Simulation method for semiconductor device, electronic equipment and storage medium

Abstract

The embodiment of the invention relates to a simulation method for a semiconductor device, electronic equipment and a storage medium. The simulation method is executed by a simulation model, the simulation model comprises a CMP physical model, the CMP physical model is configured to determine a simulation value of the surface height of a polished wafer based on input data, and the input data comprises graphic feature data and physical information; the neural network model is generated by taking the simulation value as a training sample for training, and is configured to perform parallel prediction on multiple groups of input data in the simulation process of the CMP physical model so as to generate multiple predicted values of the surface height of the polished wafer corresponding to the multiple groups of input data, and respectively taking the simulation values as corresponding simulation values of the surface heights. According to the technical scheme, the establishment and simulation process of the CMP physical model can be remarkably accelerated, the time consumption is reduced, and the simulation precision can be improved.

Description

[0001] This application is a divisional application of the Chinese Patent Application No. 202510267351.7, filed on March 6, 2025, entitled “Simulation Method for Semiconductor Device, Electronic Device and Storage Medium”, with the earliest filing date of March 6, 2025. TECHNICAL FIELD

[0002] Embodiments of the present disclosure generally relate to integrated circuits, and more particularly, to a simulation method for semiconductor device, electronic device and storage medium. BACKGROUND

[0003] Current prediction of metal thickness after chemical mechanical polishing (CMP) post-process usually greatly depends on CMP model.

[0004] The traditional CMP model is mainly a time-step physical model based on the CMP itself physical process. The model has the characteristics of completely simulating the real physical process of CMP, high modeling accuracy, and can output all relevant variables of the intermediate process to help users better understand and adjust the CMP process recipe and predict the CMP polishing result. However, the disadvantage is that the calculation speed is slow. SUMMARY

[0005] According to example embodiments of the present disclosure, a simulation scheme for semiconductor device and a training method of neural network model are provided to at least partially overcome the above or other potential drawbacks.

[0006] According to an aspect of the present disclosure, a method for generating training samples is provided. The method includes generating a first number of feature vectors based on data points representing respective pattern features on a wafer, the feature vectors being combinations of data representing different types of pattern features in the data points; generating a second number of physical vectors based on physical information related to CMP of the wafer, the physical vectors being combinations of data representing different types of physical information; and performing, by a CMP physical model, a second number of simulations based on the feature vectors and the physical vectors to generate, in each simulation, simulation values of surface height of the wafer corresponding to the first number of feature vectors as training samples.

[0007] In some embodiments, generating the first number of feature vectors based on the data points representing respective pattern features on the wafer includes determining data within respective step ranges based on value ranges and corresponding step lengths of the data points of the respective pattern features, respectively; and performing data orthogonalization on the data within the respective step ranges to generate a feature data table, wherein the feature data table includes data of the first number of feature vectors.

[0008] In some embodiments, generating the second number of physical vectors based on the physical information related to the CMP of the wafer comprises: determining data of each step range based on the value range of the data points of each physical vector and the corresponding step range, respectively; and performing data orthogonalization on the data in each step range to generate a physical information data table, wherein the data of the second number of physical vectors are included in the physical information data table.

[0009] In some embodiments, performing the second number of simulations by the CMP physical model based on the feature vector and the physical vector comprises: generating a first simulation value of the surface height of the non-trench structure and a second simulation value of the surface height of the trench structure, respectively; and determining a simulation value of the wafer surface height based on the first simulation value and the second simulation value.

[0010] In some embodiments, the pattern features at least include: a density of the pattern within a grid on the wafer; a width of the pattern; and a spacing between the patterns.

[0011] In some embodiments, the physical information at least includes the following: information of the deposited material; maximum and minimum dish-shaped defect information of each step; slurry information; long tail effect parameter information; duration information; nominal pressure information.

[0012] According to a second aspect of the present disclosure, a training method of a neural network model is provided. The method comprises: training a neural network model using training samples to determine a gradient of a loss function of the neural network model, wherein the training samples are generated by a CMP physical model based on pattern features and physical information used to determine a surface height of a ground wafer; updating parameters of the neural network model based on the gradient of the loss function; and in response to the training satisfying a predetermined condition, determining a neural network model corresponding to the corresponding updated parameters as a trained neural network model, wherein the trained neural network model outputs a predicted value of the surface height.

[0013] In a third aspect of the present disclosure, an electronic device is provided. The electronic device comprises a processor; and a memory coupled with the processor, the memory having stored therein instructions that, when executed by the processor, cause the device to perform actions comprising: training a neural network model using training samples to determine a gradient of a loss function of the neural network model, wherein the training samples are generated by a CMP physical model based on pattern features and physical information used to determine a surface height of a ground wafer; updating parameters of the neural network model based on the gradient of the loss function; and in response to the training satisfying a predetermined condition, determining a neural network model corresponding to the corresponding updated parameters as a trained neural network model, wherein the trained neural network model outputs a predicted value of the surface height.

[0014] In some embodiments, training the neural network model with the training samples to determine the gradient of the loss function of the neural network model comprises: calculating the gradient of the loss function by using a stochastic gradient descent algorithm.

[0015] In some embodiments, updating the parameters of the neural network model based on the gradient of the loss function comprises: performing gradient update on the gradient of the loss function in response to that the training is performed for a predetermined number of times; and updating the parameters of the neural network model based on the gradient update.

[0016] In some embodiments, training the neural network model with the training samples to determine the gradient of the loss function of the neural network model comprises: training the neural network model with a part of the training samples to perform gradient descent backpropagation to update the parameters; predicting the remaining part of the training samples by the trained neural network model to generate predicted values of the surface height of the polished wafer; and determining the corresponding trained neural network model as the final neural network model in response to that a comparison result of the predicted values and the simulation values of the CMP physical model satisfies a predetermined condition.

[0017] In some embodiments, the comparison result of the predicted values and the simulation values satisfies the predetermined condition comprises: a root mean square error of the predicted values and the simulation values is less than a predetermined threshold.

[0018] In some embodiments, the method further comprises: determining a loss function value of the neural network model; and wherein the training satisfies the predetermined condition comprises: a difference value of the loss function values of two consecutive iterations is less than a predetermined difference threshold; or an absolute value of a ratio of the loss function values of two consecutive iterations is greater than a predetermined ratio threshold. In some embodiments, the training samples are generated in parallel on a plurality of servers, so that the training of the neural network model is performed in parallel.

[0019] In a fourth aspect of the present disclosure, a neural network model generated according to the method of the third aspect is provided.

[0020] In a fifth aspect of the present disclosure, a simulation model is provided, comprising: a CMP physical model configured to determine simulation values of the surface height of the polished wafer based on input data, wherein the input data comprises the pattern feature data and the physical information; and a neural network model according to the fifth aspect, the neural network model is generated by training with the simulation values as training samples, and is configured to predict a plurality of groups of input data in parallel in a simulation process of the CMP physical model to generate a plurality of predicted values of the surface height of the polished wafer corresponding to the plurality of groups of input data, respectively as the corresponding simulation values of the surface height.

[0021] In some embodiments, the parallel prediction of the multiple sets of input data in the simulation process of the CMP physical model comprises: parallel acceleration of the calculation process of the neural network model by using multiple graphic processors.

[0022] In a sixth aspect of the present disclosure, a simulation method for a semiconductor device is provided, which utilizes the simulation model of the fifth aspect to perform, and the method comprises: predicting, by a neural network model in the simulation model, a simulation value of a surface height of a polished wafer based on input data, wherein the input data comprises graphic feature data and physical information.

[0023] In some embodiments, the prediction of the simulation value of the surface height of the polished wafer by the neural network model in the simulation model based on the input data comprises: parallel prediction of multiple sets of input data by the neural network model to generate multiple predicted values of the surface height of the polished wafer corresponding to the multiple sets of input data, respectively as the corresponding simulation values of the surface height.

[0024] In some embodiments, the prediction of the simulation value of the surface height of the polished wafer by the neural network model in the simulation model based on the input data comprises: generating, by the neural network model, a first predicted value of the surface height of the non-trench structure and a second predicted value of the surface height of the trench structure, respectively; and determining, by the neural network model, the simulation value of the wafer surface height based on the first predicted value and the second predicted value.

[0025] In some embodiments, further comprising: determining, by a CMP physical model in the simulation model, a condition of the surface of the wafer in the polishing process.

[0026] In a seventh aspect of the present disclosure, a neural network model is provided, comprising: an input layer configured to receive graphic feature data on a wafer and physical information related to CMP of the wafer; an intermediate layer configured to perform the method of the sixth aspect to determine a surface height value of the polished wafer; and an output layer configured to output the surface height value.

[0027] In some embodiments, the method further comprises: an embedding layer configured to map the input physical information to generate a physical vector.

[0028] In an eighth aspect of the present disclosure, a computer readable storage medium is provided, which stores a computer program that is executed by a processor to implement the above method according to the present disclosure.

[0029] It will be understood from the following description that the technical solution of the present disclosure can significantly accelerate the establishment and simulation process of the CMP physical model, reduce the related time consumption, and improve the simulation accuracy.

[0030] The following presents a summary of the disclosure in order to provide a basic understanding of some aspects of the application. This summary is not an extensive overview of the disclosure and is not intended to identify key or critical elements of the application or to delineate the scope of the application. Its sole purpose is to present some concepts of the application in a simplified form as a prelude to the more detailed description that is presented later. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A schematic diagram illustrating an example environment in which embodiments of the present disclosure can be implemented is shown;

[0032] Figure 2 A flowchart illustrating a method of generating training samples according to some embodiments of the present disclosure is shown;

[0033] Figure 3 A flowchart illustrating a method of training a neural network model according to some embodiments of the present disclosure is shown;

[0034] Figure 4 A schematic diagram illustrating a neural network architecture according to some embodiments of the present disclosure is shown;

[0035] Figure 5 A block diagram of a computing device capable of implementing a number of embodiments of the present disclosure is shown.

[0036] In the various drawings, like or corresponding elements are denoted by like or corresponding reference numerals. DETAILED DESCRIPTION

[0037] The principles of the present disclosure will now be described, by way of example only, with reference to the various example embodiments and / or examples illustrated in the drawings that are intended to explain the principles of the present disclosure and to further enable those skilled in the art to appreciate the underlying principles and / or to implement an example embodiment of the present disclosure. It will be appreciated that the description set forth herein is intended as a description of example embodiments of the present disclosure and is not intended to limit the scope of the application. It will be further appreciated that the description set forth herein is not intended to limit the scope of the application to the particular examples presented herein. It will be appreciated that the description set forth herein is intended to be illustrative only and that the application will be limited only by the appended claims.

[0038] The term "including" and variations thereof, as used in this document, mean "including, but not limited to," unless expressly specified otherwise. The term "or" means "and / or" unless expressly specified otherwise. The term "based on" means "based, at least in part, on" unless expressly specified otherwise. The terms "one example embodiment" and "an example embodiment" mean "at least one example embodiment." The term "another embodiment" means "at least one additional embodiment." The terms "a first," "a second," etc. do not require that there be only one of the indicated items.

[0039] CMP is a chemical and physical combined material removal method, which has become an indispensable step in the modern integrated circuit (IC) industry. Specifically, CMP uses chemical corrosion and mechanical force to planarize silicon wafers or other substrate materials during processing. CMP can achieve nanoscale material removal, so that the wafer surface is planarized.

[0040] The CMP model, as the name implies, is a physical model for simulating the CMP process. The model outputs some important physical indicators of each grid point of the wafer after the CMP process, such as the thickness of the trench structure, the thickness of the non-trench structure, and the thickness of the metal material after CMP. In short, it is a physical model for predicting the surface topography of the wafer after the CMP process. CMP is one of the eight major processes in the entire semiconductor industry, and the other processes after CMP will be affected by the CMP process. That is, when evaluating and controlling the process results after CMP, it is better to know the impact of the CMP process, which is of great significance to the overall semiconductor manufacturing process control and yield improvement. In addition, the CMP process itself is an indispensable step, because through the CMP model, the CMP process can be better adjusted, for example, the results of the CMP model can be used to infer which process parameters need to be modified and adjusted, and whether the design of the graphic data system (GDS) has some unreasonable places that may produce hotspots, i.e. defect places.

[0041] As mentioned earlier, the traditional CMP model has high modeling accuracy, but the disadvantage is that the calculation speed is slow. Because the time step model is calculated step by step, the result of each time step depends on the result of the previous time step, and cannot be accelerated in parallel with a graphics processing unit (GPU) or CPU. In the process of finding the optimal solution of the parameters, the running time of each attempt cannot be reduced, so building a good CMP model requires repeated adjustment and running, which consumes a lot of time and effort. In the full chip simulation stage, if a large size chip is encountered, the data volume will increase significantly, and the CMP physical model based on time step will be slow to simulate, usually taking several hours.

[0042] A CMP time step physical model is known, which is mixed by several physical and mathematical formulas. A core feature of the model is that a polishing time of each step needs to be set according to an actual CMP process, which is assumed to be t1, t2, t3, … and a unit time dt is set, where dt is usually a relatively small number, and the smaller the CMP physical model is closer to the actual CMP physical process, because the core assumption of the CMP physical model is that the polishing rate of the substrate to the CMP material remains nearly constant at a very short time dt, and when dt approaches 0, the polishing rate at this time approaches a constant value (the idea of differentiation). Therefore, if a more accurate physical model needs to be established, dt needs to be set very small, but too small dt will cause a sharp increase in calculation amount. Assuming that there are several time steps in total, each step lasts for t1, t2, …, the total time is the sum of them, which is assumed to be T, then a complete CMP calculation needs to calculate the result of T / dt time steps, usually T will be hundreds of seconds (the real physical process is so), if dt is set to 0.001s, then tens of thousands of time steps need to be calculated. In the calculation of each time step, usually not a point is involved in the operation, and a doe test needs to be performed in the modeling stage, the doe test algorithm generates Q test points for R iterations, the larger the number of Q and R is set, the closer the final found model parameters are to the global optimal solution, and in the simulation stage, it depends on the size of the full chip, and when it is large, there will be tens of millions of points that need to be simulated and calculated. A built model usually needs to be used several times, and many full chips may need to use the model for simulation. The above all shows the calculation complexity of the current CMP physical model, so it is very necessary to accelerate.

[0043] Therefore, the present disclosure provides an improved solution.

[0044] Some embodiments of the present disclosure provide an improved method for training a neural network model. The method comprises: training the neural network model using training samples to determine a gradient of a loss function of the neural network model, wherein the training samples are generated by a CMP physical model based on a pattern feature used to determine a surface height of a polished wafer and physical information; updating parameters of the neural network model based on the gradient of the loss function; and in response to the training satisfying a predetermined condition, determining the neural network model corresponding to the respective updated parameters as a trained neural network model, wherein the trained neural network model outputs a predicted value of the surface height.

[0045] Some embodiments of the present disclosure also provide an improved simulation model and a neural network model.

[0046] Embodiments of the present disclosure can significantly accelerate the establishment and simulation process of CMP physical models, reduce the relevant time consumption, and improve the simulation accuracy.

[0047] Embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0048] Figure 1 A schematic diagram of an example environment 100 in which embodiments of the present disclosure can be implemented is shown. As shown, the example environment 100 includes a computing device 110 and a client 120. Figure 1

[0049] In some embodiments, the computing device 110 can interact with the client 120. For example, the computing device 110 can receive an input message from the client 120 and output a feedback message to the client 120. In some embodiments, the input message from the client 120 can be design layout data. The computing device 110 can perform corresponding mathematical operations on the design layout data and output the corresponding operation results to the client 120.

[0050] In some embodiments, the computing device 110 can include, but is not limited to, a personal computer, a server computer, a handheld or laptop device, a mobile device (such as a mobile phone, a personal digital assistant PDA, a media player, etc.), a consumer electronic product, a small computer, a large computer, a cloud computing resource, etc.

[0051] It should be understood that the structure and function of the example environment 100 are described only for illustrative purposes and are not intended to limit the scope of the subject matter described herein. The subject matter described herein can be implemented in different structures and / or functions. The environment is only schematic and is not intended to limit the application environment of the embodiments of the present disclosure.

[0052] To more clearly explain the principles of the present disclosure scheme, the following will be described in more detail with reference to Figures 2-5

[0053] First, the CMP physical model is introduced. The essence of the CMP physical model is composed of a number of physical formulas, rather than a black box that cannot be described in mathematical language. As is known, a deep learning model has extremely strong fitting ability and can even fit a black box function relationship that cannot be described by an explicit mathematical formula. Therefore, a suitable deep learning model can be established based on deep learning to learn and approximate the CMP physical model.

[0054] ​​Next, how to build the deep learning model is specifically described. The CMP physical model is usually composed of three steps, because the real CMP process can often be divided into three stages, that is: bulk stage, which mainly grinds the uppermost metal (usually copper); touch down (tchd) stage, which is defined as exposing the buffer layer between the metal layer and the bottommost oxide (oxide); barrier removal (br) stage, which is defined as grinding the buffer layer during this period, most of the points touch the bottommost oxide material. There are different physical parameters in each stage.

[0055] Bulk process usually refers to grinding the uppermost metal material from the beginning of grinding to the end of the bulk process stage. In this stage, only metal is ground. CMP will perform a deposition process before grinding, depositing different materials, usually the uppermost layer is metal (usually copper), the middle layer is buffer layer, and the lowermost layer is oxide. The touch here is the buffer layer material, that is, in this stage, it involves from grinding only metal to starting to touch the buffer layer material. About barrier removal, as the name implies, in this layer, the main task is to remove the buffer layer. In an ideal case, the bulk process only grinds metal. The tchd stage will end when the buffer layer is touched. The br stage can completely grind the buffer layer material. But in reality, it is not so ideal. The bulk process stage may also have some areas where the material has changed from metal to buffer layer, and tchd is also impossible to control all points to touch the buffer layer material at the same time, and br is even more so. It is difficult to ensure that each point perfectly grinds off the buffer layer and does not touch the lowermost oxide.

[0056] In some embodiments, based on the real CMP process information mentioned above, for the data input into the deep learning neural network, the following information can be included: vector of Graphic Data System (GDS) features, specifically density vector, space vector, width vector, etc.; physical information vector, which can specifically include deposition material information, specifically including deposition material thickness information, nominal polishing rate information of the deposition material, maximum Dishing information of each step, minimum Dishing information, slurry information, long range effect parameter information, duration information, nominal pressure information, etc. Parameters used by the CMP physical model. Wherein Dishing refers to the difference between the height of the trench structure and the height of the non-trench structure, Erosion refers to the difference between the height of the non-trench structure of each grid point and the reference value thickness. Long range effect can be understood as the density width space etc. of the current area not only affects the polishing result of the current area, but also affects the polishing result of the surrounding area.

[0057] The input of the neural network model can be determined based on the above information. Further details are described below Figure 4 The architecture of the neural network model is introduced.

[0058] For the neural network model, training is needed through training samples. The generation method of the training sample is described below.

[0059] First refer to Figure 2 , Figure 2 The flow chart of the generation method 200 of the training sample according to some embodiments of the present disclosure is shown.

[0060] At block 202, a first number of feature vectors are generated based on data points representing various graphic features on the wafer, and the feature vectors are combinations of data representing different types of graphic features in the data points.

[0061] In some embodiments, generating the first number of feature vectors based on the data points representing the respective pattern features on the wafer can include determining data within respective step ranges based on value ranges of the data points of the respective pattern features and respective step sizes, and performing data orthogonalization on the data within the respective step ranges to generate a feature data table, wherein the feature data table includes data of the first number of feature vectors. In some embodiments, the pattern features include at least a density of patterns within a grid on the wafer, a width of the patterns, and a space between the patterns. For example, the pattern features density, width, and space have significant impact on the post-CMP polishing height, and they are significant independent variables for the CMP polishing height prediction model. Each of the three variables has a respective range. In addition, the step sizes can be divided within the ranges as needed. Details are described later.

[0062] The data points represent data of measurement points (grids) on the wafer. The data points can be provided by a user. For example, one data point represents the features density, width, and space, such as (density, width, space), where, for example, density is 0.5, width is 30 nm, and space is 40 nm. Each of density, width, and space represents a pattern feature, i.e., density, width, and space. The combination of data representing these features can constitute a vector. For example, for density, width, and space, each of the value combinations can be referred to as a vector. For example, the aforementioned (0.5, 30 nm, 40 nm). Different combinations of values of these variables can constitute multiple vectors. The combinations can be made in various ways.

[0063] One way is the Design of Experiments (DOE) verification method. DOE design, i.e., DOE data orthogonalization, is an experimental design method used to explore and verify the impact of factors on results. In DOE, experiments are usually divided into multiple combinations, each combination controls one factor, and measures its impact on the results. In this way, the impact of factors on results can be more fully understood, and the best factor combination can be determined.

[0064] In some embodiments, the DOE design can be divided into two parts, one part is the GDS feature design, i.e., the design of density, width, and space mentioned above.

[0065] The ranges and steps of density / width / space can be set first. Since all the three variables have physical meanings, there are relatively fixed ranges in most processes. The common ranges of the three variables can be counted, for example, it is known that density is a number between 0 and 1, width and space are all positive numbers, and usually a number between 0 and 50 (unit: um), and usually the maximum value does not exceed a few hundred. Thus, the data is actually generated according to the actual situation (for example, training resources, whether there are a large number of GPUs to speed up training or not).

[0066] Suppose there are k data points, and the densities are 0.1, 0.2, 0.3, …, d_k respectively, then the density vector is [0.1, 0.2, 0.3, …, d_k]. Similarly, the other vectors can be obtained.

[0067] The DOE mentioned above is to make the neural network training sufficient so as to more accurately track the CMP physical model. Specifically, the orthogonal DOE data generation can be performed on density / width / space. Orthogonal means that, assuming that density generates N possible values, width generates M, and space generates P, then the final data set contains N*M*P data entries. The characteristic of the data in it is that, for example, randomly checking the generated data set, there are several entries for a certain value of width, and the number of entries is of course N*P. In fact, the essence is an exhaustive combination, which ensures that each width has all possible density and space values. By orthogonalizing the data, the comprehensiveness of the data and the reliability of the results can be ensured.

[0068] At block 204, a second number of physical vectors are generated based on physical information related to CMP of the wafer, the physical vectors being combinations of data representing different types of physical information.

[0069] In some embodiments, the physical information at least includes the following: information of deposited material; maximum and minimum disc defect information of each step; slurry information; long tail effect parameter information; duration information; nominal pressure information.

[0070] In some embodiments, generating the second number of physical vectors based on the physical information related to the CMP of the wafer can include: determining data in each step range based on the value range of the data points of each physical vector and the corresponding step length; and data orthogonalizing the data in each step range to generate a physical information data table, wherein the physical information data table includes data of the second number of physical vectors.

[0071] In some embodiments, the physical information can be unfolded into vectors with equal length and density, space, width. In other words, the density, space, width have equal length, and the physical information can be unfolded into vectors with equal length as the density, space, width. The vectors can be combined to form a matrix, which is the input, denoted as X. The X can be used as the input of the neural network model.

[0072] Suppose the generated GDS feature data table has a length of k. Because the physical parameters are also input variables, the DOE combination table is also generated by DOE design according to the actual process conditions, for example, the pressure can be selected as small pressure, medium pressure, large pressure, and the like. Suppose the total number of physical parameter DOEs is L, then the CMP physical simulation model needs to be run L times.

[0073] At block 206, a second number of simulations are performed by the CMP physical model based on the feature vectors and the physical vectors to generate, as training samples, simulation values of the surface height of the wafer corresponding to the first number of feature vectors in each simulation.

[0074] In some embodiments, the corresponding number of simulations is performed based on the number of physical vectors.

[0075] In some embodiments, the designed GDS feature data and one of the L physical model combination tables are input each time, and then L result tables with a length of k (k is the number of data points) are generated. Because the multiple tables are independent of each other, they can be generated in parallel, greatly speeding up the generation. Specifically, suppose there are several features participating in the orthogonal, and an orthogonal table is obtained after orthogonalization. Each row of data of the orthogonal table contains the GDS features (density / space / width) and process parameters, which are input into the physical model for operation to obtain the operation results, that is, the corresponding SThickNT and SThickT. Repeat several times until each row of data generates a corresponding SThickNT and SThickT.

[0076] In some embodiments, to facilitate processing in the neural network, the data can be merged (or concatenated) by an Embeding operation, or called mapping operation. The specific implementation is as follows: assuming that the process parameter a takes the value of 0.5, then the vector of the process parameter a is [0.5, 0.5, …] (actually a copy). The vector with a length equal to density, the process vector b is similar, so each process vector can be obtained. Similar to the previous operation of concatenating density, width, these process vectors can also be concatenated behind, thus forming a final input matrix X. Assuming that density, space, width are used as GDS features, and process parameters a, b, c are used as process features, a total of k data points, then the final input matrix is a k row 6 column matrix, and since there are 3 GDS features and 3 process parameters, there are a total of 6 columns. Concatenating the data is more convenient from the engineering implementation, and in addition, since the input matrix will be multiplied by the parameter matrix of the neural network, it is easy to integrate the process information into the input, which is efficient.

[0077] In some embodiments, the second number of simulations by the CMP physical model based on the feature vector and the physical vector can include: generating a first simulation value of a surface height of the non-trench structure and a second simulation value of a surface height of the trench structure, respectively; and determining a simulation value of a wafer surface height based on the first simulation value and the second simulation value.

[0078] It should be understood that embodiments of the present disclosure are not limited thereto, for example, in some embodiments, the vectors can not be concatenated, but can be input into the neural network model as one-dimensional vectors.

[0079] It should be noted that the DOE orthogonal table only has input (input of the neural network), that is, there is no S-related column, and the S-related column refers to the label data column necessary for training the neural network, so the DOE orthogonal table needs to be input into the physical model for simulation first, and the simulation result of the physical model is obtained, so that the simulation result of the physical model becomes the label data column. The function of the DOE orthogonal table (after the label data column is generated by the simulation of the physical model) is to provide neural network training and divide a part as a validation data set to verify whether the training is in place.

[0080] In some embodiments, the result table can be: SThickNT / SThickT / SDishing / SErosion, where SThickNT and SThickT are outputs of the neural network, and the difference between the two is dishing; SErosion is the reference surface height (a certain constant) minus SThickNT.

[0081] In addition, the aforementioned parallel acceleration can have two scenarios in the present case.

[0082] One scenario is that a trained neural network is used to replace the physical model because the neural network can be accelerated in parallel. For example, there are k data points, so the vector length is k, and suppose that a genetic algorithm (GA) is to be iterated 100 times with a population size of 1000, which means that the simulation results of k data points need to be generated 1000 times for each iteration. Therefore, 1000 k-length results need to be generated for each round, and because the 1000 populations are independent of each other, GPU can be used for parallel acceleration. Because of the complexity of the CMP physical model, it is difficult to deploy it on a GPU, and at most, it can be implemented by calling the computer's ordinary multi-threading, for example, using 20 threads to accelerate, that is, the 20 threads are used to calculate the 1000 k-length results (equivalent to each thread completing 50, and everyone together completes it). However, the neural network is naturally suitable for GPU and is very easy to implement. The acceleration capability of GPU is not comparable to ordinary multi-process. GPU can implement tens of thousands of threads or even more, while ordinary multi-threading depends on the ability and size of CPU, but CPU itself cannot have many threads. In addition, when generating k-length results, because the neural network is a one-step result, k data points can be independently predicted, but the CMP physical model needs to consider environmental information and cannot be independently operated. The neural network model can be directly and completely parallel to obtain the calculation result by parallel deployment (on GPU).

[0083] Another scenario is to generate samples for training a neural network. At this time, because the samples are independent of each other, many machines can be used to generate results, and then the results are collected for training. For example, 10,000 samples are needed to train. One way is to generate one sample at a time, and another way is to start 100 servers, each of which generates part of the results, and then collects them (a total of 10,000) for training.

[0084] The following embodiments will be described in conjunction with Figure 4 . Figure 4A schematic diagram of a deep learning model according to some embodiments of the present disclosure is shown, specifically a schematic diagram of a neural network architecture 400. The architecture includes: an input layer 402; the input layer includes a physical parameter information table and a GDS feature vector (may be referred to as a graph vector), which includes a Width vector, a Space vector, and a Density vector. An embedding layer 404; usually the physical information parameters are scalars, which are mapped to vectors of the same length as density / width / space through embedding, and then these vectors can be spliced together to form an input matrix X. A first convolutional layer 406 connected to the embedding layer 404; a pooling layer 408 connected to the first convolutional layer 406; a second convolutional layer 410 connected after the pooling layer; a first fully connected layer 412 connected after the second convolutional layer 410; a first regularization layer 414 connected after the first fully connected layer 412; a second fully connected layer 418 connected after the first regularization layer 414; a second regularization layer 420 connected after the second fully connected layer 418; a third fully connected layer 422 connected after the second regularization layer 420; the output layer of the third fully connected layer 422 is connected to two fully connected layers, i.e., a fourth fully connected layer 424 and a fifth fully connected layer 426. The last two fully connected layers are each connected to an output layer, i.e., an SThickNT output layer 430 and an SThickT output layer 428. Among them, the SThickNT output layer 430 outputs the simulation value of the surface height of the non-trench structure, and the SThickT output layer 428 outputs the simulation value of the surface height of the trench structure. The simulation value of the surface height of the wafer can be determined based on the simulation value of the surface height of the non-trench structure and the simulation value of the surface height of the trench structure.

[0085] In some embodiments, the simulation value of the height of the non-trench structure on the wafer H1 can be determined; the simulation value of the height of the trench structure on the wafer H2 can be determined; and the simulation value of the surface height of the wafer can be determined based on the simulation value H1, the simulation value H2, and the density of the graph in the grid region on the wafer. Specifically, the surface height is equal to the height of the non-trench (Nontrench) structure multiplied by (1-density) plus the product of the height of the trench (Trench) structure and density. In formula: H2*density+H1*(1-density); where density represents the density of the graph in the grid region on the wafer, which ranges between 0-1.

[0086] As Figure 4As shown, this architecture also includes the ReLU activation function 416. The ReLU activation function, or Rectified Linear Unit activation function, is one of the most commonly used activation functions in Convolutional Neural Networks (CNNs) and many deep learning models. Its main function is to introduce non-linearity, enabling the model to learn and express more complex features.

[0087] As mentioned earlier, multiple result tables of length k were generated. Furthermore, after determining the training samples and constructing the neural network model, the model can be trained. The following section combines... Figure 3 The specific method for training a neural network model is described. Figure 3 A flowchart of a training method 300 for a neural network model according to some embodiments of the present disclosure is shown.

[0088] At box 302, the neural network model is trained using training samples to determine the gradient of the loss function of the neural network model, wherein the training samples are generated by the CMP physical model based on graphical features and physical information used to determine the surface height of the wafer after grinding.

[0089] In some embodiments, training samples may be pre-generated, for example, stored in a predetermined database, and may be used to train neural network models.

[0090] In some embodiments, training samples may be generated during the training of a neural network model.

[0091] In some embodiments, training a neural network model using training samples to determine the gradient of the loss function of the neural network model includes: calculating the gradient of the loss function using a stochastic gradient descent algorithm.

[0092] In some embodiments, updating the parameters of a neural network model based on the gradient of the loss function includes: updating the gradient of the loss function after a predetermined number of training iterations (e.g., 100 iterations); and updating the parameters of the neural network model based on the gradient update.

[0093] In some embodiments, training the neural network model with the training samples to determine the gradient of the loss function of the neural network model can include: training the neural network model with a part of the training samples to perform gradient descent back propagation to update the parameters; predicting the remaining part of the training samples with the trained neural network model to generate predicted values of the surface height of the ground wafer after grinding; in the case that the comparison result of the predicted values and the simulation values of the CMP physical model meets a predetermined condition, the corresponding trained neural network model can be determined as the final neural network model. That is, in this way, the training degree of the neural network can be determined. For example, assuming that the DOE orthogonal generates 10 million data, and the simulation results of the 10 million data are simulated by the physical model. Then take out 70% (also can be 80% 90%) of them as training samples for training. When 7 million training samples are each input into the neural network to update the parameters by gradient descent back propagation, the training is completed. Next, the trained neural network is used to predict the just remaining 30% of the data, and then the rmse of the neural network prediction value and the physical model simulation value is calculated. If the rmse meets a certain standard, it is considered that the training is in place. If it does not meet the standard, the DOE can be made more detailed, and some more data can be generated for further training, and the above steps can be repeated.

[0094] For the neural network, in many cases, the "input" and the "training sample" are essentially the same. The training sample refers to some data, which allows the neural network to calculate, and then updates the neural network parameters according to the calculation error. Compared with the training sample, the data set to be predicted (input) has no real corresponding value, and cannot allow the neural network to update and adjust its parameters. However, the essence is the matrix X mentioned above. In addition, the concept of "input" is wider than that of "training sample". If an input has feedback, it is a training sample, because it can be used to adjust the neural network parameters, which is also "training". If an input has no feedback, it is a data set to be predicted, because it cannot know whether the prediction result is good or bad, and cannot know whether the result of the neural network is accurate or not. Therefore, it cannot be used for training. In general, the input describes a matrix or vector that meets the requirements of the input layer of the neural network, and the training sample refers to a set of data points, each of which has a standard reference value that can be used to evaluate the goodness of the neural network prediction. The part of the data set without standard answer can be extracted as input, input into the neural network, and the neural network gives the corresponding output. Then, by comparing the standard answer and the output, the parameters are updated by back propagation. Therefore, the input can also be said to be part of the training sample.

[0095] As mentioned earlier, the physical information is obtained, and the matrix can be generated based on the physical information and the graph feature information as the input of the neural network model.

[0096] In some embodiments, the DOE is used to generate an orthogonal table, and a CMP physical model is used to simulate to obtain training samples, which contain information required by the neural network (features of GDS and results of process parameters) and simulation results given by the physical model. As mentioned above, the GDS features and process parameters can be extracted as an input X matrix and input into the neural network, and the neural network will obtain the calculation results. Then, by comparing the neural network results and the simulation results of the physical model, the loss function is calculated and back-propagated to update the network parameters, and finally a trained neural network is obtained. For example, there is a data point containing GDS features and process information, which can be input into the neural network and the physical model. The output results of the neural network are very close to the output results of the physical model, so the training is effective, and a neural network learned from the physical model is obtained.

[0097] Back to Figure 2 Continuing to describe, in some embodiments, training the neural network model with the training samples to determine the gradient of the loss function of the neural network model includes that the gradient of the loss function can be calculated using a stochastic gradient descent algorithm.

[0098] In some embodiments, the gradient of the loss function can be calculated using a stochastic gradient descent algorithm.

[0099] In some embodiments, training the neural network model with the training samples to determine the gradient of the loss function of the neural network model can include training the neural network model with a portion of the training samples to perform gradient descent back-propagation to update the parameters, and predicting the remaining portion of the training samples using the trained neural network model to generate predicted values of the surface height of the polished wafer.

[0100] In fact, the essence of the training of the neural network is to first simulate the budget using the neural network, then compare the label data, calculate the loss of the objective function, perform gradient descent, and update the parameters by back-propagation.

[0101] At block 304, the parameters of the neural network model can be updated based on the gradient of the loss function.

[0102] In some embodiments, updating the parameters of the neural network model based on the gradient of the loss function can include performing gradient updates on the gradient of the loss function for a predetermined number of times of training, such as 100 times, and updating the parameters of the neural network model based on the gradient updates. It should be understood that the number of times of training can be changed according to actual needs, and embodiments of the present disclosure are not limited in this regard.

[0103] In some embodiments, the gradient is calculated using a stochastic gradient descent algorithm and the neural network parameters are constantly updated.

[0104] At block 306, in response to the training satisfying the predetermined condition, the neural network model corresponding to the respective updated parameters is determined as a trained neural network model, wherein the trained neural network model outputs a predicted value of the surface height.

[0105] In some embodiments, the method further comprises: determining a loss function value of the neural network model; and wherein the training satisfying the predetermined condition comprises: a difference between the loss function values of two consecutive iterations is less than a predetermined difference threshold; or an absolute value of a ratio of the loss function values of two consecutive iterations is greater than a predetermined ratio threshold.

[0106] In some embodiments, the gradient is calculated using a stochastic gradient descent algorithm and the neural network parameters are updated iteratively until the loss curve approaches a plateau, at which point the training can be stopped.

[0107] In some embodiments, as mentioned previously, training the neural network model using the training samples to determine the gradient of the loss function of the neural network model can comprise: training the neural network model using a portion of the training samples to perform gradient descent backpropagation to update the parameters; and using the trained neural network model to make predictions on the remaining portion of the training samples to generate predicted values of the surface height of the polished wafer. In the case where a comparison of the predicted values and simulated values of the CMP physics model satisfies a predetermined condition, the respective trained neural network model can be determined as a final neural network model. The comparison of the predicted values and simulated values satisfying the predetermined condition can comprise: a root mean square error of the predicted values and simulated values being less than a predetermined threshold.

[0108] In some embodiments, the method further comprises: determining a loss function value of the neural network model; and wherein the training satisfying the predetermined condition comprises: a difference between the loss function values of two consecutive iterations is less than a predetermined difference threshold; or an absolute value of a ratio of the loss function values of two consecutive iterations is greater than a predetermined ratio threshold.

[0109] In some embodiments, the training samples are generated in parallel on a plurality of servers such that the training of the neural network model is performed in parallel.

[0110] In some embodiments, a trained neural network model can be generated according to the training method of the neural network described above.

[0111] In some embodiments, a neural network model is also provided, comprising: an input layer configured to receive wafer feature data and CMP-related physical information of a wafer; an intermediate layer configured to perform the simulation method of the semiconductor device described above to determine a surface height value of the polished wafer; and an output layer configured to output the surface height value.

[0112] In some embodiments, the neural network model further comprises an embedding layer configured to map the input physical information to generate a physical vector.

[0113] The above embodiments provide a trained deep learning neural network model for tracking CMP physical model, and how to use it is described later. In some embodiments, the neural network model can be used together with the CMP model.

[0114] In some embodiments, a simulation model is also provided, which comprises: a CMP physical model configured to determine a simulation value of the surface height of the polished wafer based on input data, wherein the input data comprises the pattern feature data and the physical information; and the neural network model as described above, which is trained with the simulation value as a training sample, and is configured to perform parallel prediction on multiple groups of input data in the simulation process of the CMP physical model to generate multiple prediction values of the surface height of the polished wafer corresponding to the multiple groups of input data, respectively as the corresponding simulation values of the surface height.

[0115] In some embodiments, the parallel prediction on multiple groups of input data in the simulation process of the CMP physical model comprises: using multiple graphic processors to perform parallel acceleration on the calculation process of the neural network model.

[0116] In some embodiments, a simulation method is also provided, which uses the above-mentioned simulation model to perform, and the method comprises: predicting, by the neural network model in the simulation model, a simulation value of the surface height of the polished wafer based on input data, wherein the input data comprises the pattern feature data and the physical information.

[0117] In some embodiments, predicting, by the neural network model in the simulation model, a simulation value of the surface height of the polished wafer based on input data can comprise: the neural network model performing parallel prediction on multiple groups of input data to generate multiple prediction values of the surface height of the polished wafer corresponding to the multiple groups of input data, respectively as the corresponding simulation values of the surface height.

[0118] In some embodiments, predicting, by the neural network model in the simulation model, a simulation value of the surface height of the polished wafer based on input data comprises: generating, by the neural network model, a first prediction value of the surface height of the non-trench structure and a second prediction value of the surface height of the trench structure, respectively; and determining, by the neural network model, the simulation value of the wafer surface height based on the first prediction value and the second prediction value.

[0119] In some embodiments, the method further comprises determining the condition of the surface of the wafer during the polishing process by a CMP physical model in the simulation model. Users eventually also need a physical model because sometimes they need to output the intermediate process to help them confirm some relevant details of the process. For example, if a manufacturer needs to change the process parameter settings of their substrate, they are likely to need to know the polishing conditions at some intermediate time to assist them in adjusting the parameters. Neural networks cannot give the intermediate process, and therefore the physical model can meet this need.

[0120] According to the above description, it can be known that the embodiments of the present disclosure also provide a simulation method. The method can be implemented by the simulation model described above.

[0121] In some embodiments, the simulation method can be implemented by the CMP physical model according to the DOE design principle to obtain the simulation results of SThickNT / SThickT / SDishing / SErosion of the CMP physical model.

[0122] In some embodiments, the simulation method comprises predicting, by a neural network model in the simulation model, a simulation value of the surface height of the polished wafer based on input data, wherein the input data comprises the graphic feature data and the physical information.

[0123] As can be known from the above, the current CMP physical model will have a large time and computing resource loss during training, mainly because when running a parameter combination of a certain test, a very large number of calculations are needed due to the characteristics of the time step model. The embodiments of the present disclosure provide a neural network model that has learned the complex physical relationship. The model obtains the same input as the CMP physical model, but the speed of giving the output result will be extremely fast. In fact, the time for the CMP physical model to run a test is likely to be tens of seconds, while the time for the neural network model to give the output is only zero point zero several seconds, so the operation speed and efficiency will be greatly improved.

[0124] Similarly, for the CMP physical model, its fitting takes GA genetic algorithm. GA genetic algorithm needs to set the population size of the experiment (pop size) and the number of iterations (max iter) two parameters. In the case of a single long calculation time, it cannot be set too large, otherwise the total time will be very long. GA algorithm is used to solve the parameters in CMP physical model, GA algorithm needs to constantly experiment, each experiment needs the simulation results of many experimental individuals given by CMP physical model. The iteration of GA algorithm only needs a target function, which is set as the square error between the measured data points (real label data) and the simulation results given by the simulation model. The calculation of GA algorithm is large (if you want to find the right solution), if you want to get a better solution, GA algorithm needs to get a very large number of test results. Test results refer to the simulation value of all real data points given by the CMP physical model with a given parameter combination, and then compare the difference between the simulation value given by the model and the real label value, which is a test result.

[0125] Parameter combination refers to the parameter combination generated by the undetermined parameters in the CMP physical model. For example, the CMP physical model sets 10 undetermined parameters, which are parameter combinations. Because each parameter has many possible values, the number of combinations is very large. Each additional undetermined parameter, the total possible situation of the parameter combination is exponentially increasing. The total search times of GA algorithm is the population size multiplied by the number of iterations. CMP physical model evaluation takes 10s, and deep learning model evaluation takes only 0.1s. It can be seen that it is very necessary to use deep learning to speed up.

[0126] The acceleration of the solution of the physical model parameters has been described before. Because given a data set, the simulation result of each point of the data set given by the physical model is much slower than the neural network, and in the iteration process of GA algorithm, the simulation result needs to be repeatedly simulated. If only one simulation is calculated, the difference between 10s and 0.1s is not big, if 10,000 simulations are calculated, the difference is very large. The acceleration of the simulation is the same. For a given data set, the neural network calculation is much faster than the physical model. This is not a problem. In addition, as the size of the data set increases, the advantage of the neural network becomes more and more obvious. For example, with 100 data points, the physical model needs 3s, and the neural network needs less than 0.1s. If you need to simulate 10 million data points, the physical model may need (10 million / 100)*3 seconds, and the calculation time of the neural network is not increased in this way, which may only need 2s.

[0127] In the whole process of using the CMP tool, as mentioned above, there are mainly two scenarios, one is parameter solving, and the other is global simulation. When the parameter solving is performed, the user provides tens of real data points collected to solve the parameters of the physical model, and when the global simulation is performed, the user provides tens of millions of data points to calculate the simulation value of each data point through simulation.

[0128] A set of parameter solutions can be found by minimizing the objective function, so that the result given by the simulation model is as close to the real value as possible. If deep learning acceleration is not introduced, the objective function here is the squared error of the simulation result of the physical simulation model and the real result, and if it is introduced, it becomes the squared error of the prediction result of the deep learning network and the real result.

[0129] The physical model is a large model (very complex, time-consuming to calculate, and large in calculation resources and time to solve the optimal parameters, but more accurate, and can output the simulation value of the intermediate process of the CMP. The neural network model has very fast operation speed, and because the serial process is abandoned, the calculation of the neural network can be completely accelerated in parallel. In simple terms, even without parallel acceleration (using GPU), the pure neural network prediction calculation speed is much faster than the physical model. Now, because of the characteristics of the calculation of the two, the neural network can be further accelerated by GPU parallel acceleration, which can greatly improve the speed, so the last model used for approximate replacement can greatly improve the parameter solving efficiency.

[0130] In the embodiments of the present disclosure, the single operation time is greatly reduced through the approximate replacement of the neural network model, so a very large population size and iteration number can be set to find the optimal parameter combination, and the total time does not increase compared to before, which greatly improves the probability of finding a better solution.

[0131] For simulation, the above embodiments provide a neural network model that can respond extremely fast, and it is a model that can be easily accelerated by using hardware such as GPU (compared to the CMP physical model), so even if the amount of data of the whole chip that needs to be simulated is huge, even if it is several tens of millions of rows, the original CMP physical model needs several hours, and the neural network model of the embodiments of the present disclosure only needs a few minutes.

[0132] It is known that deep learning neural network model has very strong fitting ability. However, one feature of deep learning neural network is that it requires a very large amount of data. Therefore, the defect is that the amount of training data is required to be large, and in many actual cases, there is not so much labeled data for training. In the CMP process, the manufacturer cannot provide the data level required for training a neural network due to the limitation of actual situation (for example, cost), and can usually only provide 50-100 data points, while the neural network requires tens of thousands to hundreds of thousands of data points, which is obviously impossible. In addition, as mentioned earlier, the CMP physical model is a serial calculation model, that is, the result at each time still depends on the result of the previous step, and this calculation feature leads to non-parallelism.

[0133] However, in the present application, instead of directly fitting the real CMP process with a neural network model and fitting with real measurement data (the real measurement points are usually very few), the neural network is used to track and imitate the CMP physical model. In the known CMP physical model, theoretically, it can generate labeled data infinitely, so the problem of training the neural network model is solved. In addition, in the present application, the neural network is not directly used to replace the CMP physical model, but only the feature of extremely fast calculation speed of the neural network is used to replace the return result requirement of the target function in the GA algorithm. Instead of directly using the CMP physical model to return the test results required by the GA algorithm, the trained neural network model is used to give the test results. The final purpose has not changed, which is still to obtain a more optimal parameter combination for the CMP physical model fitting. However, how to achieve this purpose, the present application ingeniously uses the neural network to accelerate.

[0134] Some embodiments of the present disclosure provide a method for generating training samples, a method for training a neural network, and a method for simulating a semiconductor device. It should be noted that the examples in the above embodiments are only for illustrating the schemes of the embodiments of the present disclosure, and are not used to limit the schemes of the present disclosure.

[0135] In some embodiments of the present disclosure, the neural network is ingeniously introduced to accelerate the parameter solving and simulation speed of the physical model by creatively using the idea of approximation substitution. The embodiments of the present disclosure can take advantage of the neural network, and are completely theoretically and practically feasible.

[0136] The present disclosure creatively introduces a deep learning / machine learning model to greatly accelerate the establishment and simulation process of the CMP physical model, greatly reduces the related time and effort, and has very high practicality in actual use.

[0137] The deep learning / machine learning model introduced in the present disclosure does not change the original precision and result of the CMP physical model loss, but greatly helps to establish and simulate it through appropriate use, and even further improves its precision. The present disclosure greatly reduces the simulation and establishment of the CMP physical model by using the deep learning model, reducing the time cost. The clever use of the deep learning model in the present disclosure can help improve the precision of the CMP physical model to a certain extent, and help find a more suitable model parameter combination.

[0138] It should be understood that the embodiments shown in the drawings are only for illustrative purposes to show the schemes of some embodiments of the present disclosure, and are not intended to limit the present disclosure. The embodiments of the present disclosure can also have various other forms.

[0139] The present disclosure also discloses an electronic device in an embodiment of the present disclosure. The electronic device includes a processor and a memory coupled with the processor, the memory having stored therein instructions which, when executed by the processor, cause the device to perform actions including training a neural network model with training samples to determine a gradient of a loss function of the neural network model, wherein the training samples are generated by a CMP physical model based on a pattern feature used to determine a surface height of a polished wafer and physical information; updating parameters of the neural network model based on the gradient of the loss function; and in response to the training satisfying a predetermined condition, determining the neural network model corresponding to the respective updated parameters as a trained neural network model, wherein the trained neural network model outputs an enhanced formula for determining the surface height.

[0140] The present disclosure also discloses a computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the method of the foregoing embodiments of the present disclosure.

[0141] Figure 5 A schematic block diagram of an electronic device in accordance with some example embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown in the electronic device, its connections and relationships, and its functions, are meant to be examples only, and are not intended to limit the implementations of the present disclosure described and / or claimed in this document.

[0142] As Figure 5As shown, the apparatus 500 includes a CPU 501 that can perform various appropriate actions and processes in accordance with a computer program stored in a read only memory (ROM) 502 or a computer program loaded from a storage unit 508 into a random access memory (RAM) 503. In the RAM 503, various programs and data required for the operation of the apparatus 500 can also be stored. The CPU 501, the ROM 502, and the RAM 503 are connected to each other through a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0143] A plurality of components in the apparatus 500, including an input unit 506, such as a keyboard, a mouse, and the like, an output unit 507, such as various types of displays, speakers, and the like, a storage unit 508, such as a magnetic disk, a magneto-optical disk, and the like, and a communication unit 509, such as a network card, a modem, a wireless communication transceiver, and the like, are connected to the I / O interface 505. The communication unit 509 allows the apparatus 500 to exchange information / data with other apparatuses through a computer network, such as the Internet, and / or various telecommunication networks.

[0144] The various processes and processes described above, such as the methods 200, 300, can be performed by the CPU 501. For example, in some embodiments, the methods 200, 300 can be implemented as a computer software program tangibly embodied in a machine readable medium, such as the storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on the apparatus 500 via the ROM 502 and / or the communication unit 509. When the computer program is loaded into the RAM 503 and executed by the CPU 501, one or more steps of the methods 200, 300 described above can be performed.

[0145] The schemes according to embodiments of the present disclosure can be method, apparatus, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for performing various aspects of the present disclosure. The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable program instructions can be downloaded to various computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network.

[0146] Embodiments of the present disclosure have been described above, the above description is exemplary only, is only optional embodiments of the present disclosure, is not exhaustive, and is not used to limit the present disclosure. Although the claims in this application have been drafted in relation to a specific combination of features, it should be understood that the scope of the present disclosure also includes any novel feature or any novel combination of features disclosed herein, explicitly or implicitly, or any generalization thereof, regardless of whether it is involved in the same solution as any of the presently claimed claims. It should be understood that new claims can be drafted to these features and / or combinations of these features during the examination of this application or any further application derived therefrom.

[0147] The selection of the terms used herein is intended to best explain the principles of the embodiments, practical application, or technical improvement in the art, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein. The present disclosure can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for training a neural network model, comprising: The neural network model is trained using training samples to determine the gradient of the loss function of the neural network model, wherein the training samples include simulated values ​​of the surface height of the wafer; The parameters of the neural network model are updated based on the gradient of the loss function; and In response to the training satisfying predetermined conditions, a neural network model corresponding to the corresponding updated parameters is determined as a trained neural network model, wherein the trained neural network model outputs a predicted value of the surface height.

2. The method of claim 1, wherein training the neural network model using training samples to determine the gradient of the loss function of the neural network model comprises: The gradient of the loss function is calculated using the stochastic gradient descent algorithm.

3. The method according to claim 2, wherein updating the parameters of the neural network model based on the gradient of the loss function includes: In response to training a predetermined number of times, the gradient of the loss function is updated. as well as The parameters of the neural network model are updated based on the gradient update.

4. The method of claim 1, wherein training the neural network model using training samples to determine the gradient of the loss function of the neural network model comprises: The neural network model is trained using a portion of the training samples to perform gradient descent backpropagation and update the parameters. The trained neural network model is used to predict the remaining portion of the training samples to generate a predicted value for the surface height of the polished wafer. as well as In response to the fact that the comparison between the predicted value and the simulated value of the chemical mechanical polishing physical model meets a predetermined condition, the corresponding trained neural network model is determined as the final neural network model.

5. The method according to claim 4, wherein the comparison result between the predicted value and the simulated value satisfies a predetermined condition, including: The root mean square error between the predicted value and the simulated value is less than a predetermined threshold.

6. The method of claim 1, further comprising: Determine the loss function value of the neural network model; Furthermore, the training satisfies predetermined conditions, including: The difference between the loss function values ​​of two consecutive iterations is less than a predetermined difference threshold; or The absolute value of the ratio of the loss function values ​​of the two iterations is greater than a predetermined ratio threshold.

7. The method of claim 1, wherein the training samples are generated by a chemical mechanical polishing physical model based on graphical features and physical information used to determine the surface height of the wafer after polishing.

8. The method according to claim 1, wherein the training samples further include graphic features and process parameters.

9. The method of claim 8, wherein the graphic features and the process parameters are arranged into an input matrix as input to the neural network model.

10. The method of claim 9, wherein the input matrix is ​​formed by concatenating the following vectors: Feature vectors, wherein the feature vectors are combinations of data representing different types of graphical features; and A process parameter vector of the same length as the feature vector, mapped from the process parameters.

11. The method of claim 1, wherein the training samples are generated in the following manner: A first number of feature vectors are generated based on data points used to represent various graphic features on the wafer, wherein the feature vectors are combinations of data points representing different types of graphic features. A second number of physical vectors are generated based on physical information related to the chemical mechanical polishing of the wafer; these physical vectors are combinations of data representing different types of physical information. The chemical mechanical polishing physical model performs a second number of simulations based on the feature vectors and the physical vectors, so as to generate simulation values ​​of the wafer surface height corresponding to the first number of feature vectors in each simulation as training samples.

12. The method of claim 1, wherein the training samples are generated in the following manner: The orthogonal array generated based on the orthogonal design verification method is used to generate the training samples by simulation of the chemical mechanical polishing physical model, wherein each row of data in the orthogonal array contains graphic features and process parameters.

13. The method according to any one of claims 1 to 12, wherein the training samples are generated in parallel on multiple servers so that the training of the neural network model is performed in parallel.

14. A neural network model generated by the method according to any one of claims 1 to 13.

15. A simulation model, comprising: A chemical mechanical polishing physical model is configured to determine a simulated value of the surface height of a wafer after polishing based on input data, wherein the input data includes graphic feature data and physical information; as well as According to claim 14, the neural network model is generated by training the simulated values ​​as training samples and is configured to perform parallel predictions on multiple sets of input data during the simulation of the chemical mechanical polishing physical model to generate multiple predicted values ​​of the surface height of the polished wafer corresponding to the multiple sets of input data, which are respectively used as the corresponding simulated values ​​of the surface height.

16. The simulation model according to claim 15, wherein parallel prediction of multiple sets of input data during the simulation of the chemical mechanical polishing physical model includes: The computation process of the neural network model is accelerated in parallel by using multiple graphics processors.

17. A simulation method for semiconductor devices, the method comprising: The simulated value of the surface height of the wafer after polishing is determined based on input data using a chemical mechanical polishing physical model, wherein the input data includes graphic feature data and physical information; as well as During the simulation of the chemical mechanical polishing physical model, the neural network model performs parallel predictions on multiple sets of input data to generate multiple predicted values ​​of the surface height of the polished wafer corresponding to the multiple sets of input data, which are respectively used as the corresponding simulated values ​​of the surface height. The neural network model is generated by training with the simulated values ​​determined by the chemical mechanical polishing physical model as training samples.

18. The method of claim 17, wherein the neural network model performs parallel predictions on multiple sets of said input data to generate multiple predicted values ​​of the surface height of the polished wafer corresponding to said multiple sets of input data, comprising: The neural network model generates a first predicted value for the surface height of the non-groove structure and a second predicted value for the surface height of the groove structure, respectively. as well as The simulated value of the wafer surface height is determined by the neural network model based on the first predicted value and the second predicted value.

19. The method of claim 17, wherein determining the simulated value of the wafer surface height by the neural network model based on the first predicted value and the second predicted value comprises: The simulated value of the wafer surface height is determined based on the first predicted value, the second predicted value, and the density of the pattern within the grid region on the wafer.

20. The method of claim 17, further comprising: The surface condition of the wafer during the polishing process is determined using the chemical mechanical polishing physical model.

21. A neural network model, comprising: The input layer is configured to receive pattern feature data on the wafer and physical information related to the chemical mechanical polishing of the wafer; An intermediate layer is configured to perform the method of any one of claims 17 to 20 to determine the surface height value of the polished wafer; as well as The output layer is configured to output the surface height value.

22. The method of claim 21, further comprising: The embedding layer is configured to map the input physical information to generate physical vectors.

23. An electronic device, comprising: processor; as well as A memory coupled to a processor, containing instructions stored therein, which, when executed by the processor, cause the device to perform actions, including: The neural network model is trained using training samples to determine the gradient of the loss function of the neural network model, wherein the training samples include simulated values ​​of the surface height of the wafer; The parameters of the neural network model are updated based on the gradient of the loss function; and In response to the training satisfying predetermined conditions, a neural network model corresponding to the corresponding updated parameters is determined as a trained neural network model, wherein the trained neural network model outputs an enhanced formula for determining the surface height.

24. 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 13 and 17 to 20.

Images