodiments, computing device 300 may also deploy a physical model 320 related to the ion implantation process, which can simulate the collision process of ions and atoms during ion implantation, thereby predicting the post-collision ion positions and determining the ion concentration distribution in the device using Monte Carlo methods. In one example, physical model 320 may be based on the Lindhard-Scharff-Schiott theory. It is understood that computing device 300 may also not deploy physical model 320, or physical model 320 may be deployed on another computing device.

[0028] Figure 4 A schematic flowchart of a method 400 for training a neural network model 310 according to an embodiment of the present disclosure is shown. The ion implantation process involved in method 400 can be performed in... Figure 1 Exemplary scenario 100 and Figure 2 The method 400 is implemented in the exemplary scenario 200. Furthermore, the method 400 can be executed by the computing device 300.

[0029] At box 401, computing device 300 acquires first training data and second training data related to the ion implantation process. For example, during the ion implantation process, the implanted ion 220 collides with a target atom 211 of the implantation device 110. The first training data relates to the states of the implanted ion 110 and the target atom 211 before the collision, and the second training data relates to the states of the implanted ion 110 and the target atom 211 after the collision. As an example, the first and second training data may relate to a single or multiple collisions between the ion and the atom, and include parameters related to the collision process and their changes. The first and second training data may be ion implantation process data acquired in a highly accurate manner, such as data acquired through a high-precision physical model. The first and second training data may be acquired directly by computing device 300 from an external device. Additionally or alternatively, computing device 300 may receive external data and acquire the first and second training data by processing the external data.

[0030] At box 402, computing device 300 inputs first training data into neural network model 310. That is, computing device 300 can input the first training data as input feature data into the neural network model 310 to be trained. For example, the first training data can be in the form of a feature vector and include one or more feature parameters. Neural network model 310 can perform forward propagation on the input data and output a prediction result based on the input data.

[0031] At box 403, computing device 300 updates neural network model 310 based on the difference between the output of neural network model 310 and the second training data. As mentioned earlier, the second training data represents collision results with higher accuracy or precision. Therefore, the current neural network model 310 can be updated based on the difference between the predicted results output by neural network model 310 and the second training data, for example, by adjusting the weights and biases of neural network model 310. By updating neural network model 310, it can continuously approach a model that can more accurately predict collision results.

[0032] It is understandable that the processes in boxes 401 to 403 can be repeated. Thus, computing device 300 can acquire more training data and repeatedly execute boxes 401 to 403 to continuously train the neural network model 310 until the model reaches the desired performance. For example, a training sample set can be collected and built, which can then be used to train the neural network model 310 until training of all samples in the training sample set is complete. In this case, the first training data and the second training data can be used to train the sample data in the training sample set. For example, the first training data and the second training data belong to the same sample, and boxes 401 to 403 represent the training process for a single sample. In some examples, a larger dataset can be built, which may include a training sample set for training the neural network model, a validation sample set for validating the neural network model, and a test sample set for testing the neural network model. For example, the validation sample set can be used to optimize the hyperparameters of the neural network, select from multiple models with different architectures trained on the training sample set, and implement other validation functions, thereby further updating and optimizing the neural network model. For example, the test sample set can be used to independently evaluate the performance of the neural network model to determine whether it meets the requirements. When validating or evaluating a neural network model using a validation or test set, the standard deviation can be determined based on the predicted values ​​to further calculate the confidence level. For example, multiple predicted values ​​can be generated through multiple forward propagations, and the standard deviation of these predicted values ​​can be calculated to evaluate the model's confidence. The smaller the standard deviation, the higher the confidence of the model in the prediction results. To generate multiple predicted values, the Monte Carlo Dropout method can be used. This method estimates model uncertainty by maintaining the activation state of the Dropout layer during testing. Multiple forward propagations are performed on the input data, with the Dropout layer randomly dropping some neurons each time, and the predicted values ​​from each forward propagation are recorded. As an example, the following equation can be used to calculate the average as the final prediction result:

[0033] (1)

[0034] Where y t Let represent the predicted value of the t-th forward propagation, where T represents the number of forward propagations. This represents the average of the predicted values. The standard deviation of all predicted values ​​can then be calculated based on the following equation:

[0035] (2)

[0036] in The standard deviation represents the model's uncertainty. After determining the standard deviation, the confidence level can be calculated using the following equation:

[0037] (3)

[0038] Where max_stdev represents the maximum value of the standard deviation of the predicted values, i.e. Therefore, trained neural network models can be evaluated during validation and testing based on confidence levels.

[0039] The validation and test sets can be collected in the same way as the training set. For example, the dataset can be randomly divided into training, validation, and test sets in a certain proportion to ensure that various collision conditions are evenly distributed across the sets. In one example, the training set could comprise 70% of the dataset, the validation set 15%, and the test set 15%. More broadly, the data in the validation and test sets can also be considered for training and updating the neural network model; therefore, the first and second training data can also be sample data from the validation and test sets.

[0040] In some embodiments of this disclosure, the first training data may include the energy of the injected ion 220 before the collision, the energy of the target atom 211 before the collision, the position of the injected ion 220 before the collision, the position of the target atom 211 before the collision, the velocity of the injected ion 220 before the collision, and the velocity of the target atom 211 before the collision. The second training data may include the energy loss of the injected ion 220 after the collision, the energy increment of the target atom 211 after the collision, the deflection angle of the injected ion 220 after the collision, and the deflection angle of the target atom 211 after the collision. Depending on actual needs, the first training data and the second training data may also include only one or more of the above items; for example, the first training data may not include the energy of the target atom 211 before the collision. Furthermore, it is understood that the above description of the first training data and the second training data is not limiting; in some cases, the first training data and the second training data may also include other parameters not listed above.

[0041] In some embodiments of this disclosure, the positions of the injected ion 220 and the target atom 211 before the collision are represented by their relative positions relative to the target atom 211. Alternatively or additionally, the velocities of the injected ion 220 and the target atom 211 before the collision are represented by their relative velocities. Specifically, to facilitate the training of the neural network model, the first training data can be optimized. By processing or converting the positions of the ion 220 and the atom 211 to their relative positions relative to the atom 211, and by processing or converting the velocities of the ion 220 and the atom 211 to their relative velocities relative to the atom 211, the amount of input data to the neural network can be effectively reduced, avoiding data redundancy. In this way, the complexity of the neural network can be reduced without affecting the training results. In one example, the optimized first training data may include the energy of the injected ion 220 (e.g., 10 keV), the relative position of the injected ion 220 with respect to the target atom 211 (e.g., the relative position of the injected ion 220 in a coordinate system constructed with the target atom 211 as the origin), and the relative velocity of the injected ion 220 with respect to the target atom 211.

[0042] In some embodiments of this disclosure, the acquired first and second training data satisfy the following conditions: the implanted ion is of a predetermined type; the target atom is of a predetermined type; and the energy of the implanted ion before collision is within a predetermined range. Specifically, the training data for the neural network model can be collected and acquired based on the ion type, atom type, and energy range. Thus, more targeted data can be used to train the corresponding neural network model. As an example only, ion types may include: aluminum, antimony, arsenic, boron, carbon, fluorine, gallium, germanium, indium, nitrogen, phosphorus, silicon, etc. The material types of the device to be implanted may include: silicon, germanium, silicon carbide (SiC), gallium arsenide (GaAs), gallium nitride (GaN), indium phosphide (InP), silicon-germanium alloy (SiGe), diamond, silicon dioxide (SiO2), and silicon nitride (Si3N4). Energy ranges can include: below 10keV, 10keV-100keV, 100keV-500keV, 500keV-1000keV, and above 1000keV.

[0043] For example, data from scenarios where boron is the implanted ion, silicon is the target atom, and the boron ion energy is greater than 10 keV can be collected as training data to form a boron-silicon-high-energy (greater than 10 keV) dataset or sample set. Similarly, datasets (or sample sets) for boron-silicon-low-energy (not greater than 10 keV) and arsenic-silicon dioxide-50 keV-200 keV (50 keV ≤ energy ≤ 200 keV) can be formed. The first and second training data can come from such specific datasets or sample sets. During ion implantation, the collision process will differ significantly if the types of ions and atoms and the ion energies are different. Therefore, using training data with the same type of ions and atoms in the same energy range helps improve training efficiency and obtain a more efficient and accurate dedicated neural network in subsequent neural network training. Alternatively, the first and second training data can also satisfy one or two of the above three conditions. That is, the training data used for the neural network model can be collected and obtained based on any one or any two of the ion type, atom type, and energy type. For example, data collected using boron as the implanted ion and silicon as the target atom can be used as training data to form a boron-silicon dataset or sample set, regardless of ion energy. Collecting training data considering only one or both of the ion type, atom type, and energy type also helps in training relatively efficient and accurate neural networks. Collecting specific training datasets for training dedicated neural network models can complete the training process with fewer iterations and provide better prediction accuracy. For instance, training low-energy and high-energy neural networks with 100,000 data points each on a low-energy dataset and a high-energy dataset of 100,000 data points respectively requires 80 iterations for each network to achieve 98% accuracy in predicting both high and low energy data. However, training a neural network on a mixed dataset of the same amount of high and low energy data requires 300 iterations to achieve the same accuracy.

[0044] In some embodiments of this disclosure, the first training data further includes at least one of the following: the mass of the implanted ion 220, the atomic number of the implanted ion 220, the mass of the target atom 211, the atomic number of the target atom 211, and the atomic density of the material of the device 110 to be implanted. Specifically, the training data for the neural network may also disregard ion and atom types, thereby training a general-purpose neural network model rather than a dedicated neural network model. In this case, the first training data may include feature parameters indicating the types and densities of the implanted ions and target atoms, such as mass, atomic number, and atomic density. Compared to dedicated neural network models, general-purpose neural network models may require more training data and iterations to achieve performance and accuracy comparable to dedicated neural network models. However, general-purpose neural network models are applicable to the simulation needs of new materials, new processes, and new device structures, and can be applied to more application scenarios, thus possessing stronger versatility and scalability.

[0045] Figure 5 A schematic flowchart of a process 500 for acquiring first training data and second training data according to an embodiment of the present disclosure is shown. Process 500 may be implemented at block 401 of method 400.

[0046] At box 501, computing device 300 generates second initial data based on first initial data and using a physical model 320 of the ion implantation process. The first initial data relates to the states of the implanted ion 220 and the target atom 211 before the collision, and the second initial data relates to the states of the implanted ion 220 and the target atom 211 after the collision. Specifically, training data can utilize... Figure 3 The physical model 320 is used to generate the data. For example, the Monte Carlo method can be used to randomly generate initial ion positions, and a large number of ion implantation simulations can be performed using a collision physical model (such as a model based on the Lindhard-Scharff-Schiott theory). This allows recording data such as the energy, position, and velocity of the implanted ions and target atoms before the collision, and determining or calculating data such as the energy loss and deflection angle of the implanted ions and the energy increment and deflection angle of the target atoms caused by the collision based on the physical model. The relevant data before the collision can be used as the first initial data, and the relevant data after the collision can be used as the second initial data. In one example, for a dedicated neural network model, a dedicated physical model can be used to generate data. This dedicated physical model considers factors such as ion type, atom type, ion energy, implantation angle, and device material density, thereby generating more efficient and targeted data for subsequent training.

[0047] At box 502, computing device 300 generates first training data and second training data based on first initial data and second initial data, respectively. The first initial data and second initial data can be processed to form the first training data and second training data, or they can be directly used as the first training data and second training data. In one example, the data generated using physical model 320 can form a dataset or sample set, or the data generated using physical model 320 can be processed to form a dataset or sample set. Thus, the first training data and second training data can be obtained from the dataset or sample set generated by the physical model. In some embodiments, computing device 300 can normalize the first initial data and second initial data. Computing device 300 can use the normalized first initial data as the first training data and the normalized second initial data as the second training data. As an example, a specified maximum value can be used as the normalization denominator for the corresponding parameter, or the maximum value of the corresponding parameter in the dataset generated using the physical model can be used as the normalization denominator for that parameter, thereby normalizing the parameter in the first initial data and second initial data to a value in the range of 0 to 1. For example, the energy of the implanted ion before collision is normalized using a specified maximum energy or the maximum energy value in the dataset as the normalization denominator; the relative position of the implanted ion relative to the target atom is normalized using a specified maximum position or the maximum relative position value in the dataset as the normalization denominator; the relative velocity of the implanted ion relative to the target atom is normalized using a specified maximum velocity or the maximum relative velocity value in the dataset as the normalization denominator; the mass of the implanted ion and the mass of the target atom are normalized using a specified maximum mass or the maximum mass value in the dataset as the normalization denominator; the atomic number of the implanted ion and the atomic number of the target atom are normalized using a specified maximum atomic number or the maximum atomic number value in the dataset as the normalization denominator; the atomic density of the material of the device to be implanted is normalized using a specified maximum atomic density or the maximum atomic density value in the dataset as the normalization denominator; the energy loss of the implanted ion and the energy increment of the target atom are normalized using a specified maximum energy or the maximum energy value in the dataset as the normalization denominator; and π The deflection angles of the injected ions and the target atoms are normalized using radians as the normalization denominator. Normalizing the data effectively improves the training efficiency and stability of neural network models.

[0048] Figure 6A schematic diagram of a neural network model 310 according to an embodiment of the present disclosure is shown. As an example, the neural network model 310 may be a specialized neural network model, trained using a dataset or sample set, for example, a specific ion type, a specific target atom type, and / or a specific energy range. Alternatively, the neural network model 310 may be a general-purpose neural network model, in which case data and parameters related to factors such as the injected ion type and the target atom type may be added to the training data for training.

[0049] like Figure 6 As shown, the neural network model 310 may include an input layer 311, an intermediate layer 312, and an output layer 313. Specifically, the neural network model 310 may have a multi-layer neural network architecture. The input layer 311 is responsible for receiving the first training data, such as the data feature vector before the collision. The input layer 311 may have multiple neurons, for example, these neurons may be used to receive input features such as the energy of the injected ion, the relative position of the injected ion relative to the target atom, and the relative velocity of the injected ion relative to the target atom. The intermediate layer 312 has multiple fully connected layers, each composed of multiple neurons. The number of layers in the intermediate layer 312 and the number of neurons in each layer can be optimized according to experimental results and model performance. In one example, the number of layers in the intermediate layer 312 may be between 3 and 5, and the number of neurons in each layer may be between 64 and 256. The output layer 313 contains multiple neurons, for example, these neurons may correspond to output features such as the energy loss of the injected ion, the deflection angle of the injected ion, the energy increment of the target atom, and the deflection angle of the target atom.

[0050] In some embodiments of this disclosure, the activation function of the intermediate layer 312 includes the ReLU (Rectified LinearUnit) function, and / or the activation function of the output layer 313 includes the Sigmoid function. Specifically, an activation function is a mathematical function on a neuron in a neural network that determines whether the neuron should be "activated" or "triggered." The main function of an activation function is to perform a nonlinear transformation on the input signal from the neuron in the previous layer and then output it to the next layer. The choice of activation function is crucial to the performance of the neural network. By using the ReLU function in the intermediate layer, the vanishing gradient problem can be effectively solved, accelerating the training of the neural network model. The ReLU function can be expressed by the following equation:

[0051] (4)

[0052] The max function represents the maximum value operation, and x represents the variable input to the ReLU function.

[0053] Using the Sigmoid function in the output layer 313 maps the neural network's output values ​​to the [0,1] interval, thus facilitating the processing of normalized output data. The Sigmoid function can be expressed by the following equation:

[0054] (5)

[0055] In some embodiments of this disclosure, computing device 300 can update the neural network model based on the difference between the output of neural network model 310 and second training data, using a loss function including a mean squared error function. A loss function is a function used to measure the difference between model predictions and actual values. The mean squared error function can be used to evaluate the difference between the current neural network model's predictions and the second training data, thereby helping to determine how to adjust the parameters of the neural network model to narrow the gap between the predicted and actual values. The mean squared error function can be expressed by the following equation:

[0056] (6)

[0057] MSE (Mean Squared Error) represents the mean squared error. y i This represents the actual value in the second training data. This represents the predicted value output by the neural network model, where N represents the number of samples. Backpropagation can be used to calculate the gradient of the loss function (e.g., mean squared error) with respect to each model parameter (e.g., weights and biases). The calculated gradient indicates how the model parameters should be adjusted to reduce the value of the loss function (e.g., MSE). The neural network model can then use optimizers (e.g., Adam, RMSprop) to adjust the model parameters based on the gradient, thereby optimizing model performance. After repeated training and parameter optimization, the neural network model can replace the physical model to predict collisions during ion implantation, significantly reducing simulation time, improving computational efficiency, and meeting the demands of rapid iteration in modern semiconductor processes.

[0058] Using the training methods described above, multiple specific versions of neural network models can be trained for common ion implantation types, material types, and implantation energies to meet the ion implantation prediction needs under these specific conditions. Furthermore, to cope with variable and non-specific simulation conditions, a general-purpose neural network model can also be trained, enabling it to handle ion implantation predictions under different scenarios. This improves the accuracy and efficiency of ion implantation simulations, and by selecting the neural network model best suited to the current conditions, ensures the reliability of the prediction results.

[0059] Figure 7A schematic flowchart of a method 700 for simulating an ion implantation process according to an embodiment of the present disclosure is shown. The ion implantation process involved in method 700 can be performed in... Figure 1 Exemplary scenario 100 and Figure 2 The ion implantation process involved in the simulation method 700 may be different from or the same as the ion implantation process involved in the training method 400. Furthermore, method 700 may be executed by computing device 300.

[0060] At box 701, the neural network model generated or trained using method 400 generates second data based on first data. The first data is associated with the states of the implanted ion 220 and the target atom 211 before the collision, and the second data is associated with the states of the implanted ion 220 and the target atom 211 after the collision. In some embodiments, the first data may include at least one of the following: the energy of the implanted ion 220 before the collision, the energy of the target atom 211 before the collision, the position of the implanted ion 220 before the collision, the position of the target atom 211 before the collision, the velocity of the implanted ion 220 before the collision, and the velocity of the target atom 211 before the collision. When the neural network model is a general neural network model, the first data may also include at least one of the following: the atomic mass of the implanted ion, the atomic number of the implanted ion, the atomic mass of the target atom, the atomic number of the target atom, and the material density of the device to be implanted. If the neural network model is optimized for the input training data during training, the first data also needs to be optimized, for example, by converting the positions of the implanted ion and the target atom into relative positions of the implanted ion and the target atom, and by converting the velocities of the implanted ion and the target atom into relative velocities of the implanted ion and the target atom. If the neural network model normalizes the input training data during training, the first data also needs to be normalized. The normalization method is similar to that of the training data, and therefore will not be described in detail. In some embodiments, the second data includes at least one of the following: the energy loss of the injected ion 220 after the collision, the energy increment of the target atom 211 after the collision, the deflection angle of the injected ion 220 after the collision, and the deflection angle of the target atom 211 after the collision. Based on the second data, parameters such as the position, velocity, and energy of the ion and atom after the collision can be determined through further calculation. It is understood that, similar to the training data, the first data and the second data may also include other types of parameters, and this disclosure does not limit this.

[0061] At box 702, the simulation results of the ion implantation process are generated based on the second data. The second data output by the neural network model can predict the collision results, thereby updating the ion's running state and ultimately determining the ion's stopping position. As an example, the neural network model can output the predicted collision event results, such as the ion's energy loss and deflection angle, the target atom's energy increment and deflection angle, etc. Since the ion's energy, velocity, and position before the collision are known, the ion's energy, velocity, and position after the collision can be calculated based on the ion's energy loss, ion deflection angle, target atom's energy increment, and target atom deflection angle. This determines the position of the target atom in the next collision, and this is used as input features to predict the next collision using the neural network model. If the first data input to the neural network model is normalized, the second data needs to be denormalized to obtain the corresponding data. The above simulation process can be repeated until the predicted ion stops after continuously losing energy. Thus, the neural network model can be used to simulate the physical process of a large number of incident ions continuously colliding with material atoms during ion implantation, thereby obtaining ion implantation simulation results with high accuracy and shorter simulation time. In some embodiments, the simulation results of the ion implantation process include the concentration distribution of implanted ions in the device to be implanted. The ion concentration distribution in the device can be determined by identifying the stopping positions of a large number of ions after a collision. Therefore, the process parameters of ion implantation can be adjusted and optimized based on the simulation results, thereby achieving improvements in semiconductor processes at a lower cost. It is understood that, in addition to ion concentration distribution, the simulation results of the neural network model may also include other results of the ion implantation process, such as lattice damage or defect distribution.

[0062] In some embodiments of this disclosure, the computing device 300 can select a desired neural network model from multiple neural network models based on at least one of the implanted ion type, the target atom type, and the energy of the implanted ion. The computing device 300 can use the selected neural network model to generate second data based on the first data. Specifically, as described above, different datasets can be used to train dedicated neural network models. For example, a boron-silicon-low energy (not greater than 10 eV) dataset can be used to train a boron-silicon-low energy neural network model, which can train multiple high-performance dedicated neural network models with less data and less time. Thus, when performing simulations using neural network models, the corresponding neural network model can be selected based on the ion type, material atom type, and energy range to be simulated. For example, for ion implantation of low-energy boron ions colliding with silicon atoms, a boron-silicon-low energy neural network model can be used for prediction. Furthermore, when the device material is a compound, multiple neural network models can be used for prediction. For example, in a scenario where the implanted ion is arsenic and the device material is silicon dioxide, an arsenic-oxygen dataset is needed to train an arsenic-oxygen neural network model, and an arsenic-silicon dataset is needed to train an arsenic-silicon neural network model. Therefore, when predicting arsenic ion implantation into silica materials, an arsenic-silica neural network, including an arsenic-oxygen neural network model and an arsenic-silicon neural network model, can be used to predict collision and ion implantation results.

[0063] In some cases, if a corresponding neural network model is unavailable, the closest possible model can be chosen, for example, if one or both of the ion type, material atom type, and energy range are the same. Alternatively, a general-purpose neural network model can be selected if a corresponding model is not available. When using a general-purpose neural network model, the first data input to the model must also include parameters related to the injected ion type, target atom type, and energy range. For example, assuming the injected ions include element A (mass 1) and element C (mass 3), and the target atoms include element X (mass 12) and element Z (mass 14), a general-purpose neural network model can be trained using the AX, AZ, CX, and CZ datasets (these datasets need to include parameters such as the injected ion type, target atom type, and energy range). This general neural network model can predict well the collision of an implanted ion of element B (mass 2) with the target atom of element Y (mass 13), or the collision of an implanted ion of element B with the target atom of element X or element Z, or the collision of an implanted ion of element C with the target atom of element Y, or the collision of an implanted ion of element B with the compound of elements X and Y, the compound of elements X and Z, or the compound of elements X, Y, and Z, etc.

[0064] Figure 8A schematic flowchart of an additional process 800 of a method 700 for simulating an ion implantation process according to an embodiment of the present disclosure is shown. The additional process 800 may be performed before or after blocks 701 and 702 of method 700, or may be performed in parallel with blocks 701 and 702.

[0065] At box 801, computing device 300 determines the confidence level of the second data based on the standard deviation of the second data generated multiple times. Computing device 300 can use a neural network model trained according to the aforementioned method to generate the second data multiple times based on the first data. Thus, the standard deviation can be calculated based on these second data to determine the confidence level. That is, multiple predicted values ​​can be generated using multiple forward propagations of the neural network model, and the confidence level can be determined by calculating the standard deviation of these predicted values ​​to evaluate the neural network model. The confidence level can be calculated or determined using equations (1) to (3) described above, and therefore will not be repeated. In some embodiments, for each generation of the second data, the second data can be generated based on the first data using a neural network model after randomly discarding at least one neuron. For example, the Monte Carlo Dropout method can be used to discard some neurons in each forward propagation. Thus, different neural network structures can be created in each forward propagation, thereby improving the reliability of the confidence level.

[0066] At box 802, computing device 300 determines whether the determined confidence level is lower than a threshold. For example, the threshold can be set to 90%.

[0067] At box 803, if the confidence level is below a threshold, the computing device 300 generates third data based on the first data using a physical model 320 of the ion implantation process. This third data is correlated with the states of the implanted ions and the target atoms after the collision. In other words, if the confidence level is not high enough, it indicates that the reliability of the neural network model's prediction is poor. If the confidence level is not lower than the threshold, process 800 can end.

[0068] At box 804, computing device 300 adds the first data and the third data as new data to the training dataset or sample set, which includes the data used when training the neural network model.

[0069] At box 805, computing device 300 determines that the amount of new data in the training dataset or sample set exceeds a predetermined proportion.

[0070] At box 806, if the amount of new data exceeds a predetermined proportion, computing device 300 retrains the neural network model based on the training dataset including the new data. Excessive new data implies that the current neural network's prediction reliability is poor, or that it is not adept at predicting and simulating the current ion implantation scenario. Therefore, the neural network model can be retrained using a dataset including the new data, for example, using the methods described above. For example, the predetermined proportion could be 5%, in which case, if the proportion of new data to the dataset exceeds 5%, the neural network model can be retrained. If the amount of new data does not exceed the predetermined proportion, process 800 can end.

[0071] At box 807, using a retrained neural network model, computing device 300 generates fifth data based on the fourth data. The fourth data relates to the states of the injected ions and the target atom before the collision, and the fifth data relates to the states of the injected ions and the target atom after the collision. In other words, the retrained neural network model can replace the previous neural network model and be used for subsequent predictions and simulations.

[0072] In this way, active learning strategies can be applied to neural network models, giving them adaptive learning capabilities and enabling them to continuously improve simulation accuracy and efficiency over time. Furthermore, this neural network model can continuously expand the coverage of data predictions. Therefore, it can continuously adapt to the simulation needs of new materials, new processes, and new device structures, exhibiting strong versatility and scalability.

[0073] This disclosure provides an ion implantation process simulation technology driven by a neural network algorithm. This technology, by constructing a trajectory prediction model based on a neural network, achieves efficient and accurate simulation of ion trajectory, energy loss, and final spatial distribution during ion implantation. It can be widely applied in fields such as integrated circuit manufacturing process optimization and semiconductor process simulation software. The neural network model significantly reduces simulation time, improves computational efficiency, and meets the demands of rapid iteration in modern semiconductor processes. Furthermore, some embodiments of this disclosure support multiple ion types and material types, enabling rapid adaptation to new materials, new processes, and complex three-dimensional device structures. Some embodiments of this disclosure also provide an adaptive learning mechanism for the neural network model, allowing the system to continuously expand the coverage of training data, particularly increasing sample density in the edge and sparse regions of the parameter space, thereby improving the model's adaptability and versatility.

[0074] Figure 9 A schematic block diagram of an example device 900 that can be used to implement embodiments of the present disclosure is shown. Device 900 may be... Figure 3 An example implementation of computing device 300 in the example, and used for implementation Figure 4 Method 400 Figure 5 Process 500 Figure 7 Method 700 and Figure 8 The process is 800.

[0075] like Figure 9 As shown, device 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) 902 or loaded from storage unit 908 into random access memory (RAM) 903. RAM 903 may also store various programs and data required for the operation of device 900, such as the datasets or sample sets mentioned above. CPU 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0076] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0077] Processing unit 901 executes the methods or processes described above, method 400, process 500, method 700, and process 800. For example, in some embodiments, method 400, process 500, method 700, and process 800 may be implemented as a computer software program or computer program product tangibly contained in a machine-readable medium, such as a non-transient computer-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program may be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by CPU 901, one or more steps of method 400, process 500, method 700, and process 800 described above may be performed. Alternatively, in other embodiments, CPU 901 may be configured to execute method 400, process 500, method 700, and process 800 by any other suitable means (e.g., by means of firmware).

[0078] Those skilled in the art will understand that the various steps of the methods disclosed above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using device-executable program code, which can then be stored in a storage device for execution by the computing device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this disclosure is not limited to any particular combination of hardware and software.

[0079] It should be understood that although several devices or sub-devices of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more devices described above can be embodied in one device. Conversely, the features and functions of one device described above can be further divided and embodied by multiple devices.

[0080] The above description is merely an optional embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for training a neural network model, comprising: Acquire first training data and second training data related to the ion implantation process, in which the implanted ion collides with the target atom of the device to be implanted during the ion implantation process. The first training data is associated with the state of the implanted ion and the target atom before the collision, and the second training data is associated with the state of the implanted ion and the target atom after the collision. The first training data is input into the neural network model; as well as The neural network model is updated based on the difference between the output of the neural network model and the second training data.

2. The method according to claim 1, wherein the first training data includes at least one of the following: the energy of the injected ion before the collision, the energy of the target atom before the collision, the position of the injected ion before the collision, the position of the target atom before the collision, the velocity of the injected ion before the collision, and the velocity of the target atom before the collision; and / or The second training data includes at least one of the following: the energy loss of the injected ion after the collision, the energy increment of the target atom after the collision, the deflection angle of the injected ion after the collision, and the deflection angle of the target atom after the collision.

3. The method according to claim 2, wherein: The positions of the injected ions and the target atoms before the collision are represented by their relative positions to the target atoms before the collision; and / or The velocity of the injected ion and the velocity of the target atom before the collision are represented by the relative velocity of the injected ion relative to the target atom before the collision.

4. The method according to claim 2, wherein the first training data further includes at least one of the following: the mass of the implanted ion, the atomic number of the implanted ion, the mass of the target atom, the atomic number of the target atom, and the atomic density of the material of the device to be implanted.

5. The method of claim 1, wherein the acquired first training data and second training data satisfy at least one of the following: The implanted ions are of a predetermined type; The target atom is of a predetermined type; and The energy of the injected ions before the collision is within a predetermined range.

6. The method according to any one of claims 1 to 5, wherein acquiring the first training data and the second training data related to the ion implantation process comprises: Based on the first initial data, a second initial data is generated using a physical model of the ion implantation process. The first initial data is associated with the state of the implanted ion and the target atom before the collision, and the second initial data is associated with the state of the implanted ion and the target atom after the collision. as well as The first training data and the second training data are generated based on the first initial data and the second initial data, respectively.

7. The method of claim 6, wherein generating the first training data and the second training data based on the first initial data and the second initial data respectively comprises: Normalize the first initial data and the second initial data; The first normalized initial data is used as the first training data, and the second normalized initial data is used as the second training data.

8. The method according to any one of claims 1 to 5, wherein the neural network model comprises an input layer, an intermediate layer, and an output layer. The activation function of the intermediate layer includes the ReLU function, and / or the activation function of the output layer includes the Sigmoid function.

9. The method according to any one of claims 1 to 5, wherein updating the neural network model based on the difference between the output of the neural network model and the second training data comprises: Based on the difference between the output of the neural network model and the second training data, the neural network model is updated using a loss function, which includes a mean squared error function.

10. A method for simulating an ion implantation process, the method comprising: Using a neural network model generated by the method according to any one of claims 1 to 9, second data is generated based on first data, wherein the implanted ion collides with the target atom of the device to be implanted during the ion implantation process, the first data is associated with the state of the implanted ion and the target atom before the collision, and the second data is associated with the state of the implanted ion and the target atom after the collision; as well as The simulation results of the ion implantation process are generated based on the second data.

11. The method of claim 10, further comprising: Based on at least one of the implanted ion type, target atom type, and implanted ion energy, a neural network model is selected from a plurality of neural network models generated by the method according to any one of claims 1 to 9. Generating the second data based on the first data includes: using a selected neural network model to generate the second data based on the first data.

12. The method of claim 10, wherein the first data includes at least one of the following: the energy of the implanted ion before collision, the energy of the target atom before collision, the position of the implanted ion before collision, the position of the target atom before collision, the velocity of the implanted ion before collision, the velocity of the target atom before collision, the mass of the implanted ion, the atomic number of the implanted ion, the mass of the target atom, the atomic number of the target atom, and the atomic density of the material of the device to be implanted, and / or The second data includes at least one of the following: the energy loss of the injected ion after the collision, the energy increment of the target atom after the collision, the deflection angle of the injected ion after the collision, and the deflection angle of the target atom after the collision.

13. The method of claim 10, wherein the simulation results of the ion implantation process include the concentration distribution of implanted ions in the device to be implanted.

14. The method of claim 10, wherein generating the second data based on the first data comprises: Using a neural network model generated by the method according to any one of claims 1 to 9, the second data is generated multiple times based on the first data; The method further includes: The confidence level of the second data is determined based on the standard deviation of the second data generated multiple times. In response to the determined confidence level being lower than a threshold, third data is generated based on the first data using a physical model of the ion implantation process, the third data being associated with the state of the implanted ion and the target atom after the collision; The first data and the third data are added as new data to the training dataset, which includes the data used when training the neural network model. In response to the amount of new data in the training dataset exceeding a predetermined proportion, the neural network model is retrained based on the training dataset including the new data; and Using the retrained neural network model, fifth data is generated based on fourth data, which is associated with the states of the injected ions and target atoms before the collision, and the fifth data is associated with the states of the injected ions and target atoms after the collision.

15. The method of claim 14, wherein generating the second data multiple times based on the first data comprises: For each generation of the second data, the second data is generated based on the first data using a neural network model generated by the method according to any one of claims 1 to 9 after randomly discarding at least one neuron.

16. An electronic device comprising: processor; as well as A memory coupled to a processor, the memory having instructions stored therein, which, when executed by the processor, cause the device to perform the method according to any one of claims 1 to 9 and 10 to 15.

17. A computer-readable storage medium storing machine-executable instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 9 and 10 to 15.

18. A computer program product comprising machine-executable instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 9 and 10 to 15.

Images