A CMOS process optimization method and system based on particle swarm optimization algorithm

Abstract

The present invention discloses a CMOS process optimization method and system based on a particle swarm optimization algorithm. The present invention uses an electrical property prediction neural network model, a process prediction neural network model, a particle swarm optimization algorithm, and a weighted fitness function to optimize the CMOS process. The newly introduced process parameter prediction model can solve the problem of process deviation in actual manufacturing. The introduction of the weighted fitness function allows users to customize the optimization direction of electrical parameters, solving the problem of fixed optimization direction in the previous process. In addition, the optimization method introduces process parameter reconstruction error in the weighted fitness function, allowing users to control the degree of optimization and adjust the reliability of the optimization results.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor integrated circuits, and in particular to a CMOS process optimization method and system based on a particle swarm optimization algorithm. Background Art

[0002] With the rapid development of chip manufacturing technology, a variety of manufacturing processes have been applied to CMOS production. The process parameters of multiple processes have different degrees of influence on the electrical parameters of the device. Among them, the on-state current I on and the off-state current I off It is an important electrical parameter that reflects the working performance of the device. on and low I off MOSFET devices often represent better performance. Therefore, how to improve I on While reducing I off It is a very important issue.

[0003] like Figure 1 As shown in the figure, existing artificial intelligence-assisted process optimization methods usually include two parts: A. Model training; B. Process optimization. Among them, the input of the electrical prediction model is process parameters and the output is electrical parameters. Based on the fitness function obtained from the predicted values ​​of the electrical parameters, the optimization algorithm searches for the optimal process parameters within the search range and constructs an optimization model. However, the data used in the above schemes are often simulation data, and the process deviation problem in actual manufacturing is not taken into account. Process deviation causes one process to correspond to multiple sets of different device electrical parameters, and the electrical parameter prediction model in the above schemes usually cannot capture this feature. In addition, the fitness function of the above schemes is often fixed, which means that when optimizing multiple target electrical parameters that have a constrained relationship, there is a bias in the optimization direction. Summary of the Invention

[0004] In order to solve the problems existing in the prior art, the present invention provides a CMOS process optimization method and system based on a particle swarm optimization algorithm.

[0005] In a first aspect, the present invention provides a CMOS process optimization method based on a particle swarm optimization algorithm, specifically:

[0006] Step 1. Determine the optimization targets of CMOS process parameters and corresponding device electrical parameters;

[0007] Step 2. Collect CMOS production line data to obtain sample data sets corresponding to the different process parameters;

[0008] Step 3. Construct a neural network model for forward prediction of device electrical parameters and a neural network model for reverse prediction of process parameters, and train the forward model and the reverse model;

[0009] Step 4. Obtaining predicted values ​​of electrical parameters based on the forward model;

[0010] Step 5. Obtain the reconstruction error of the process parameters based on the electrical parameter prediction value and the inverse model;

[0011] Step 6. Use the particle swarm optimization algorithm to optimize the target electrical parameters and the reconstruction error of the process parameters to build a process optimization model;

[0012] Step 7. Output the optimized electrical parameters and corresponding process parameters according to the optimization model.

[0013] In a second aspect, the present invention provides a CMOS process optimization system based on a particle swarm optimization algorithm, comprising:

[0014] A target determination module is used to determine the optimization targets of CMOS process parameters and corresponding device electrical parameters;

[0015] A data collection module is used to collect CMOS production line data and obtain sample data sets corresponding to the different process parameters;

[0016] The forward and reverse model acquisition module is used to build a neural network model for forward prediction of device electrical parameters and a neural network model for reverse prediction of process parameters, and to train the forward and reverse models;

[0017] An electrical parameter prediction module, used to obtain electrical parameter prediction values ​​based on a forward model;

[0018] A reconstruction error acquisition module is used to obtain the reconstruction error of the process parameters based on the electrical parameter prediction value and the reverse model;

[0019] The process optimization model construction module is used to optimize the target electrical parameters and the reconstruction error of the process parameters using the particle swarm optimization algorithm to build a process optimization model;

[0020] The output module is used to output the optimized electrical parameters and corresponding process parameters according to the optimization model.

[0021] The present invention provides a process prediction model that accounts for process deviations, i.e., multiple sets of different electrical parameters corresponding to the same process parameters. This model, combined with a weighted fitness function, adjusts the optimization trend, thereby yielding optimized results with varying trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Design a method flow chart for an existing device;

[0023] Figure 2This is the overall flow chart of the CMOS process optimization method of this application;

[0024] Figure 3 This is a specific flow chart of the CMOS process optimization method according to an embodiment of the present application;

[0025] Figure 4 is the on-state current I of MOSFET under different initial processes on and the MOSFET's off-state current I off Relationship diagram. DETAILED DESCRIPTION

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

[0027] like Figure 2 As shown, the method of the present invention comprises the following steps:

[0028] Step 1. Determine the optimization targets of CMOS process parameters and corresponding device electrical parameters.

[0029] Step 2: Collect CMOS production line data to obtain sample data sets corresponding to the different process parameters.

[0030] Step 3. Construct a neural network model for forward prediction of device electrical parameters and reverse prediction of process parameters, and train the forward model and reverse model.

[0031] Step 4. Obtain the predicted values ​​of electrical parameters based on the forward model.

[0032] Step 5. Obtain the reconstruction error of the process parameters based on the electrical parameter prediction value and the inverse model.

[0033] Step 6. Use the particle swarm optimization algorithm to optimize the target electrical parameters and the reconstruction error of the process parameters to build a process optimization model.

[0034] Step 7. Output the optimized electrical parameters and corresponding process parameters according to the optimization model.

[0035] The present invention utilizes an electrical property prediction neural network model, a process prediction neural network model, a particle swarm optimization algorithm, and a weighted fitness function to optimize CMOS processes. The newly introduced process parameter prediction model can account for process deviations in actual manufacturing, while the introduction of a weighted fitness function allows users to customize the direction of electrical parameter optimization, resolving the previous issue of fixed optimization directions. Furthermore, this optimization method incorporates process parameter reconstruction errors into the weighted fitness function, allowing users to control the degree of optimization and adjust the reliability of the optimization results.

[0036] The neural network model for forward prediction of device electrical parameters and reverse prediction of process parameters is constructed and trained, including:

[0037] 3-1. Divide the sample data set obtained from the production line into a training set and a validation set according to the preset ratio.

[0038] 3-2. Set the input of the forward prediction model to process parameters and the output to electrical parameters; set the input of the reverse prediction model to electrical parameters and the output to process parameters.

[0039] 3-3. Use the training set to train multiple neural networks. Use preset indicators to evaluate the prediction effect of the trained model on the validation set. Determine the values ​​of the hyperparameters during the training process of the neural network and obtain a neural network model that meets the indicators.

[0040] The electrical parameter prediction values ​​are obtained based on the forward model, including:

[0041] 4-1. Input the sample data set into the forward prediction electrical parameter model to obtain the predicted values ​​of the electrical parameters.

[0042] The reconstruction error of process parameters obtained based on the predicted values ​​of electrical parameters and the reverse model includes:

[0043] 5-1. Input the electrical parameters predicted in step 4 into the reverse prediction process parameter model to obtain the predicted values ​​of the process parameters.

[0044] 5-2. Use the preset index to measure the difference between the input process parameters and the predicted process parameters as the reconstruction error of the process parameters.

[0045] The particle swarm optimization algorithm is used to optimize the electrical parameter optimization target and obtain the process optimization model, including:

[0046] 6-1. Initialize sample data and obtain the initialized particle swarm.

[0047] 6-2. Use forward and reverse models to obtain electrical parameter predictions and reconstruction errors.

[0048] 6-3. Construct a weighted fitness function based on the predicted values ​​of electrical parameters and reconstruction errors.

[0049] 6-4. Calculate the individual historical optimal position of the particle, the historical optimal position of the population, and the optimal fitness of the population

[0050] 6-5. Update the individual position vector and individual velocity vector of each particle.

[0051] 6-6. If the optimization termination condition is met, the iteration is stopped and the optimized process parameters are output; if the preset requirements are not met, return to 6-2 and restart the optimization.

[0052] Example:

[0053] The I on and I off For example, the optimization of Figure 3 As shown, the process is divided into two parts: A. Model training and B. Process optimization.

[0054] A. Model training:

[0055] A-1. Determine the process parameters and target electrical parameters that need to be optimized;

[0056] A-2. Collect data from a 55-nm CMOS production line and construct a sample dataset.

[0057] A-3. Train the model to obtain the electrical property prediction model and process prediction model.

[0058] B. Process optimization:

[0059] B-1. Initialize process parameters and obtain the initialized particle swarm;

[0060] B-2. Input the process parameters into the electrical prediction model to obtain the predicted values ​​of the electrical parameters;

[0061] B-3. ​​Input the predicted values ​​of the electrical parameters into the process prediction model to obtain the predicted values ​​of the process parameters;

[0062] B-4. Calculate the process parameter reconstruction error based on the process parameters input into the electrical prediction model and the process parameter prediction values ​​obtained by the process prediction model;

[0063] B-5. Reconstruction error and target electrical parameters based on process parameters (I on ,I off ) Calculate the fitness function;

[0064] B-6. Use particle swarm optimization algorithm to optimize process parameters;

[0065] B-7. If the process parameters meet the preset requirements or the optimization process reaches the maximum number of iterations, the optimized process parameters and electrical parameters are output. Otherwise, return to B-2 to continue optimization.

[0066] Furthermore, based on the above embodiments, when determining the 55-nm CMOS process parameters and electrical parameters, the process parameters involve gate oxide layer preparation, n-type polysilicon gate pre-doping, polysilicon gate etching, Offset Spacer preparation (offset sidewall preparation), pocket injection, Spacer preparation (sidewall preparation), source / drain injection and multiple annealing processes in the device manufacturing process; the electrical parameters are the test items of WAT (wafer acceptance test) in the production line.

[0067] Furthermore, based on the above embodiment, production line data is collected to obtain sample data sets corresponding to different process parameters.

[0068] Furthermore, based on the above embodiment, a neural network model for forward prediction of device electrical parameters and reverse prediction of process parameters is constructed, and the forward model and reverse model are trained, including:

[0069] 3-1. Set the input of the forward electrical parameter prediction model as process parameters and the output as electrical parameters; set the input of the reverse process parameter prediction model as electrical parameters and the output as process parameters.

[0070] Here, the inverse process parameter prediction model takes electrical parameters as input and outputs process parameters as output. Compared to the forward electrical prediction model, the input and output of this inverse model are reversed. This setup can account for situations where multiple sets of different electrical parameters correspond to the same process parameters due to process deviations.

[0071] 3-2. Divide the sample data set obtained from the production line into a training set and a validation set according to the preset ratio.

[0072] 3-3. Use the training set to train the forward neural network model for predicting electrical parameters (the reverse neural network model for predicting process parameters) under different hyperparameters.

[0073] 3-4. Compare the predicted results with the real data, use MSE as the loss function, and Adam as the optimizer to train the neural network.

[0074] 3-5. Use 1-MAPE (mean absolute percentage error) as the accuracy rate to evaluate whether the prediction model meets the requirements.

[0075] If the model meets the preset accuracy requirements, model training is complete. If not, adjust the model's hyperparameters and train the model again.

[0076] Furthermore, based on the above embodiment, obtaining the predicted value of the electrical parameter based on the forward model includes:

[0077] 4-1. Input the process parameters in the 55nm-CMOS production line data set into the forward prediction electrical parameter model to obtain the corresponding electrical parameter prediction values.

[0078] Furthermore, based on the above embodiment, the reconstruction error of the process parameters is obtained based on the electrical parameter prediction value and the reverse model, including:

[0079] 5-1 Input the electrical property prediction values ​​from step 4 into the reverse prediction process parameter model to obtain the process parameter prediction values, and calculate the root mean square error (RMSE) between the predicted values ​​and the process parameters input into the forward model. The RMSE is used as the reconstruction error of the process parameters.

[0080] Furthermore, based on the above embodiment, the electrical parameter target is optimized using a particle swarm optimization algorithm to obtain an optimization model, including:

[0081] 6-1. Initialize sample data and obtain the initialized particle swarm.

[0082] Here, assume that there are N optimization tasks, where the dimension corresponding to the i-th dimension is D i For each D i Set the search space to [min i ,max i ]. In the subsequent optimization process, D i The search range will not exceed the set interval.

[0083] Set the speed range of each particle [V min ,V max ], randomly initialize the position and velocity of each particle.

[0084] Set the maximum number of iterations T of the algorithm.

[0085] 6-2. Use forward and reverse models to obtain electrical parameter predictions and reconstruction errors.

[0086] Here, if optimization has not yet begun, the process parameters in the initialization sample are input into the forward model to obtain the corresponding electrical parameter prediction values. If optimization has already begun, the process parameters after the previous iteration are input into the forward model to obtain the electrical parameter prediction values.

[0087] The electrical property predictions were input into the reverse model to obtain the process parameter predictions. The RMSE between these predictions and the process parameters input into the forward model was calculated. The RMSE was used as the reconstruction error of the process parameters.

[0088] 6-3. Construct a weighted fitness function based on the predicted values ​​of electrical parameters and reconstruction errors.

[0089] Here, the electrical parameter optimization target is I on and I off Based on the predicted value and the reconstruction error of the process parameters at this time, the fitness J function of the particle swarm optimization is constructed as follows:

[0090] J=w Ion *(-I on )+w Ioff *I off +w res*RMSE(P opti ,P pred )

[0091]

[0092] Among them, I on is the on-state current of the device, I off is the off-state current of the device, P opti P is the process parameter input for the forward prediction model of electrical parameters. pred It is the process parameter prediction value output by the reverse prediction process parameter model. Ion ,w Ioff ,w res Corresponding to I on , I off and the weight of the reconstruction error

[0093] Here, the optimization algorithm reduces -I on , I off and reconstruction error, thereby reducing the fitness function. on Lowering means higher on-state current, I off Reduction means lower off-state current. Reduction of reconstruction error means P opti With P pred The closer they are, the better the consistency of the forward and reverse models, the lower the degree of optimization, and the more reliable the optimization results.

[0094] Here, the weighted fitness function assigns a weight to each optimization item. A higher weight value indicates a higher bias of the optimization item, thereby achieving customization of the optimization direction.

[0095] 6-4. Calculate the individual historical optimal position of the particle, the historical optimal position of the population, and the optimal fitness of the population

[0096] Here, the current position of each particle is calculated and compared with the historical best fitness. If the current position fitness is better, the optimal position of the particle and the corresponding optimal fitness are updated.

[0097] Compare the fitness of each particle's optimal position with the fitness of the population's optimal position. If there is a better fitness, update the population's optimal position and optimal fitness.

[0098] 6-5. Update each individual position vector and individual velocity vector.

[0099] The formula for updating the individual particle position is:

[0100]

[0101] in, is the position of the i-th particle at the d-th iteration. is the velocity of the i-th particle at the d-1th iteration.

[0102] The speed update formula is:

[0103]

[0104] Among them, w is the inertia weight of velocity, pbest i d is the individual optimal position of the i-th particle after d iterations, gbest d is the optimal position of the population after d iterations. c1 and c2 are learning factors, and r1 and r2 are uniformly distributed random numbers between [0,1].

[0105] 6-6. If the optimization termination condition is met, the iteration is stopped and the optimized process parameters are output; if the preset requirements are not met, return to 6-2 to continue the iteration.

[0106] 55-nm CMOS process optimization results:

[0107] Figure 4 The data point group in (a) represents the I under different initial processes. on and I off The expected result of achieving process optimization is reflected in the overall right shift of the data point group.

[0108] When w Ion =w Ioff When it increases, the optimization process tends to sacrifice the reliability of the optimization results in order to obtain a higher degree of optimization. Figure 4 (b) With w Ion =w Ioff As increases, the median of the data point group shifts to the right more obviously, that is, the data point group gradually shifts to the right.

[0109] When w res is a fixed value, w Ioff Gradually decrease, w Ion As it gradually increases, the optimization results tend to sacrifice I off Optimization to obtain higher I on . Reflected in Figure 4 In (c), the median of the data point group moves toward the upper right corner, that is, the data point group shows a trend of moving to the upper right corner.

[0110] In addition, if Figure 4 As shown in (d), the boundary points of the data point group will change with w Ion =w Ioff The gradual increase moves to the lower right, that is, high I on , low I offThis indicates that this method can optimize the process and obtain devices with better performance.

[0111] Based on the concept of the above method embodiment, the embodiment of the present application further provides a CMOS process optimization system based on a particle swarm optimization algorithm, comprising:

[0112] A target determination module is used to determine the optimization targets of CMOS process parameters and corresponding device electrical parameters;

[0113] A data collection module is used to collect CMOS production line data and obtain sample data sets corresponding to the different process parameters;

[0114] The forward and reverse model acquisition module is used to build a neural network model for forward prediction of device electrical parameters and a neural network model for reverse prediction of process parameters, and to train the forward and reverse models;

[0115] An electrical parameter prediction module, used to obtain electrical parameter prediction values ​​based on a forward model;

[0116] A reconstruction error acquisition module is used to obtain the reconstruction error of the process parameters based on the electrical parameter prediction value and the reverse model;

[0117] The process optimization model construction module is used to optimize the target electrical parameters and the reconstruction error of the process parameters using the particle swarm optimization algorithm to build a process optimization model;

[0118] The output module is used to output the optimized electrical parameters and corresponding process parameters according to the optimization model.

[0119] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A CMOS process optimization method based on a particle swarm optimization algorithm, characterized by: Step 1. Determine the optimization targets of CMOS process parameters and corresponding device electrical parameters; Step 2. Collect CMOS production line data and obtain sample data sets corresponding to different process parameters; Step 3. Construct a neural network model for forward prediction of device electrical parameters and a neural network model for reverse prediction of process parameters, and train the forward model and the reverse model; The input of the neural network model for forward prediction of device electrical parameters is process parameters, and the output is electrical parameters; the input of the neural network model for reverse prediction of process parameters is electrical parameters, and the output is process parameters; Step 4. Obtaining predicted values ​​of electrical parameters based on the forward model; Step 5. Obtain the reconstruction error of the process parameters based on the electrical parameter prediction value and the inverse model; Step 6. Use the particle swarm optimization algorithm to optimize the electrical parameter prediction value and the reconstruction error of the process parameters to build a process optimization model; Step 7. Output the optimized electrical parameters and corresponding process parameters according to the optimization model.

2. The CMOS process optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that: The CMOS process parameters in step 1 include gate oxide layer preparation parameters, n-type polysilicon gate pre-doping parameters, polysilicon gate etching parameters, offset spacer preparation parameters, pocket implantation parameters, spacer preparation parameters, source and / or drain implantation parameters, and multi-pass annealing process parameters.

3. The CMOS process optimization method based on particle swarm optimization algorithm according to claim 2, characterized in that: The electrical parameters described in step 1 are the test items of WAT in the production line.

4. The CMOS process optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that: MSE is used as the loss function and Adam is used as the optimizer to train the neural network model in step 3.

5. The CMOS process optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that: Step 5 is specifically as follows: input the electrical parameter prediction value in step 4 into the reverse prediction process parameter neural network model to obtain the process parameter prediction value, calculate the root mean square error between the process parameter and the process parameter input into the forward prediction device electrical parameter neural network model, and use the root mean square error as the reconstruction error of the process parameter.

6. The CMOS process optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that: The electrical parameter prediction values ​​in step 6 are the on-state current and the off-state current; based on their predicted values ​​and the corresponding process parameter reconstruction errors, a fitness function of the particle swarm optimization is constructed.

7. A CMOS process optimization system based on particle swarm optimization algorithm, characterized in that: include: A target determination module is used to determine the optimization targets of CMOS process parameters and corresponding device electrical parameters; Data collection module, used to collect CMOS production line data and obtain sample data sets corresponding to different process parameters; The forward and reverse model acquisition module is used to build a neural network model for forward prediction of device electrical parameters and a neural network model for reverse prediction of process parameters, and to train the forward and reverse models; The input of the neural network model for forward prediction of device electrical parameters is process parameters, and the output is electrical parameters; the input of the neural network model for reverse prediction of process parameters is electrical parameters, and the output is process parameters; An electrical parameter prediction module, used to obtain electrical parameter prediction values ​​based on a forward model; A reconstruction error acquisition module is used to obtain the reconstruction error of the process parameters based on the electrical parameter prediction value and the reverse model; The process optimization model construction module is used to optimize the electrical parameter prediction values ​​and the reconstruction errors of the process parameters using the particle swarm optimization algorithm to build a process optimization model; The output module is used to output the optimized electrical parameters and corresponding process parameters according to the optimization model.

Images