Device structure and process condition optimization method and apparatus, and electronic device

Abstract

The invention provides a device structure and process condition optimization method and apparatus, and an electronic device, and belongs to the technical field of computers, and the method comprises the steps of obtaining initial device structure data and initial process condition data; inputting the initial device structure data and the initial process condition data into one or more first performance index prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in a training result; determining a data pair composed of the first device structure data and the first process condition data as an individual in the population, and determining a first performance prediction result as the fitness of the individual; and solving the multi-objective function based on an NSGA-II algorithm to obtain a Pareto frontier solution set of the population, and taking the Pareto frontier solution set as an optimization result of a device structure and process conditions. The method is used for solving the problems of long design time and high cost in the existing device structure optimization process.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and specifically to a device structure and process condition optimization method, a device structure and process condition optimization device, an electronic device, a machine-readable storage medium and a computer program product. Background Art

[0002] The existing semiconductor device structure optimization design process is to first carry out process deviation and structure optimization design based on the existing baseline product, then use TCAD simulation software to model the process device simulation, simulate the device performance indicators, screen out device structures and process conditions that may have excellent performance, conduct actual small-scale tape-out verification, test device performance indicators, and determine the optimized device structure and process conditions.

[0003] However, the existing device structure optimization design requires device process simulation and device electrical simulation. This simulation process is highly dependent on the professional knowledge and experience of personnel, and simulation modeling is difficult and takes a long time. Subsequent chip verification and related Design of Experiment (DOE) analysis are also required to manually find the optimal or better device process conditions and structural dimensions. It can be seen that the existing device structure optimization process has the problems of long design time and high cost. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a device structure and process condition optimization method, apparatus and electronic equipment to solve the problems of long design time and high cost in the existing device structure optimization process.

[0005] In order to achieve the above object, an embodiment of the present invention provides a device structure and process condition optimization method, comprising:

[0006] Acquire initial device structure data and initial process condition data;

[0007] Inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and filtering out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results; the target device structure data includes a plurality of device structure data, and the target process condition data includes a plurality of process condition data;

[0008] Determine a data pair consisting of first device structure data and first process condition data as an individual in a population, and determine a first performance prediction result as the fitness of the individual; the first device structure data is any one of the multiple device structure data, and the first process condition data is any one of the multiple process condition data; the first performance prediction result is obtained by inputting the first device structure data and the first process condition data into a second performance indicator prediction model; the second performance indicator prediction model is obtained by training based on the target device structure data, the target process condition data, and the performance data label associated with the target device structure data and the target process condition data;

[0009] The multi-objective function is solved based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0010] Optionally, the inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises:

[0011] Inputting the initial device structure data and the initial process condition data into a first performance indicator prediction model for model training;

[0012] Determine the initial device structure data whose weight of the training result is greater than or equal to the set weight threshold as the target device structure data, and determine the initial process condition data whose weight of the training result is greater than or equal to the set weight threshold as the target process condition data.

[0013] Optionally, the inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises:

[0014] Inputting the initial device structure data and the initial process condition data into a plurality of first performance indicator prediction models for model training;

[0015] Performing an average operation based on weight values ​​of the initial device structure data in a plurality of the first performance indicator prediction models to obtain a first average weight of the initial device structure data;

[0016] Performing an average operation based on weight values ​​of the initial process condition data in a plurality of the first performance indicator prediction models to obtain a second average weight of the initial process condition data;

[0017] Initial device structure data whose first average weight is greater than or equal to a set weight threshold is determined as the target device structure data, and initial process condition data whose second average weight is greater than or equal to a set weight threshold is determined as the target process condition data.

[0018] Optionally, solving the multi-objective function based on the NSGA-II algorithm to obtain a Pareto front solution set of the population as an optimization result of the device structure and process conditions includes:

[0019] Repeat the following steps until the termination condition is set:

[0020] performing non-dominated sorting on all individuals in the population to divide all individuals in the population into a plurality of non-dominated layers;

[0021] Calculate the crowding degree of each individual in the non-dominated layer;

[0022] Generate offspring population based on selection, crossover and mutation operations;

[0023] Merging the population and the progeny population to form a combined population;

[0024] All individuals in the joint population are non-dominated and sorted, and individuals that meet set conditions are selected from the joint population according to the non-dominated level and the crowding degree to form a new generation population.

[0025] Optionally, performing non-dominated sorting on all individuals in the population includes:

[0026] Initialize the number of dominated individuals and the set of dominated individuals of each individual;

[0027] Traverse all individuals in the population and update the number of dominant individuals and the set of dominated individuals of each individual according to the dominance relationship;

[0028] Find out all individuals that are not dominated by any other individuals, and classify the individuals that are not dominated by any other individuals into the first layer;

[0029] Repeat the following steps until all individuals are sorted into the corresponding levels:

[0030] Continue to find all the individuals that are not dominated by other individuals among the remaining unstratified individuals in the population and assign them to the second layer.

[0031] Optionally, calculating the crowding degree of each individual in the non-dominated layer includes:

[0032] Initialize the crowding degree of each individual;

[0033] Sort the individuals in each non-dominated layer by the objective function value;

[0034] The crowding degree of the last sorted boundary individual in each non-dominated layer is set to infinity, and the crowding degree of the non-boundary individuals in each non-dominated layer is calculated according to the difference in the objective function values ​​of adjacent individuals.

[0035] Optionally, the initial device structure data includes at least one of gate width, gate length, number of fingers and pitch; the initial process condition data includes at least one of deposition rate and deposition temperature.

[0036] Optionally, the first performance indicator prediction model selects any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model and a lightweight gradient boosting machine model.

[0037] Optionally, the second performance indicator prediction model selects any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model and a lightweight gradient boosting machine model.

[0038] On the other hand, an embodiment of the present invention further provides a device structure and process condition optimization device, comprising:

[0039] An acquisition module, used for acquiring initial device structure data and initial process condition data;

[0040] A screening module, used for inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training result; the target device structure data includes a plurality of device structure data, and the target process condition data includes a plurality of process condition data;

[0041] A determination module, used to determine a data pair consisting of first device structure data and first process condition data as an individual in a population, and to determine a first performance prediction result as the fitness of the individual; the first device structure data is any one of the multiple device structure data, and the first process condition data is any one of the multiple process condition data; the first performance prediction result is obtained by inputting the first device structure data and the first process condition data into a second performance indicator prediction model; the second performance indicator prediction model is obtained by training based on the target device structure data, the target process condition data, and the performance data label associated with the target device structure data and the target process condition data;

[0042] A construction module, used for constructing a multi-objective function based on all individuals in the population;

[0043] The solution module is used to solve the multi-objective function based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0044] Optionally, the inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises:

[0045] Inputting the initial device structure data and the initial process condition data into a first performance indicator prediction model for model training;

[0046] Determine the initial device structure data whose weight of the training result is greater than or equal to the set weight threshold as the target device structure data, and determine the initial process condition data whose weight of the training result is greater than or equal to the set weight threshold as the target process condition data.

[0047] Optionally, the inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises:

[0048] Inputting the initial device structure data and the initial process condition data into a plurality of first performance indicator prediction models for model training;

[0049] Performing an average operation based on weight values ​​of the initial device structure data in a plurality of the first performance indicator prediction models to obtain a first average weight of the initial device structure data;

[0050] Performing an average operation based on weight values ​​of the initial process condition data in a plurality of the first performance indicator prediction models to obtain a second average weight of the initial process condition data;

[0051] Initial device structure data whose first average weight is greater than or equal to a set weight threshold is determined as the target device structure data, and initial process condition data whose second average weight is greater than or equal to a set weight threshold is determined as the target process condition data.

[0052] Optionally, solving the multi-objective function based on the NSGA-II algorithm to obtain a Pareto front solution set of the population as an optimization result of the device structure and process conditions includes:

[0053] Repeat the following steps until the termination condition is set:

[0054] performing non-dominated sorting on all individuals in the population to divide all individuals in the population into a plurality of non-dominated layers;

[0055] Calculate the crowding degree of each individual in the non-dominated layer;

[0056] Generate offspring population based on selection, crossover and mutation operations;

[0057] Merging the population and the progeny population to form a combined population;

[0058] All individuals in the joint population are non-dominated and sorted, and individuals that meet set conditions are selected from the joint population according to the non-dominated level and the crowding degree to form a new generation population.

[0059] Optionally, performing non-dominated sorting on all individuals in the population includes:

[0060] Initialize the number of dominated individuals and the set of dominated individuals of each individual;

[0061] Traverse all individuals in the population and update the number of dominant individuals and the set of dominated individuals of each individual according to the dominance relationship;

[0062] Find out all individuals that are not dominated by any other individuals, and classify the individuals that are not dominated by any other individuals into the first layer;

[0063] Repeat the following steps until all individuals are sorted into the corresponding levels:

[0064] Continue to find all the individuals that are not dominated by other individuals among the remaining unstratified individuals in the population and assign them to the second layer.

[0065] Optionally, calculating the crowding degree of each individual in the non-dominated layer includes:

[0066] Initialize the crowding degree of each individual;

[0067] Sort the individuals in each non-dominated layer by the objective function value;

[0068] The crowding degree of the last sorted boundary individual in each non-dominated layer is set to infinity, and the crowding degree of the non-boundary individuals in each non-dominated layer is calculated according to the difference in the objective function values ​​of adjacent individuals.

[0069] Optionally, the initial device structure data includes at least one of gate width, gate length, number of fingers and pitch; the initial process condition data includes at least one of deposition rate and deposition temperature.

[0070] Optionally, the first performance indicator prediction model selects any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model and a lightweight gradient boosting machine model.

[0071] Optionally, the second performance indicator prediction model selects any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model and a lightweight gradient boosting machine model.

[0072] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned device structure and process condition optimization method when executing the program.

[0073] On the other hand, the present invention also provides a machine-readable storage medium having a computer program stored thereon, which implements the above-mentioned device structure and process condition optimization method when executed by a processor.

[0074] On the other hand, the present invention also provides a computer program product, including a computer program, which implements the above-mentioned device structure and process condition optimization method when executed by a processor.

[0075] Through the above technical scheme, the present invention trains the first performance index prediction model by fusing the initial device structure data and the initial process condition data, screens out the target device structure data and target process condition data that affect each performance index of the device through the first performance index prediction model, and performs secondary model training based on the second performance index prediction model; then uses the NSGA-Ⅱ algorithm to solve according to the pre-set multi-objective function to obtain the Pareto front solution set as the optimization result of the device structure and process conditions. The embodiment of the present invention can significantly reduce the number of TCAD simulations and the number of actual process tape-outs during device structure optimization design, shorten the design time and reduce costs, and accelerate the progress of device research and development.

[0076] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0078] Figure 1 This is one of the flow diagrams of the device structure and process condition optimization method provided by the present invention;

[0079] Figure 2 This is the second flow chart of the device structure and process condition optimization method provided by the present invention;

[0080] Figure 3 It is a schematic diagram of the Pareto optimal frontier distribution of the multi-objective function provided by the present invention;

[0081] Figure 4 It is a schematic diagram of the elite strategy of the NSGA-Ⅱ algorithm provided by the present invention;

[0082] Figure 5 It is a schematic diagram of the Pareto front solution set for two objective functions provided by the present invention;

[0083] Figure 6 It is a schematic diagram of the structure of the device structure and process condition optimization device provided by the present invention;

[0084] Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0085] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0086] Method Embodiment

[0087] Please refer to Figure 1 , an embodiment of the present invention provides a device structure and process condition optimization method, comprising:

[0088] Step 100: Acquire initial device structure data and initial process condition data.

[0089] The electronic device can obtain initial device structure data and initial process condition data based on data sources such as historical simulation data and test data. The initial device structure data and the initial process condition data include multiple data respectively. The initial device structure data includes at least one of gate width, gate length, number of fingers and pitch; the initial process condition data includes at least one of deposition rate and deposition temperature. For example, the initial device structure data may include multiple types of device structure data such as the gate width W, gate length L, number of fingers (the number of transistors connected in parallel in the MOS tube), pitch (the distance between adjacent pins or pads) of the device. The initial process condition data may include deposition rate, deposition temperature, etc. The data pair consisting of the initial device structure data and the initial process condition data also has corresponding label data, i.e., device performance index data. The device performance index data may include one or more of the gate turn-on voltage Vtlin in the linear working area, the gate turn-on voltage Vtsat in the saturated working area, the drain current Idlin in the linear working area, the saturated drain current Idsat, the turn-off current Idoff, the breakdown voltage BVon in the on state, the breakdown voltage BVoff in the off state, and the substrate current Isub.

[0090] Step 200: input the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screen out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results.

[0091] Among them, the first performance indicator prediction model can be a traditional machine learning model or a neural network model. The traditional machine learning model used includes but is not limited to a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model, a lightweight gradient boosting machine model, etc. That is, the first performance indicator prediction model selects any one of the random forest model, the extreme gradient boosting tree model, the gradient boosting decision tree model and the lightweight gradient boosting machine model.

[0092] In the embodiment of the present invention, it is necessary to perform feature screening on the initial device structure data and the initial process condition data. For example, the electronic device inputs the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screens out target device structure data and target process condition data with higher importance based on the weight of the initial device structure data and the weight of the initial process condition data in the training results.

[0093] Take the random forest model as an example. The electronic device inputs the initial device structure data and the initial process condition data into the random forest model for model training. By training a random forest model, the importance of the initial device structure data and the initial process condition data is obtained using the feature_importances_ attribute. This feature_importances attribute returns an array representing the importance score (i.e., weight) of each initial device structure data and initial process condition data. These scores can be used to understand which features have a greater impact on the prediction results of the model. By screening the initial device structure data and the initial process condition data, target device structure data and target process condition data with a larger importance score (larger weight) are obtained. For example, the importance scores of initial device structure data a, initial device structure data b, initial device structure data c, initial device structure data d, initial device structure data e, initial process condition data f, initial process condition data g, initial process condition data h, initial process condition data i and initial process condition data j are 0.15, 0.22, 0.07, 0.03, 0.08, 0.11, 0.19, 0.09, 0.07 and 0.09 respectively. The electronic device selects the initial device structure data a, initial device structure data b and initial device structure data d corresponding to 0.15, 0.22 and 0.19 as the target device structure data, and selects the initial process condition data g corresponding to 0.11 as the target process condition data.

[0094] Step 300, determine a data pair consisting of first device structure data and first process condition data as an individual in a population, and determine a first performance prediction result as the fitness of the individual; the first device structure data is any one of the multiple device structure data, and the first process condition data is any one of the multiple process condition data; the first performance prediction result is obtained after inputting the first device structure data and the first process condition data into a second performance indicator prediction model; the second performance indicator prediction model is trained based on the target device structure data, the target process condition data, and the performance data labels associated with the target device structure data and the target process condition data.

[0095] The target device structure data includes a plurality of device structure data, and the target process condition data includes a plurality of process condition data. When the target device structure data and the target process condition data are obtained, please refer to Figure 2 , the electronic device performs model training again based on the target device structure data and the target process condition data to obtain a second performance indicator prediction model (or device process structure fusion prediction model). The embodiment of the present invention uses the second performance indicator prediction model to construct the fitness function of the NSGA-Ⅱ algorithm. That is, the embodiment of the present invention uses the target device structure data and the target process condition data as the initial population. The second performance indicator prediction model is used as the fitness function of each individual in the initial population to calculate the fitness of each individual. By using the performance indicator prediction results of the second performance indicator prediction model as the fitness of individuals in the population, it is helpful to accelerate the evaluation process of individuals in the population, thereby improving the calculation speed.

[0096] The electronic device uses the data pairs consisting of the single device structure data in the target device structure data and the single process condition data in the target process condition data as individuals in the population, and uses the first performance prediction result output by the second performance indicator prediction model as the fitness of the individual, providing a basis for the subsequent non-dominated sorting and congestion calculation of the NSGA-Ⅱ algorithm.

[0097] The second performance indicator prediction model is obtained by training based on the target device structure data, the target process condition data, and the performance data label associated with the target device structure data and the target process condition data. The second performance indicator prediction model can adopt the same traditional machine learning model or neural network model as the first performance indicator prediction model. Among them, the traditional machine learning model used includes but is not limited to a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model, a lightweight gradient boosting machine model, etc. That is, the second performance indicator prediction model selects any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model, and a lightweight gradient boosting machine model. The second performance indicator prediction model is based on the target device structure data and the target process condition data as input data, and the performance data label is used as the label data for supervised training to obtain a trained second performance indicator prediction model. The performance data label may include one or more of the gate turn-on voltage Vtlin in the linear working area, the gate turn-on voltage Vtsat in the saturated working area, the drain current Idlin in the linear working area, the saturated drain current Idsat, the turn-off current Idoff, the breakdown voltage BVon in the on state, the breakdown voltage BVoff in the off state, and the substrate current Isub.

[0098] Step 400: construct a multi-objective function based on all individuals in the population.

[0099] The embodiment of the present invention constructs a multi-objective function based on a multi-objective optimization problem, that is, a function including multiple objective functions. The multi-objective function can be set according to actual conditions. For example, in one embodiment, assuming that under the premise of satisfying performance (Performance), it is necessary to simultaneously obtain two indicators of minimizing the leakage current Isub of the device and maximizing the breakdown voltage BVoff as the required performance indicators, then the multi-objective function as the optimization criterion can be expressed as:

[0100]

[0101] stg(X)≤0;

[0102] As shown in the above formula, f(X)=[f I (X), -f B (X)] represents the two objective functions of device performance (minimizing the leakage current of the device and maximizing the breakdown voltage), and Ω is the feasible region of X. g(X) is the performance constraint condition, that is, the replaced device still meets the performance index requirements. f1(X) is the minimum leakage current, -f B (X) is the maximum breakdown voltage, and f(X) is the multi-objective function.

[0103] Step 500: Solve the multi-objective function based on the NSGA-II algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0104] The electronic device uses the NSGA-Ⅱ algorithm to solve the multi-objective function to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions. In other embodiments, other multi-objective optimization algorithms such as a multi-objective evolutionary algorithm based on decomposition (MOEA-D) can also be used to solve the multi-objective function to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0105] The present invention trains a first performance index prediction model by fusing initial device structure data and initial process condition data, screens out target device structure data and target process condition data that affect each performance index of the device through the first performance index prediction model, and performs secondary model training based on the second performance index prediction model; then uses the NSGA-Ⅱ algorithm to solve according to a pre-set multi-objective function to obtain a Pareto front solution set as the optimization result of the device structure and process conditions. The embodiment of the present invention can significantly reduce the number of TCAD simulations and the number of actual process tape-outs during device structure optimization design, shorten design time and reduce costs, and accelerate device research and development progress.

[0106] In other aspects of the embodiments of the present invention, step 200, inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, includes:

[0107] Step 210: Input the initial device structure data and the initial process condition data into a first performance indicator prediction model for model training.

[0108] Step 230, determine the initial device structure data whose weight of the training result is greater than or equal to the set weight threshold as the target device structure data, and determine the initial process condition data whose weight of the training result is greater than or equal to the set weight threshold as the target process condition data.

[0109] The embodiment of the present invention needs to perform feature screening before model training. The electronic device can use the traditional machine learning model (such as the random forest model) to give a weight list (importance score list) of factors affecting each performance index of the device, and screen out important factors greater than or equal to the set weight threshold according to the weight, so as to perform model training again and obtain a second performance index prediction model (device process structure fusion prediction model). For example, the importance scores of initial device structure data a, initial device structure data b, initial device structure data c, initial device structure data d, initial device structure data e, initial process condition data f, initial process condition data g, initial process condition data h, initial process condition data i and initial process condition data j are 0.15, 0.22, 0.07, 0.03, 0.08, 0.11, 0.19, 0.09, 0.07, 0.09 respectively. The weight threshold is set to 0.15. The electronic device selects the initial device structure data a and the initial device structure data b corresponding to 0.15 and 0.22 as the target device structure data, and selects the initial process condition data g corresponding to 0.19 as the target process condition data. Thus, the embodiment of the present invention selects the target device structure data and the target process condition data of high importance from the initial device structure data and the initial process condition data, so as to perform model training again.

[0110] The embodiment of the present invention selects initial device structure data greater than or equal to a set weight threshold as the target device structure data, and determines initial process condition data whose weight of the training result is greater than or equal to the set weight threshold as the target process condition data. Feature screening is performed from the initial device structure data and the initial process condition data to select important factors of the device performance index, so as to improve the robustness of the performance index prediction through the second performance index prediction model.

[0111] In other aspects of the embodiments of the present invention, step 200, inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, includes:

[0112] Step 220: Input the initial device structure data and the initial process condition data into a plurality of first performance indicator prediction models respectively for model training.

[0113] Step 240 , performing an average operation based on the weight values ​​of the initial device structure data in a plurality of the first performance indicator prediction models to obtain a first average weight of the initial device structure data.

[0114] Step 260: Perform an averaging operation based on the weight values ​​of the initial process condition data in the plurality of the first performance indicator prediction models to obtain a second average weight of the initial process condition data.

[0115] Step 280, determine the initial device structure data whose first average weight is greater than or equal to the set weight threshold as the target device structure data, and determine the initial process condition data whose second average weight is greater than or equal to the set weight threshold as the target process condition data.

[0116] The embodiment of the present invention performs feature screening before model training. The electronic device can use a variety of machine learning models to sort the importance of features multiple times. By averaging, the initial device structure data and initial process condition data whose average weight is greater than or equal to the set weight threshold are determined, and used as the target device structure data and target process condition data for subsequent model training so as to perform model training again, and obtain a second performance indicator prediction model (device process structure fusion prediction model). For example, the embodiment of the present invention adds the initial device structure data and the initial process condition data to the random forest model, the extreme gradient boosting tree model and the gradient boosting decision tree model, respectively, calculates the weight average based on the weight of each model, and then compares it with the set weight threshold, and determines the initial device structure data and the initial process condition data greater than or equal to the set weight threshold as the target device structure data and the target process condition data.

[0117] For example, the initial device structure data a and the initial process condition data f are added to the random forest model, the extreme gradient boosting tree model and the gradient boosting decision tree model respectively, and the output weights of the random forest model, the extreme gradient boosting tree model and the gradient boosting decision tree model are 0.15, 0.22; 0.23, 0.25; 0.18, 0.23 respectively. Then the first average weight of the initial device structure data a is (0.15+0.23+0.18) / 3=0.19; then the second average weight of the initial process condition data f is (0.22+0.25+0.23) / 3=0.23. When the weight threshold is set to 0.20, the first average weight of the initial device structure data a is less than the set weight threshold, and it does not belong to the target device structure data. The second average weight of the initial process condition data f is greater than the set weight threshold, and it belongs to the target device structure data.

[0118] It should be noted that, in other embodiments, a weighted average operation may be performed based on the weight values ​​of the initial device structure data in multiple first performance indicator prediction models to obtain a first weighted average weight of the initial device structure data. A weighted average operation may be performed based on the weight values ​​of the initial process condition data in multiple first performance indicator prediction models to obtain a second weighted average weight of the initial process condition data. The first weighted average weight and the second weighted average weight are then compared with a set weight threshold.

[0119] The embodiment of the present invention calculates an average weight by combining the weights of multiple first performance indicator prediction models to screen out target device structure data and target process condition data. The embodiment of the present invention combines the importance scores of multiple first performance indicator prediction models to perform feature screening from initial device structure data and initial process condition data, further improving the accuracy of screening out important factors of device performance indicators, thereby further improving the robustness of performance indicator prediction through the second performance indicator prediction model.

[0120] In other aspects of the embodiments of the present invention, step 500, solving the multi-objective function based on the NSGA-II algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions, includes:

[0121] Repeat the following steps until the termination condition is set:

[0122] Step 510: Perform non-dominated sorting on all individuals in the population to divide all individuals in the population into multiple non-dominated layers.

[0123] Step 520: Calculate the crowding degree of each individual in the non-dominated layer.

[0124] Step 530: Generate a descendant population based on the selection operation, the crossover operation, and the mutation operation.

[0125] Step 540: merge the population and the offspring population to form a joint population.

[0126] Step 550: perform non-dominated sorting on all individuals in the joint population, and select individuals that meet set conditions from the joint population according to the non-dominated level and the crowding degree to form a new generation population.

[0127] The NSGA-Ⅱ algorithm adopts a fast non-dominated sorting algorithm, and its computational complexity is greatly reduced compared with NSGA. It also adopts crowding and crowding comparison operators to replace the need to specify the shared radius shareQ, and uses it as the winning criterion in the same-level comparison after quick sorting, so that the individuals in the quasi-Pareto domain can be expanded to the entire Pareto domain and evenly distributed, maintaining the diversity of the population. It also expands the sampling space by introducing the elite strategy, prevents the loss of the best individuals, and improves the algorithm's computing speed and robustness.

[0128] The non-dominated sorting is to stratify a population of size n by using a non-dominated sorting algorithm. Step 510, performing non-dominated sorting on all individuals in the population, includes:

[0129] Step 511, initialize the number of dominated individuals and the set of dominated individuals of each individual;

[0130] Step 512: traverse all individuals in the population, and update the number of dominating individuals and the set of dominated individuals of each individual according to the domination relationship;

[0131] Step 513, finding out all individuals that are not dominated by any other individuals, and classifying the individuals that are not dominated by any other individuals into the first layer;

[0132] Repeat the following steps until all individuals are sorted into the corresponding levels:

[0133] Step 514: Continue to find all the individuals that are not dominated by other individuals in the remaining unstratified individuals in the population and assign them to the second layer.

[0134] Specifically, the non-dominated sorting of all individuals in the population can be performed according to the following steps:

[0135] (1)Suppose i=1;

[0136] (2) For all j = 1, 2, ..., n, and j ≠ i, according to the above definition, compare the individual x i and individual x j the relationship of domination and non-domination between them;

[0137] (3) If there is no individual x j Better than x i , then x i Labeled as nondominant individuals;

[0138] (4) Let i = i + 1 and go to step (2) until all non-dominated individuals are found.

[0139] The set of non-dominated individuals obtained through the above steps is the first non-dominated layer of the population (i.e., the first layer described in step 513). Then, these marked non-dominated individuals are ignored (i.e., these individuals are no longer compared in the next round), and steps (1)-(4) are followed to obtain the second non-dominated layer (i.e., the second layer described in step 514). This process is repeated in this way until the entire population is stratified.

[0140] In order to obtain an estimate of the crowding of solutions around a specific solution in the population, we calculate the crowding of each objective function value (f1(X), -f B (X)) calculates the average distance between the two points on either side of this point (this individual). This value is used as an estimate of the perimeter of the cuboid with the nearest neighbors as vertices (called the crowding coefficient). Please refer to Figure 3 , the crowding coefficient of the ith solution on its frontier is the length of the cuboid around it (e.g. Figure 3 The calculation of crowding degree ensures the diversity of the population.

[0141] The calculation of the crowding coefficient requires sorting the population in ascending order according to the size of each objective function value (that is, if the first-level non-dominated layer is obtained, it is sorted according to the size of the objective function value, and then the crowding is calculated). Therefore, for each objective function value, the boundary solution (the solution with the maximum and minimum values) is assigned a value of infinite distance. All other intermediate solutions are assigned to be equal to the normalized absolute difference of the function values ​​of two adjacent solutions. The calculation method is the same for other objective function values. All crowding coefficient values ​​are calculated by adding the distance values ​​of each individual target, and each fitness is normalized before calculating the crowding coefficient.

[0142] In other aspects of the embodiments of the present invention, step 520, calculating the crowding degree of each individual in the non-dominated layer, includes:

[0143] Step 521, initializing the crowding degree of each individual;

[0144] Step 522, sorting the individuals in each non-dominated layer according to the objective function value;

[0145] Step 523: Set the crowding degree of the last sorted boundary individual in each non-dominated layer to infinity, and calculate the crowding degree of the non-boundary individuals in each non-dominated layer according to the difference in objective function values ​​of adjacent individuals.

[0146] Specifically, the congestion degree of each individual in the non-dominated layer is calculated by the following steps:

[0147] ① Initialize the congestion degree i of each point d Set to 0;

[0148] ② For each objective function value, perform non-dominated sorting on the population, making the crowding degree of the two boundary individuals at the boundary infinite;

[0149] ③Calculate the crowding degree of other non-boundary individuals:

[0150] Where: i d represents the congestion degree of point i, represents the jth objective function value at point i+1, Represents the jth objective function value of point i-1. m is the number of objective functions. Among them, the boundary individual refers to the individual at the front or back of the sorted individual sequence. For non-boundary individuals, calculate their crowding. The calculation of crowding is usually based on the difference in the objective function values ​​of adjacent individuals. The specific steps are as follows:

[0151] For each objective function k (k = 1, 2, ..., m, m is the number of objective functions): calculate the maximum value f of the objective function in the current non-dominated layer kmax and the minimum value f kmin For each non-boundary individual i, calculate the difference between it and its neighboring individuals on the objective function k, and divide it by the difference between the maximum and minimum values ​​of the objective function to obtain the contribution of the objective function to the crowding degree. Add up the contribution values ​​of all objective functions to the crowding degree to obtain the crowding degree of individual i.

[0152] After the previous fast non-dominated sorting and crowding calculation, each individual i in the population has two attributes: the non-dominated order i determined by the non-dominated sorting rank (level, i.e., which level) and congestion degree i d Based on these two properties, we can define the crowding comparison operator: when individual i is compared with another individual j, individual i wins as long as any of the following conditions is met.

[0153] ① If the non-dominated layer of individual i is better than the non-dominated layer of individual j, that is, i rank <j rank ;

[0154] ② If they have the same rank, and individual i has a larger crowding distance than individual j, that is, i rank =j rank And i d >j d The first condition ensures that the selected individual belongs to the better non-inferior class. The second condition selects the individual that is in the less crowded area (with a larger crowding degree i) of two individuals that are tied in the same non-inferior class according to their crowding distance. d ). The winning individual enters the next operation.

[0155] Please refer to Figure 4 Next, we first generate the new population Q of the tth generation t , and the parent P t Merge to form population R t , the population size is 2N. Then R t Perform non-dominated sorting to generate a series of non-dominated sets Z1 and calculate the crowding degree. Since both the offspring and parent individuals are included in R t Then the non-dominated set Z1 after non-dominated sorting is the best, so first add Z1 to P t+1 In the new parent population. If the size is less than N, continue to P t+1 , add Z2 to the population. Until Z3 is added, the population size exceeds N, then it is selected and added according to the crowding comparison operator, so that P t+1 The number of individuals reaches N. Then a new offspring population Q is generated through genetic operators (selection, crossover, mutation) t+1 .

[0156] When faced with a multi-objective function such as f(X) = [f I (X), -f B When [X] represents the two objective functions of device performance (minimizing leakage current and maximizing breakdown voltage), f1(X) is the minimum leakage current. B (X) is the maximum breakdown voltage. Therefore, there is a certain trade-off between these two performance indicators. This trade-off can be well described by the boundary of the feasible region Ω of PF. Please refer to Figure 5 , the horizontal and vertical axes represent the values ​​of the two objective functions, and the f corresponding to the solution represented by point E I (X) and -f B (X) must be smaller than the objective function value corresponding to the solution represented by C and D points, then E is said to dominate C and D. Figure 5 The solid points (such as A and B) are Pareto optimal points, and these solid points constitute the Pareto frontier solution set.

[0157] In optimization problems involving two objectives, its Pareto front (PF) is usually a line. However, when dealing with optimization problems with multiple objectives, the Pareto front is usually presented as a hypersurface. This is because in multi-objective optimization problems, there are multiple conflicting objective functions, so that different solutions may have different advantages and disadvantages on different objectives. Therefore, the Pareto front solution set often contains multiple non-dominated solutions, forming a high-dimensional solution space. Such a solution space usually cannot be represented in a simple linear way, but is presented as a complex hypersurface structure. This also makes the solution of multi-objective optimization problems more complex and challenging, because it is necessary to find a set of solutions on this high-dimensional hypersurface that cannot be dominated by other solutions on all objectives, thus forming a Pareto front solution set.

[0158] In summary, the present invention trains the first performance index prediction model by fusing the initial device structure data and the initial process condition data, screens out the target device structure data and target process condition data that affect each performance index of the device through the first performance index prediction model, and performs secondary model training based on the second performance index prediction model; then the NSGA-Ⅱ algorithm is used to solve according to the pre-set multi-objective function to obtain the Pareto front solution set as the optimization result of the device structure and process conditions. The embodiment of the present invention can significantly reduce the number of TCAD simulations and the number of actual process tape-outs during device structure optimization design, shorten the design time and reduce costs, and accelerate the progress of device research and development.

[0159] Device Embodiment

[0160] Please refer to Figure 6 On the other hand, an embodiment of the present invention further provides a device structure and process condition optimization device, comprising:

[0161] An acquisition module 601 is used to acquire initial device structure data and initial process condition data;

[0162] A screening module 602 is used to input the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screen out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results; the target device structure data includes a plurality of device structure data, and the target process condition data includes a plurality of process condition data;

[0163] A determination module 603 is used to determine a data pair consisting of first device structure data and first process condition data as an individual in a population, and to determine a first performance prediction result as the fitness of the individual; the first device structure data is any one of the multiple device structure data, and the first process condition data is any one of the multiple process condition data; the first performance prediction result is obtained by inputting the first device structure data and the first process condition data into a second performance indicator prediction model; the second performance indicator prediction model is obtained by training based on the target device structure data, the target process condition data, and the performance data label associated with the target device structure data and the target process condition data;

[0164] A construction module 604, configured to construct a multi-objective function based on all individuals in the population;

[0165] The solution module 605 is used to solve the multi-objective function based on the NSGA-II algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0166] Optionally, the inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises:

[0167] Inputting the initial device structure data and the initial process condition data into a first performance indicator prediction model for model training;

[0168] Determine the initial device structure data whose weight of the training result is greater than or equal to the set weight threshold as the target device structure data, and determine the initial process condition data whose weight of the training result is greater than or equal to the set weight threshold as the target process condition data.

[0169] Optionally, the inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises:

[0170] Inputting the initial device structure data and the initial process condition data into a plurality of first performance indicator prediction models for model training;

[0171] Performing an average operation based on weight values ​​of the initial device structure data in a plurality of the first performance indicator prediction models to obtain a first average weight of the initial device structure data;

[0172] Performing an average operation based on weight values ​​of the initial process condition data in a plurality of the first performance indicator prediction models to obtain a second average weight of the initial process condition data;

[0173] Initial device structure data whose first average weight is greater than or equal to a set weight threshold is determined as the target device structure data, and initial process condition data whose second average weight is greater than or equal to a set weight threshold is determined as the target process condition data.

[0174] Optionally, solving the multi-objective function based on the NSGA-II algorithm to obtain a Pareto front solution set of the population as an optimization result of the device structure and process conditions includes:

[0175] Repeat the following steps until the termination condition is set:

[0176] performing non-dominated sorting on all individuals in the population to divide all individuals in the population into a plurality of non-dominated layers;

[0177] Calculate the crowding degree of each individual in the non-dominated layer;

[0178] Generate offspring population based on selection, crossover and mutation operations;

[0179] Merging the population and the progeny population to form a combined population;

[0180] All individuals in the joint population are non-dominated and sorted, and individuals that meet set conditions are selected from the joint population according to the non-dominated level and the crowding degree to form a new generation population.

[0181] Optionally, performing non-dominated sorting on all individuals in the population includes:

[0182] Initialize the number of dominated individuals and the set of dominated individuals of each individual;

[0183] Traverse all individuals in the population and update the number of dominant individuals and the set of dominated individuals of each individual according to the dominance relationship;

[0184] Find out all individuals that are not dominated by any other individuals, and classify the individuals that are not dominated by any other individuals into the first layer;

[0185] Repeat the following steps until all individuals are sorted into the corresponding levels:

[0186] Continue to find all the individuals that are not dominated by other individuals among the remaining unstratified individuals in the population and assign them to the second layer.

[0187] Optionally, calculating the crowding degree of each individual in the non-dominated layer includes:

[0188] Initialize the crowding degree of each individual;

[0189] Sort the individuals in each non-dominated layer by the objective function value;

[0190] The crowding degree of the last sorted boundary individual in each non-dominated layer is set to infinity, and the crowding degree of the non-boundary individuals in each non-dominated layer is calculated according to the difference in the objective function values ​​of adjacent individuals.

[0191] Optionally, the initial device structure data includes at least one of gate width, gate length, number of fingers and pitch; the initial process condition data includes at least one of deposition rate and deposition temperature.

[0192] Optionally, the first performance indicator prediction model selects any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model and a lightweight gradient boosting machine model.

[0193] Optionally, the second performance indicator prediction model selects any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model and a lightweight gradient boosting machine model.

[0194] The device structure and process condition optimization device includes a processor and a memory. The above-mentioned acquisition module 601, screening module 602, determination module 603, construction module 604 and solution module 605 are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize corresponding functions.

[0195] The processor includes a kernel, which calls the corresponding program unit from the memory. There can be one or more kernels.

[0196] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0197] Figure 7 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 7As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the device structure and process condition optimization method, the method comprising: obtaining initial device structure data and initial process condition data; inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training result; the target device structure data includes a plurality of device structure data, and the target process condition data includes a plurality of process condition data; determining a data pair consisting of the first device structure data and the first process condition data as an individual in the population, and determining a first performance prediction result as the individual fitness; the first device structure data is any one of the multiple device structure data, and the first process condition data is any one of the multiple process condition data; the first performance prediction result is obtained after the first device structure data and the first process condition data are input into the second performance indicator prediction model; the second performance indicator prediction model is trained based on the target device structure data, the target process condition data and the performance data label associated with the target device structure data and the target process condition data; a multi-objective function is constructed based on all individuals in the population; the multi-objective function is solved based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0198] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0199] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a machine-readable storage medium. When the computer program is executed by a processor, the computer can execute a device structure and process condition optimization method, which includes: obtaining initial device structure data and initial process condition data; inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results; the target device structure data includes multiple device structure data, and the target process condition data includes multiple process condition data; determining a data pair consisting of the first device structure data and the first process condition data as individuals in a population, and determining a first performance prediction result as the fitness of the individual; the first device structure data is any one of the multiple device structure data, and the first process condition data is any one of the multiple process condition data; the first performance prediction result is obtained after the first device structure data and the first process condition data are input into a second performance indicator prediction model; the second performance indicator prediction model is trained based on the target device structure data, the target process condition data, and the performance data labels associated with the target device structure data and the target process condition data; a multi-objective function is constructed based on all individuals in the population; the multi-objective function is solved based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0200] On the other hand, the present invention also provides a machine-readable storage medium having a computer program stored thereon, which is implemented when the processor executes the device structure and process condition optimization method, the method comprising: obtaining initial device structure data and initial process condition data; inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results; the target device structure data includes multiple device structure data, and the target process condition data includes multiple process condition data; determining a data pair consisting of the first device structure data and the first process condition data as an individual in a population, and determining the first device structure data and the first process condition data as an individual in a population. A performance prediction result is used as the fitness of the individual; the first device structure data is any one of the multiple device structure data, and the first process condition data is any one of the multiple process condition data; the first performance prediction result is obtained after the first device structure data and the first process condition data are input into a second performance indicator prediction model; the second performance indicator prediction model is trained based on the target device structure data, the target process condition data, and the performance data labels associated with the target device structure data and the target process condition data; a multi-objective function is constructed based on all individuals in the population; the multi-objective function is solved based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

[0201] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0202] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0203] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing device structure and process conditions, characterized in that: include: Acquire initial device structure data and initial process condition data; Inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results; The target device structure data includes a plurality of device structure data, and the target process condition data includes a plurality of process condition data; Determine a data pair consisting of first device structure data and first process condition data as an individual in a population, and determine a first performance prediction result as the fitness of the individual; The first device structure data is any one of the plurality of device structure data, and the first process condition data is any one of the plurality of process condition data; The first performance prediction result is obtained by inputting the first device structure data and the first process condition data into a second performance indicator prediction model; The second performance indicator prediction model is obtained by training based on the target device structure data, the target process condition data, and the performance data label associated with the target device structure data and the target process condition data; constructing a multi-objective function based on all individuals in the population; The multi-objective function is solved based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

2. The device structure and process condition optimization method according to claim 1, characterized in that: The step of inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises: Inputting the initial device structure data and the initial process condition data into a first performance indicator prediction model for model training; Determine the initial device structure data whose weight of the training result is greater than or equal to the set weight threshold as the target device structure data, and determine the initial process condition data whose weight of the training result is greater than or equal to the set weight threshold as the target process condition data.

3. The device structure and process condition optimization method according to claim 1, characterized in that: The step of inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises: Inputting the initial device structure data and the initial process condition data into a plurality of first performance indicator prediction models for model training; Performing an average operation based on weight values ​​of the initial device structure data in a plurality of the first performance indicator prediction models to obtain a first average weight of the initial device structure data; Performing an average operation based on weight values ​​of the initial process condition data in a plurality of the first performance indicator prediction models to obtain a second average weight of the initial process condition data; Initial device structure data whose first average weight is greater than or equal to a set weight threshold is determined as the target device structure data, and initial process condition data whose second average weight is greater than or equal to a set weight threshold is determined as the target process condition data.

4. The device structure and process condition optimization method according to claim 1, characterized in that: The multi-objective function is solved based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions, including: Repeat the following steps until the termination condition is set: performing non-dominated sorting on all individuals in the population to divide all individuals in the population into a plurality of non-dominated layers; Calculate the crowding degree of each individual in the non-dominated layer; Generate offspring population based on selection, crossover and mutation operations; Merging the population and the progeny population to form a combined population; All individuals in the joint population are non-dominated and sorted, and individuals that meet set conditions are selected from the joint population according to the non-dominated level and the crowding degree to form a new generation population.

5. The device structure and process condition optimization method according to claim 4, characterized in that: The non-dominated sorting of all individuals in the population comprises: Initialize the number of dominated individuals and the set of dominated individuals of each individual; Traverse all individuals in the population and update the number of dominant individuals and the set of dominated individuals of each individual according to the dominance relationship; Find out all individuals that are not dominated by any other individuals, and classify the individuals that are not dominated by any other individuals into the first layer; Repeat the following steps until all individuals are sorted into the corresponding levels: Continue to find all the individuals that are not dominated by other individuals among the remaining unstratified individuals in the population and assign them to the second layer.

6. The device structure and process condition optimization method according to claim 4, characterized in that: The calculating the crowding degree of each individual in the non-dominated layer includes: Initialize the crowding degree of each individual; Sort the individuals in each non-dominated layer by the objective function value; The crowding degree of the last sorted boundary individual in each non-dominated layer is set to infinity, and the crowding degree of the non-boundary individuals in each non-dominated layer is calculated according to the difference in the objective function values ​​of adjacent individuals.

7. The device structure and process condition optimization method according to claim 1, characterized in that: The initial device structure data includes at least one of gate width, gate length, number of fingers and pitch; the initial process condition data includes at least one of deposition rate and deposition temperature.

8. The device structure and process condition optimization method according to claim 1, characterized in that: The first performance indicator prediction model is selected from any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model, and a lightweight gradient boosting machine model.

9. The device structure and process condition optimization method according to claim 1, characterized in that: The second performance indicator prediction model is selected from any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model, and a lightweight gradient boosting machine model.

10. A device structure and process condition optimization device, characterized in that: include: An acquisition module, used for acquiring initial device structure data and initial process condition data; A screening module, used for inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results; The target device structure data includes a plurality of device structure data, and the target process condition data includes a plurality of process condition data; A determination module, configured to determine a data pair consisting of first device structure data and first process condition data as an individual in a population, and to determine a first performance prediction result as the fitness of the individual; The first device structure data is any one of the plurality of device structure data, and the first process condition data is any one of the plurality of process condition data; The first performance prediction result is obtained by inputting the first device structure data and the first process condition data into a second performance indicator prediction model; The second performance indicator prediction model is obtained by training based on the target device structure data, the target process condition data, and the performance data label associated with the target device structure data and the target process condition data; A construction module, used for constructing a multi-objective function based on all individuals in the population; The solution module is used to solve the multi-objective function based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions.

11. The device structure and process condition optimization device according to claim 10, characterized in that: The step of inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises: Inputting the initial device structure data and the initial process condition data into a first performance indicator prediction model for model training; Determine the initial device structure data whose weight of the training result is greater than or equal to the set weight threshold as the target device structure data, and determine the initial process condition data whose weight of the training result is greater than or equal to the set weight threshold as the target process condition data.

12. The device structure and process condition optimization device according to claim 10, characterized in that: The step of inputting the initial device structure data and the initial process condition data into one or more first performance indicator prediction models for model training, and screening out target device structure data and target process condition data based on the weight of the initial device structure data and the weight of the initial process condition data in the training results, comprises: Inputting the initial device structure data and the initial process condition data into a plurality of first performance indicator prediction models for model training; Performing an average operation based on weight values ​​of the initial device structure data in a plurality of the first performance indicator prediction models to obtain a first average weight of the initial device structure data; Performing an average operation based on weight values ​​of the initial process condition data in a plurality of the first performance indicator prediction models to obtain a second average weight of the initial process condition data; Initial device structure data whose first average weight is greater than or equal to a set weight threshold is determined as the target device structure data, and initial process condition data whose second average weight is greater than or equal to a set weight threshold is determined as the target process condition data.

13. The device structure and process condition optimization device according to claim 10, characterized in that: The multi-objective function is solved based on the NSGA-Ⅱ algorithm to obtain the Pareto front solution set of the population as the optimization result of the device structure and process conditions, including: Repeat the following steps until the termination condition is set: performing non-dominated sorting on all individuals in the population to divide all individuals in the population into a plurality of non-dominated layers; Calculate the crowding degree of each individual in the non-dominated layer; Generate offspring population based on selection, crossover and mutation operations; Merging the population and the progeny population to form a combined population; All individuals in the joint population are non-dominated and sorted, and individuals that meet set conditions are selected from the joint population according to the non-dominated level and the crowding degree to form a new generation population.

14. The device structure and process condition optimization device according to claim 13, characterized in that: The non-dominated sorting of all individuals in the population comprises: Initialize the number of dominated individuals and the set of dominated individuals of each individual; Traverse all individuals in the population and update the number of dominant individuals and the set of dominated individuals of each individual according to the dominance relationship; Find out all individuals that are not dominated by any other individuals, and classify the individuals that are not dominated by any other individuals into the first layer; Repeat the following steps until all individuals are sorted into the corresponding levels: Continue to find all the individuals that are not dominated by other individuals among the remaining unstratified individuals in the population and assign them to the second layer.

15. The device structure and process condition optimization device according to claim 13, characterized in that: The calculating the crowding degree of each individual in the non-dominated layer includes: Initialize the crowding degree of each individual; Sort the individuals in each non-dominated layer by the objective function value; The crowding degree of the last sorted boundary individual in each non-dominated layer is set to infinity, and the crowding degree of the non-boundary individuals in each non-dominated layer is calculated according to the difference in the objective function values ​​of adjacent individuals.

16. The device structure and process condition optimization device according to claim 10, characterized in that: The initial device structure data includes at least one of gate width, gate length, number of fingers and pitch; the initial process condition data includes at least one of deposition rate and deposition temperature.

17. The device structure and process condition optimization device according to claim 10, characterized in that: The first performance indicator prediction model is selected from any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model, and a lightweight gradient boosting machine model.

18. The device structure and process condition optimization device according to claim 10, characterized in that: The second performance indicator prediction model is selected from any one of a random forest model, an extreme gradient boosting tree model, a gradient boosting decision tree model, and a lightweight gradient boosting machine model.

19. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the device structure and process condition optimization method described in any one of claims 1 to 9 is implemented.

20. A machine-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the device structure and process condition optimization method described in any one of claims 1 to 9 is implemented.

21. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the device structure and process condition optimization method described in any one of claims 1 to 9 is implemented.

Images