CN-121983190-A - Optimization method and device for laminated material plasma etching process

CN121983190ACN 121983190 ACN121983190 ACN 121983190ACN-121983190-A

Abstract

The invention discloses a method and a device for optimizing a laminated material plasma etching process, relates to the technical field of semiconductor simulation, and aims to solve the problems of low efficiency, insufficient precision and poor universality in the simulation of laminated materials in the prior art. The optimization method comprises the steps of obtaining physical parameters to be optimized of laminated materials in a plasma etching process and a value space of the physical parameters, carrying out initial sampling on the value space based on a Bayesian optimization method and a first Monte Carlo simulation method to obtain initial sampling data, training a neural network proxy model based on the initial sampling data, selecting candidate points through an acquisition function to obtain a candidate point set, adopting a second Monte Carlo simulation method to process the candidate point set and obtain a loss function value set, carrying out optimization iteration based on the loss function value set until a preset condition is met to obtain a target physical parameter combination, and enabling the preset condition to achieve preset optimization round or loss function value convergence.

Inventors

CHEN RUI
SONG HONGLI

Assignees

北京知识产权运营管理有限公司
中国科学院微电子研究所

Dates

Publication Date: 20260505
Application Date: 20251219

Claims (10)

1. The optimization method of the laminated material plasma etching process is characterized by comprising the following steps: Obtaining physical parameters to be optimized of the laminated material in a plasma etching process and a value space of the physical parameters; performing initial sampling on the value space based on a Bayesian optimization method and a first Monte Carlo simulation method to obtain initial sampling data; Training a neural network agent model based on the initial sampling data and selecting candidate points through an acquisition function to obtain a candidate point set; Processing the candidate point set by adopting a second Monte Carlo simulation method and obtaining a loss function value set; And carrying out optimization iteration based on the loss function value set until a preset condition is met, so as to obtain a target physical parameter combination, wherein the preset condition is that a preset optimization round or loss function value convergence is achieved.
2. The method for optimizing a stacked material plasma etching process according to claim 1, wherein the initial sampling of the value space based on a bayesian optimization method and a first monte carlo simulation method to obtain initial sampling data comprises: n rounds of uniform sampling are carried out in the value taking space, and each round of sampling generates M initial parameter combinations, wherein each initial parameter combination comprises specific numerical values of physical parameters, and N and M are both larger than zero; for M initial parameter combinations, performing etching simulation by adopting the first Monte Carlo simulation method to obtain simulation results of each initial parameter combination; Analyzing the simulation result of each initial parameter combination, and extracting simulation values of measurement indexes related to etching; And constructing the initial sampling data based on the simulation value and the real electron microscope characterization data.
3. The method of optimizing a plasma etching process of a laminate material according to claim 2, wherein the measurement indicators include an etching maximum depth, a difference between the maximum width and the etched opening, a depth at which the maximum width is located, a bottom width, and a bottom bump height; based on the simulation value and the real electron microscope characterization data, constructing the initial sampling data comprises the following steps: determining the true value of each measurement index based on the true electron microscope characterization data; Substituting the simulation value and the true value of each measurement index into a mean square error function calculation formula: obtaining a loss function value corresponding to the single initial parameter combination, wherein, -A value of said loss function; For measuring the number of indicators; Is the first Simulation values of the individual measurement indexes; Is the first The true value of each measurement index; And constructing the initial sampling data consisting of the initial parameter combinations and the loss function values by corresponding the initial parameter combinations and the corresponding loss function values one by one.
4. The method of optimizing a laminate material plasma etching process of claim 3, further comprising, prior to training a neural network proxy model based on the initial sample data, constructing the neural network proxy model; The construction of the neural network proxy model comprises the step of integrating a plurality of feedforward neural networks in parallel to obtain the neural network proxy model.
5. The method of optimizing a stacked material plasma etching process of claim 4, wherein prior to parallel integration of a plurality of feedforward neural networks, said method further comprises constructing a single feedforward neural network; constructing a single feed-forward neural network includes: The feedforward neural network is obtained by sequentially setting an input layer, a first full-connection layer, a first RELU activation function, a first Dropout regularization module, a second full-connection layer, a second RELU activation function, a second Dropout regularization module, a third full-connection layer, a third RELU activation function and an output layer, wherein the output layer comprises a fourth full-connection layer and an output module.
6. The method of claim 4, wherein training a neural network proxy model based on the initial sample data and selecting candidate points by an acquisition function to obtain a set of candidate points comprises: Training the neural network agent model by utilizing the initial sampling data and combining a learning rate scheduler and an early stopping mechanism to obtain a trained model; randomly sampling from the value space to generate a plurality of potential parameter combinations to form a candidate pool; Inputting each potential parameter combination in the candidate pool into the trained model to obtain a prediction mean value and prediction uncertainty of each potential parameter combination, wherein the prediction mean value is the mean value of a plurality of feedforward neural network loss value prediction results, and the prediction uncertainty is the standard deviation of the feedforward neural network loss value prediction results; and inputting the prediction mean value and the prediction uncertainty of each potential parameter combination into the acquisition function, and generating a plurality of non-overlapping candidate points by combining Constant Liar strategies to form the candidate point set.
7. The method of claim 6, wherein inputting the predicted mean and predicted uncertainty for each potential parameter combination to the collection function and combining Constant Liar strategies to generate a plurality of non-overlapping candidate points, comprising the candidate point set, comprises: The prediction mean and prediction uncertainty for each potential parameter combination is input to the EI acquisition function: calculating EI function value of each potential parameter combination, and screening out potential parameter combination with maximum EI function value as first candidate point, ; A cumulative distribution function that is a standard normal distribution; Probability density function of standard normal distribution; Is a prediction mean value; Is the optimal loss function value; Non-negative hyper-parameters for exploration and utilization of balance for regulation; To predict uncertainty; Based on the Constant Liar strategy, determining an optimal loss function value in the point selection initial data as a virtual loss function value of the first candidate point, and supplementing the first candidate point and the corresponding virtual loss function value as a first virtual sample to the point selection initial data to form first virtual update data; screening out a second candidate point through the EI acquisition function based on the first virtual update data; Determining an optimal loss function value in the first virtual update data as a virtual loss function value of the second candidate point based on the Constant Liar strategy, and supplementing the second candidate point and the corresponding virtual loss function value into the first virtual update data to form second virtual update data; and screening out the last candidate point by using the EI acquisition function to obtain the candidate point set.
8. The method of claim 6, wherein processing the candidate set of points and obtaining a set of loss function values using a second monte carlo simulation method comprises: executing etching simulation by adopting the second Monte Carlo simulation method aiming at each candidate point in the candidate point set to obtain a simulation result of each candidate point; Analyzing the simulation result of each candidate point, and extracting the simulation value of the measurement index; calculating a loss function value of each candidate point by combining the true value of each measurement index based on the mean square error function calculation formula; And integrating the loss function values of all the candidate points to form the loss function value set.
9. The method of claim 8, wherein performing optimization iterations based on the set of loss function values comprises: the method comprises the steps of obtaining a candidate point set, a history sampling data, a real accumulated sampling data, a historical sampling data and a calculation result, wherein the candidate point set and the corresponding loss function value are supplemented to the history sampling data to obtain real accumulated sampling data; retraining a neural network proxy model based on the real accumulated sampling data to obtain updated model output; And determining an updated candidate point set through the acquisition function and the Constant Liar strategy based on the updated model output, processing the updated candidate point set through the second Monte Carlo simulation method to obtain an updated loss function value set, and continuing to perform optimization iteration based on the updated loss function value set until a preset condition is met.
10. An optimizing device for a laminated material plasma etching process, which is characterized by comprising: The acquisition module is used for acquiring physical parameters to be optimized of the laminated material in the plasma etching process and a value space of the physical parameters; the sampling module is used for carrying out initial sampling on the value space based on a Bayesian optimization method and a first Monte Carlo simulation method to obtain initial sampling data; the acquisition module is used for training a neural network agent model based on the initial sampling data and selecting candidate points through an acquisition function to obtain a candidate point set; the processing module is used for processing the candidate point set by adopting a second Monte Carlo simulation method and obtaining a loss function value set; And the iteration module is used for carrying out optimization iteration based on the loss function value set until a preset condition is met to obtain a target physical parameter combination, wherein the preset condition is that a preset optimization round is reached or the loss function value converges.

Description

Optimization method and device for laminated material plasma etching process Technical Field The invention relates to the technical field of semiconductor simulation, in particular to a method and a device for optimizing a laminated material plasma etching process. Background In the field of integrated circuit manufacturing, the laminated material has become a core foundation for coping with challenges of process miniaturization and structure complexity by virtue of the advantages of the laminated material in cooperation with multi-material characteristics, for example, the Si/SiGe laminated material is grown alternately by periodicity of Si layers and SiGe layers, and the functional division of a carrier channel and a sacrificial layer is realized by utilizing the difference of lattice mismatch and etching selectivity, so that the laminated material is widely applied to advanced device preparation. The plasma etching is used as a key process of pattern transfer, substrate atoms are removed through a physical and chemical cooperative mechanism, and the precise optimization of process parameters directly determines the quality of an etching section and the performance of a device. The simulation optimization of the current laminated material plasma etching process mainly depends on three methods of manual tuning, experimental verification and atomic level simulation, and the methods have obvious technical bottlenecks that firstly, the manual tuning is excessively dependent on the history experience and data accumulation of engineers, non-physical solutions are easily generated due to subjective deviation, the optimization process is long in time consumption and is easily trapped in local optimization, nonlinear response caused by new materials or gas proportion change cannot be adapted, secondly, the experimental verification shows that the result is accurate, but the single etching needs a special reactor and electron microscope for characterization, the cost is high, the resolution is difficult to quantify the sub-nano etching mechanism, the data acquisition period is seriously delayed by the process development rhythm, and thirdly, the atomic level simulation is extremely slow in operation speed and cannot efficiently finish the targeted tuning of physical parameters by means of molecular dynamics, first principle and the like. The problems directly lead to out-of-control of the etched characteristic profile, the defects of neck blockage, overlarge etching depth deviation and the like, and the defects of complete failure in innovative application scenes such as three-dimensional memory devices and the like, so that a high-efficiency, accurate and general etching process parameter optimization scheme is needed. Disclosure of Invention The invention aims to provide an optimization method and device for a laminated material plasma etching process, which are used for solving the problems of low efficiency, insufficient precision and poor universality in the simulation of laminated materials in the prior art. In order to achieve the above object, the present invention provides the following technical solutions: In a first aspect, the present invention provides a method for optimizing a laminate material plasma etching process, comprising: Obtaining physical parameters to be optimized of the laminated material in a plasma etching process and a value space of the physical parameters; Performing initial sampling on the value space based on a Bayesian optimization method and a first Monte Carlo simulation method to obtain initial sampling data; training a neural network agent model based on initial sampling data, and selecting candidate points through an acquisition function to obtain a candidate point set; Processing the candidate point set by adopting a second Monte Carlo simulation method and obtaining a loss function value set; And carrying out optimization iteration based on the loss function value set until a preset condition is met, so as to obtain a target physical parameter combination, wherein the preset condition is that a preset optimization round is reached or the loss function value is converged. Optionally, performing initial sampling on the value space based on a bayesian optimization method and a first monte carlo simulation method to obtain initial sampling data, including: Uniformly sampling N rounds in a value space, wherein each round of sampling generates M initial parameter combinations, each initial parameter combination comprises specific numerical values of each physical parameter, and N and M are both larger than zero; for M initial parameter combinations, performing etching simulation by adopting a first Monte Carlo simulation method to obtain a simulation result of each initial parameter combination; analyzing the simulation result of each initial parameter combination, and extracting simulation values of measurement indexes related to etching; And constructing initial sampling data ba