A Machine Learning-Based Method for Optimizing Split-Gate Memory Parameters

Abstract

The present invention discloses a method for optimizing split-gate memory parameters based on machine learning. First, a parameter space is constructed with the design parameters of the split-gate memory, and multi-dimensional Latin hypercube sampling is used for initial data acquisition. Subsequently, TCAD simulations are carried out to obtain electrical performance parameters. Then, according to the correlation between the input parameters and the electrical performance, the sampling strategy is optimized based on the correlation, and a new parameter combination is selected for TCAD simulation to obtain an optimized data set. Next, the data is preprocessed, the training set, validation set, and test set are divided, and a machine learning model is constructed. Finally, the trained model is used to replace part of the TCAD simulations to achieve rapid prediction of electrical performance, thereby accelerating the parameter optimization process of the split-gate memory and improving the design efficiency and accuracy. The method of the present invention can effectively reduce the number of required simulations, reduce the consumption of computing resources, and provide an efficient optimization scheme for the design of memory devices.

Description

Technical Field

[0001] The present invention belongs to the technical field of semiconductor devices and manufacturing, and particularly relates to a method for optimizing split-gate memory parameters based on machine learning. Background Art

[0002] In traditional integrated circuit design, technology computer-aided design (TCAD) simulation tools are widely used in device design and optimization. TCAD simulation technology provides designers with accurate device performance predictions by simulating and analyzing the electrical, thermal, and mechanical characteristics of semiconductor devices, helping to design and optimize integrated circuits (ICs). TCAD tools are mainly used in the design process to calculate performance parameters such as the current-voltage characteristics (I-V characteristics), switching characteristics, threshold voltage, and leakage current of devices, so as to optimize device structures and process parameters.

[0003] However, TCAD simulation usually requires a large amount of computing resources and long computing time, especially in the case of involving multiple design parameters. For complex device structures, such as split-gate memories, due to the huge parameter space, the simulation time and computing resource requirements are extremely high. Even with advanced parallel computing technology, TCAD simulation still cannot meet the requirements of efficient and rapid optimization.

[0004] Split-gate memory is a memory cell widely used in non-volatile memory (NVM) technology, and its structure is as Figure 1 shown. It has a high storage density and good write / read performance. The optimization of split-gate memory usually involves multiple design parameters, such as front gate length (FG_length), front gate oxide thickness (FG_ox), source / drain implantation concentration (LDD_IMP3), back gate oxide thickness (EP_ox), etc. In order to improve the performance of the memory, such as parameters like Ion (on-state current), Ioff (off-state current), Vp (programming voltage), Ve (erase voltage), etc., these design parameters need to be precisely adjusted to ensure the optimal working state of the device. Traditional optimization methods rely on experiments and TCAD simulation, but due to large amounts of calculation, long time, and high cost, engineers face the problem of efficiently optimizing in a large-scale design space.

[0005] In recent years, machine learning (ML) technology has been increasingly widely used in integrated circuit design and optimization. Compared with traditional TCAD simulation methods, machine learning methods can significantly reduce the computing time and data requirements, and can directly extract the complex relationship between device performance and design parameters from data without a complex physical model. However, it still has the following defects:

[0006] (1) Data sampling problem: To train an accurate machine learning model, a large amount of training data is usually required. When dealing with complex device designs, it is very difficult to generate effective training data, especially when the parameter space is very extensive.

[0007] (2) Model accuracy and interpretability: At present, many machine learning models still lack sufficient interpretability. It is very difficult for designers to understand why a machine learning model selects a certain combination of parameters, which may limit the practical application of the model. Summary of the Invention

[0008] The first object of the present invention is to propose a method for optimizing the parameters of a split-gate memory in view of the deficiencies of the prior art. Through effective sampling, correlation analysis, machine learning modeling, and replacement of the simulation model, the present invention greatly accelerates the process of parameter optimization, reduces the computational cost, and improves the efficiency and accuracy of the design.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] Step S1: Define the design parameters of the split-gate memory and their physical value ranges, and construct a parameter space;

[0011] Step S2: Use multi-dimensional Latin hypercube sampling in the parameter space to generate multiple groups of initial parameter combinations that evenly cover the parameter space; perform TCAD simulation on each group of initial parameter combinations to obtain the first simulation results of the electrical performance parameters;

[0012] Step S3: Calculate the Spearman correlation coefficient between the design parameters and the first simulation results, then calculate the sampling weights of each design parameter, and adjust the number of interval divisions of the parameter space according to the sampling weights to optimize the sampling strategy;

[0013] Step S4: Resample in the parameter space according to the optimized sampling strategy to obtain parameter combinations, perform TCAD simulation on the parameter combinations to obtain the second simulation results of the electrical performance parameters; construct a data set with the parameter combinations as input features and the second simulation results as labels, and divide the data set into a training set, a validation set, and a test set;

[0014] Step S5: Construct a machine learning model, preprocess the data set, and then use the preprocessed data set to train, validate, and test the machine learning model, and output the prediction results of the electrical performance parameters.

[0015] The second object of the present invention is to provide a device for optimizing the parameters of a split-gate memory, including the following modules:

[0016] A data acquisition module, which combines multi - dimensional Latin hypercube sampling, Spearman correlation coefficient, and TCAD simulation to obtain the design parameter combinations and simulation results of the split - gate memory;

[0017] A parameter optimization module, which inputs the design parameter combinations and simulation results into a trained machine - learning model and outputs the prediction results of electrical performance parameters.

[0018] The third object of the present invention is to provide an electronic device, including a processor and a memory. The memory stores machine - executable instructions that can be executed by the processor, and the processor executes the machine - executable instructions to implement the above - mentioned method.

[0019] The fourth object of the present invention is to provide a machine - readable storage medium. The machine - readable storage medium stores machine - executable instructions, and when the machine - executable instructions are called and executed by a processor, the machine - executable instructions cause the processor to implement the above - mentioned method.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] By using multi - dimensional Latin hypercube sampling (Latin Hypercube Sampling, LHS) and calculating the Spearman correlation coefficient, the present invention effectively reduces the amount of data required for training the model, ensures the diversity and representativeness of the sampled data, analyzes the influence of different parameters on the device performance, and automatically assigns weights to each parameter, thereby optimizing the sampling strategy of the parameter space.

[0022] The present invention uses a trained machine - learning model to replace the traditional TCAD simulation, significantly accelerating the parameter optimization process, reducing the calculation time and resource consumption. By training the machine - learning model, the present invention can quickly screen and optimize the optimal parameter combinations in a large - scale parameter space, greatly improving the efficiency and accuracy of the memory design. Through this method, the present invention can, within a short time, use machine learning to replace the cumbersome TCAD simulation process to achieve the rapid optimization of the split - gate memory, reduce the consumption of computing resources, and significantly improve the design efficiency. Description of the Drawings

[0023] Figure 1The structure of a split-gate memory, where EP is the Epitaxial Layer, ONO is the Oxide-Nitride-Oxide, CG is the Control Gate, FG is the Floating Gate, SG1 is the Select Gate 1, SG2 is the Select Gate 2, SRC is the Source, LDD is the Lightly Doped Drain, and BL is the Bit Line.

[0024] Figure 2 This is the flowchart of the present invention.

[0025] Figure 3 This is the structural diagram of the machine learning model of the present invention. Detailed implementation manners

[0026] The present invention will be further described below with reference to the accompanying drawings.

[0027] As Figure 2 shown, the present invention relates to a parameter optimization method based on machine learning, particularly a technique for replacing the traditional TCAD simulation model to optimize the parameters of a split-gate memory. By means of effective sampling, correlation analysis, machine learning modeling, and replacement of the simulation model, the present invention greatly accelerates the parameter optimization process, reduces the computational cost, and improves the design efficiency and accuracy.

[0028] Specifically, the present invention includes the following steps:

[0029] Step S1: Define the design parameters of the split-gate memory and their physical value ranges, and construct a parameter space;

[0030] The design parameters and their physical value ranges are shown in Table 1 below:

[0031] Table 1 Input parameters and value ranges

[0032] Serial number Parameter name Value range Interpretation 1 FG_length [60, 80] Floating gate length 2 FG_ox [8, 10] Floating gate oxide layer thickness 3 SG-FG_ONO [10, 15] ONO thickness between floating gate and select gate 4 FG-CG_N [3, 5] Nitride thickness between floating gate and control gate 5 FG-CG_O [2, 3] Oxide thickness between floating gate and control gate 6 SG_THK [90, 110] Select gate thickness 7 SG_ox [13, 15] Select gate oxide layer thickness 8 LDD_IMP3 [15, 20] Lightly doped drain implantation dose 9 EP_ox [8, 10] Erase gate oxide layer thickness

[0033] Step S2: Obtain parameter combinations by using multi-dimensional Latin hypercube sampling (LHS) within the parameter space: In this embodiment, 100 groups of initial parameter combinations that uniformly cover the parameter space are generated.

[0034] The goal of LHS is to ensure that each variable is sampled uniformly within its distribution range by dividing the range of each input variable, thus avoiding the "aggregation" or "omission" phenomena that may occur in traditional Monte Carlo sampling.

[0035] For each dimension i in the d-dimensional space, the parameter range [ai,bi] is divided into n intervals with a width of In the i-th dimension, suppose an interval j is randomly selected, and its corresponding point is:

[0036]

[0037] where j∈{1,2,…,n} represents the center of the jth interval in that dimension.

[0038] According to the requirements of the embodiment of the present invention, d=9 is set (a total of 9 input parameters), and the corresponding value of i is {1, 2, 3, 4, 5, 6, 7, 8, 9}. As shown in Table 1, when i is 1, the parameter range is is [60, 80]. When i is 2, the parameter range is The parameter range is [8, 10]. When i takes other values, the parameter range is similar. We want to sample 100 sets of data, so n=100, and divide each parameter range into 100 intervals.

[0039] For the parameter FG_length, the sampling points are

[0040]

[0041] For the parameter FG_ox, the sampling points are

[0042]

[0043] Similarly, the other parameters are sampled. Then, the 100 sampling points of each parameter are randomly arranged, and finally the randomly arranged sampling points are sequentially combined into 100 groups of initial parameter combinations. Take the first five groups of sampled data as an example, as shown in Table 2:

[0044] Table 2 Initial sampling data of the first five groups

[0045] FG_length (Floating gate length) FG_ox (Floating gate oxide layer thickness) SG-FG_ONO (ONO thickness between floating gate and select gate) FG-CG_N (Nitride thickness between floating gate and control gate) FG-CG_O (Oxide thickness between floating gate and control gate) SG_THK (Select gate thickness) SG_ox (Select gate oxide layer thickness) LDD_IMP3 (Lightly doped drain implantation dose) EP_ox (Erase gate oxide layer thickness) 72.34 9.12 12.45 4.23 2.56 98.76 14.23 17.89 9.01 65.78 8.45 11.23 3.89 2.12 105.34 13.78 16.45 8.56 78.12 9.78 14.56 4.56 2.89 92.34 14.56 18.23 9.45 69.45 8.89 13.12 3.45 2.34 102.12 13.45 17.12 8.89 74.56 9.34 12.89 4.12 2.67 96.78 14.12 16.78 9.23

[0046] Then, a TCAD simulation (Sentaurus TCAD) was performed on the 100 sets of initial data to obtain the output corresponding to each set of input parameters, i.e., the first simulation results of the electrical performance parameters. The outputs included: Ion (on-state current), Ioff (off-state current), Vp (programming voltage), and Ve (erase voltage).

[0047] An example of simulation output data is shown in Table 3 below:

[0048] Table 3 Simulation output data example

[0049] Ioff (Off-state current) Ion (On-state current) Vp (Programming voltage) Ve (Erase voltage) 3.85E-11 1.73E-05 -1.54828 3.88976 4.20E-11 1.69E-05 -1.62352 3.89842 4.16E-11 1.68E-05 -1.90497 3.84419 5.26E-11 1.66E-05 -1.76001 3.84333 5.70E-11 1.64E-05 -2.0828 3.82126 5.53E-11 1.61E-05 -1.73298 3.8244 4.41E-11 1.62E-05 -2.11759 3.74379

[0050] Step S3: Calculate the Spearman correlation coefficient between the parameter combination and the first simulation result, adjust the sampling distribution based on the correlation, and optimize the sampling strategy;

[0051] The Spearman correlation coefficient is particularly suitable for measuring the correlation between ordinal data or ranked data and is also widely used in the analysis of non-linear relationships. It calculates the strength of the relationship between variables by sorting the data and based on the sorting results.

[0052] For two variables and , first sort the values of each variable and assign sequential numbers according to their rankings.

[0053] For variable X, sort its values in ascending order to obtain the corresponding rankings .

[0054] For variable Y, sort its values in ascending order to obtain the corresponding rankings .

[0055] For each pair of sample points , calculate the ranking difference in the two variables:

[0056]

[0057] where di is the ranking difference of the i-th sample point in variables X and Y.

[0058] Finally, the Spearman correlation coefficient ρ is calculated by the following formula:

[0059]

[0060] where n is the total number of sample data, is the sum of the squares of the ranking differences of all sample points.

[0061] The value range of the Spearman correlation coefficient ρ is between [−1, 1]:

[0062] ρ = 1 indicates a perfect positive correlation (i.e., there is a perfect monotonic relationship between the two variables).

[0063] ρ = −1 indicates a perfect negative correlation (i.e., there is a perfect monotonic reverse relationship between the two variables).

[0064] ρ = 0 indicates no monotonic relationship.

[0065] The calculation process of the Spearman correlation coefficient mainly includes sorting the data, calculating the ranking differences, and calculating the correlation based on the ranking differences. The formula calculation process can effectively measure the strength of the correlation between variables. In this step, the strength of the correlation between the measurement parameter combination and the electrical performance parameters is measured.

[0066] Specifically in the present invention, the Spearman correlation coefficient is calculated for the input parameters and the electrical performance parameters. For example, if it is desired to calculate the correlation coefficient between the input parameter FG_length and the output variable Ion, then the values of FG_length and the values of Ion

[0067] are sorted in ascending order, and sequential numbers are assigned according to their rankings. .

[0068] For the variable FG_length, its values are sorted in ascending order to obtain the corresponding rankings .

[0069] For each pair of sample points , the ranking differences in the two variables are calculated:

[0070]

[0071] where is the ranking difference of the i-th sample point in FG_length and Ion.

[0072] Finally, the Spearman correlation coefficient ρ between FG_length and Ion is calculated by the following formula:

[0073]

[0074] where n is the total number of sample data, which is 100 in this embodiment, is the sum of the squares of the ranking differences of all sample points. The Spearman correlation coefficients between FG_length and Ioff, Vp, and Ve are calculated in the same way. The average value of the absolute values of the Spearman correlation coefficients between FG_length and Ion, Ioff, Vp, and Ve is taken as the correlation coefficient between FG_length and the output result , that is, the following formula:

[0075]

[0076] where 0.1 is a smoothing term to avoid the correlation coefficient being 0. Similarly, the correlation coefficients between other parameters and the output result can be calculated, and the sampling weights are calculated according to the following formula .

[0077]

[0078] Among them, i = {1, 2, 3, …, 9}, which is the serial number of the input parameter.

[0079] Step S4: Resample in the parameter space according to the optimized sampling strategy to obtain parameter combinations, perform TCAD simulations on the parameter combinations to obtain the second simulation results of the electrical performance parameters; preprocess the parameter combinations and the second simulation results and construct a data set, and then divide it into a training set, a validation set and a test set.

[0080] The specific optimized sampling strategy is as follows: according to the sampling weights calculated in step S3, adjust the number of intervals of each input parameter. The number of intervals of high-weight parameters is more, the interval width is narrower, and the sampling density is greater; the number of intervals of low-weight parameters is less, the interval width is wider, and the sampling density is smaller. The number of sampling intervals for each parameter is as follows:

[0081]

[0082] Among them is the sampling weight of the parameter, is the total number of intervals, which is set to 2500. Divide each parameter according to the corresponding number of intervals, and perform random sampling within each interval. Finally, randomly combine each parameter point to obtain 2500 groups of final sampling data, and complete the corresponding simulation in TCAD to obtain the data set.

[0083] Step S5: Construct a machine learning model, perform normalization processing on the data set to ensure that each input data is on the same scale before training, and then divide it into a training set, a test set and a validation set. Train, validate and test the machine learning model. The preprocessed data passes through two hidden layers, and finally de-normalize after training is completed to obtain the final output result.

[0084] Specifically, divide the 2500 groups of experimental data according to 80% for the training set, 10% for the test set, and 10% for the validation set. The model structure for machine learning is as Figure 3As shown, it includes an input layer, a shared layer, a multi-task layer, and an output layer. Among them, the input layer receives 9 input parameters. The shared layer extracts the common features of the input parameters through two fully connected layers (using 128 and 256 neurons respectively), and introduces a batch normalization layer and a Dropout layer to improve the model stability and prevent overfitting. The multi-task layer independently designs branches for each output variable. Each branch contains a fully connected layer (64 neurons) and a linear output layer, which are used to predict specific performance parameters. After denormalization, it is restored to the original data scale, and the predicted values of the on-state current, off-state current, programming voltage, and erasing voltage are output respectively at the output layer. This model reduces the model complexity through the shared feature layer, and at the same time improves the prediction accuracy through independent task branches, and is applicable to multi-task regression problems.

[0085] The above are only the preferred embodiments of one or more embodiments of this specification, and are not intended to limit one or more embodiments of this specification. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of one or more embodiments of this specification shall be included within the scope of protection of one or more embodiments of this specification.

Claims

1. A method for optimizing split-gate memory parameters based on machine learning, characterized in that, It includes the following steps: Step S1: Define the design parameters of the split-gate memory and their physical value ranges, and construct a parameter space; Step S2: Adopt multi-dimensional Latin hypercube sampling within the parameter space to generate multiple groups of initial parameter combinations that uniformly cover the parameter space; Perform TCAD simulation on each group of initial parameter combinations to obtain the first simulation results of the electrical performance parameters; Step S3: Calculate the sampling weights of each design parameter according to the correlation between the design parameters and the first simulation results, adjust the number of interval divisions of the parameter space according to the sampling weights, and optimize the sampling strategy; Step S4: Resample within the parameter space according to the optimized sampling strategy to obtain parameter combinations, perform TCAD simulation on the parameter combinations, and obtain the second simulation results of the electrical performance parameters; Construct a data set with the parameter combinations as input features and the second simulation results as labels, and divide the data set into a training set, a validation set, and a test set; Step S5: Construct a machine learning model, preprocess the data set, and then use the preprocessed data set to train, validate, and test the machine learning model, and output the prediction results of the electrical performance parameters.

2. The method for optimizing split-gate memory parameters based on machine learning according to claim 1, characterized in that The design parameters include: front gate length FG_length, front gate oxide thickness FG_ox, source-front gate oxide thickness SG-FG_ONO, front gate-control gate forward current FG-CG_N, front gate-control gate oxide thickness FG-CG_O, source thickness SG_THK, source oxide thickness SG_ox, LLDD implantation concentration DD_IMP3, and back gate oxide thickness EP_ox.

3. The method for optimizing split-gate memory parameters based on machine learning according to claim 1, wherein The electrical performance parameters include: on-state current Ion, off-state current Ioff, programming voltage Vp, and erase voltage Ve.

4. The method for optimizing split-gate memory parameters based on machine learning according to claim 1, wherein The specific optimization sampling strategy is as follows: Adjust the number of intervals of each design parameter according to the sampling weights, as shown in the following formula: ； wherein is the sampling weight of the design parameter, is the total number of intervals; Divide each design parameter according to the corresponding number of intervals, then perform random sampling within each interval, and finally randomly combine each parameter point.

5. The method for optimizing split-gate memory parameters based on machine learning according to claim 1, wherein The machine learning model includes an input layer, a shared layer, a multi-task layer, and an output layer connected in series in sequence; The input layer is used to receive input data and perform normalization preprocessing on the input data; The shared layer extracts general features through multiple fully connected layers and provides shared information for subsequent prediction tasks; The multi-task layer, based on the output of the shared layer, distributes the features to different prediction tasks and calculates the predicted values of each electrical performance parameter respectively; The output layer is used to perform denormalization processing on the predicted values and then output the prediction results of the electrical performance parameters.

6. The method for optimizing split-gate memory parameters based on machine learning according to claim 5, wherein The shared layer consists of two layers of fully connected layers, and uses a non-linear activation function for feature transformation to extract general features for subsequent tasks to share, and at the same time uses batch normalization and Dropout to prevent overfitting; The multi-task layer inputs the output of the shared layer into four independent sub-networks, each sub-network corresponding to a prediction task; each sub-network consists of two layers of fully connected layers and an activation function.

7. A split-gate memory parameter optimization device for performing the method according to any one of claims 1-6, characterized in that, It includes the following modules: A data acquisition module, configured to combine multi-dimensional Latin hypercube sampling, Spearman correlation coefficient and TCAD simulation to obtain a design parameter combination and simulation results of a split-gate memory; A parameter optimization module, which inputs the design parameter combination and simulation results into a trained machine learning model and outputs a prediction result of electrical performance parameters.

8. An electronic device, comprising a processor and a memory, characterized in that, The memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the method according to any one of claims 1-6.

9. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are called and executed by the processor, the machine-executable instructions cause the processor to implement the method according to any one of claims 1-6.

Images