Method, device and equipment for estimating reaction etching selection ratio of laminated material

Abstract

The invention discloses a method, a device and equipment for estimating a reaction etching selection ratio of a laminated material, relates to the technical field of etching in microelectronic manufacturing, and is used for solving the problem that a micro-scale etching selection ratio is difficult to efficiently obtain in the prior art. Comprising the steps of obtaining a basic data set of a laminated material; integrating and screening the basic data set to obtain an initial sample; based on the initial sample, optimizing the basic data set by adopting an iterative optimization mechanism of a machine learning dynamic situation function of active learning iterative sampling to obtain a final potential function; and performing molecular dynamics simulation based on the final potential function, and estimating the reaction etching selection ratio of the laminated material in combination with an adaptive time step algorithm. According to the invention, a high-precision machine learning potential function can be formed and a dynamic model of an etching process can be completed; and quantitatively calculating the reaction rate difference of each layer of material, and deducing the etching selection ratio.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of etching technology in microelectronic manufacturing, and in particular to a method, device and equipment for estimating the reaction etching selectivity of a laminated material. BACKGROUND

[0002] There are mainly the following simulation modeling methods for etching: firstly, a simple model based on geometric topology. This kind of model has low accuracy and cannot completely match the actual etching process. Secondly, a characteristic scale model using Monte Carlo and finite element methods has been applied to vertical etching processes, but has not been effectively applied to lateral selective etching processes. Finally, there are modeling methods based on the microscopic process of atomic motion. One method is based on quantum chemical calculation of etching reaction activation energy and estimates the reaction probability by combining the Arrhenius formula. This method can only provide qualitative results and cannot be used as a quantitative basis for discrimination. Another method is based on molecular dynamics simulation of atomic motion processes, which can accurately reproduce atomic-level reactions in the etching process, but has problems of long calculation time and large resource consumption in the simulation of large-scale systems. In recent years, machine learning potential functions have been developed as a supplementary means to make up for the low accuracy of traditional molecular dynamics simulation and the low efficiency of density functional theory calculations. The accuracy and effectiveness of machine learning potential functions are closely related to the quality of the training data set. At present, due to the difficulty in obtaining experimental data, it is difficult to obtain a large amount of high-quality data. The development of potential function training methods based on DFT and high-throughput technology has become the mainstream paradigm. However, high-throughput calculations based on density functional theory (DFT) still face problems such as high computational cost and redundant data dimensions, making it difficult to meet the demand for large-scale high-quality data sets for machine learning potential functions. In addition, traditional random sampling methods often ignore the representativeness of sample distribution when selecting training data, resulting in insufficient prediction ability of the model in complex potential energy surfaces or local extremum regions. In particular, in the multi-phase interface and heterostructure of laminated materials, traditional methods are difficult to accurately capture the material-specific dynamics behavior in the etching process, thereby affecting the estimation accuracy of the key process parameter of etching selectivity.

[0003] Therefore, there is an urgent need to provide a more reliable scheme for estimating the reaction etching selectivity of a laminated material. SUMMARY

[0004] The present application aims to provide a method, device and equipment for estimating the reaction etching selectivity of a laminated material, which solves the problem of difficulty in efficiently obtaining microscale etching selectivity in the prior art.

[0005] To achieve the above object, the present application provides the following technical scheme:

[0006] In a first aspect, the present application provides a method for estimating the reaction etching selectivity of a laminated material, the method comprising:

[0007] obtaining a basic data set of the laminated material; the basic data set at least includes a key reaction configuration, an energy barrier and an activation parameter;

[0008] performing integrated screening on the basic data set to obtain an initial sample;

[0009] based on the initial sample, using an iterative optimization mechanism of a machine learning dynamic potential function with active learning iterative sampling to optimize the basic data set to obtain a final potential function;

[0010] based on the final potential function, performing molecular dynamics simulation and combining an adaptive time step algorithm to estimate a reaction etching selectivity ratio of the laminated material.

[0011] Optionally, based on the initial sample, using an iterative optimization mechanism of a machine learning dynamic potential function with active learning iterative sampling to optimize the basic data set to obtain a final potential function, comprising:

[0012] training an initial neural evolution potential function based on the initial sample, and generating an extrapolated configuration set by introducing a multi-dimensional ensemble condition; the multi-dimensional ensemble condition at least includes a temperature gradient disturbance, a pressure fluctuation and a random noise injection;

[0013] establishing a structure differentiation evaluation function, comparing atomic displacement matrices of the extrapolated configuration set and the initial sample, and screening abnormal configurations deviating from a ground state distribution;

[0014] performing DFT accuracy verification on the abnormal configurations, fusing energy convergence results with an original data set to obtain an enhanced training set;

[0015] using a potential function training module to perform multiple rounds of potential function iteration on the enhanced training set, triggering a convergence condition when a differentiation rate of the extrapolated configuration set is lower than a first preset threshold for three consecutive times, and outputting a final potential function.

[0016] Optionally, based on the final potential function, performing molecular dynamics simulation and combining an adaptive time step algorithm to estimate a reaction etching selectivity ratio of the laminated material, comprising:

[0017] constructing interface reaction path information of different laminations;

[0018] constructing a machine learning potential function of different lamination systems according to the interface reaction path information and a classical potential function strategy;

[0019] generating a dedicated potential function using an active learning method;

[0020] deduce through a laminated machine learning potential function;

[0021] Calculate etching output and estimate reactive etching selectivity.

[0022] Optionally, the basic data set is integrated and screened to obtain an initial sample, including:

[0023] Using the CINEB method, a basic dataset containing key reaction configurations, energy barriers, and activation parameters was constructed through first-principles calculations. This basic dataset enabled three-dimensional visualization of the reaction pathway through a preset platform, yielding high-confidence initial samples for potential function training.

[0024] Optionally, construct interface reaction path information for different stacks, including:

[0025] Define the initial state and final state in the absence and presence of reactants, respectively, and establish a substrate model;

[0026] The NEB method is used to calculate saddle points on the reaction path. Each saddle point on the reaction path is evaluated to determine whether it meets the convergence criteria. The lowest energy path of the reaction path is found and the key transition state is identified. The atomic information of the reaction path is output; the atomic information includes the initial and final states of the reaction process and the existing saddle point structures of multiple transition states.

[0027] Optionally, machine learning potential functions for different stacked systems can be constructed based on the interface reaction path information and classical potential function strategies, including:

[0028] Obtaining alloy potential function training data set files of substrates and reaction products and potential function training data set files of gas kinetic reactions; the alloy potential function training data set files include energy and force calculation results associated with different laminated materials;

[0029] Based on the interface reaction path information, the energy distribution of different states in the reaction path is obtained by traversing the insertion point energy information obtained by CINEB through sampling;

[0030] Through the density functional theory self-consistent field calculation method, the single-point energy calculation of the traversed atomic structure is performed to obtain the energy and force data of each atomic configuration;

[0031] The neural evolution potential function is used to train the structural data of alloy and gas systems and the structural information file generated by the NEP method to obtain the alloy potential function suitable for a variety of stacked systems.

[0032] Optionally, use active learning methods to generate a dedicated potential function, including:

[0033] In the initial stage, the neural evolution potential function trained based on the basic dataset enters the small-scale extrapolation experiment stage; the small-scale extrapolation experiment stage adopts a multi-physics field coupling disturbance strategy, constructs an extreme ensemble condition in a preset temperature gradient and a preset pressure fluctuation range, and introduces a preset amplitude of Gaussian white noise to stimulate the structural response of the material in a non-equilibrium state;

[0034] Through high-temperature quenching dynamics simulation of the preset platform, an extrapolation configuration set containing metastable configurations is collected;

[0035] In the configuration screening link, a dual evaluation system based on geometric structure and electronic state characteristics is established, and the atomic root mean square deviation between the extrapolation configuration set and the basic dataset is calculated;

[0036] When the atomic root mean square deviation exceeds a second preset threshold, a configuration marker is triggered; the charge density distribution is reduced by kernel principal component analysis, and the abnormal point of the electronic structure is identified by Mahalanobis distance to obtain an abnormal configuration deviating from the ground state distribution;

[0037] Perform plane wave cutoff energy convergence test, K-point grid density scanning and force convergence threshold verification to obtain verified configurations;

[0038] Extract the energy matrix, force matrix and Hessian matrix from the verified configurations, and convert them into standardized data blocks;

[0039] Perform lossless conversion on the standardized data blocks, and fuse the new and old datasets by a consistent hashing algorithm to obtain an enhanced training set;

[0040] Generate a dedicated potential function using an active learning method.

[0041] Optionally, the extrapolation is performed by a stacked machine learning potential function, including:

[0042] Determine the input parameters of the etching dynamics model of the stacked structure according to the distribution state of the incident particles at the interface; the distribution state at least includes the number, incident angle and velocity distribution;

[0043] Determine whether the element system described by the potential function file contains a stacked material, and obtain a determination result;

[0044] If the determination result indicates that the element system described by the potential function file contains a stacked material, the different stacks are selected to be merged for unified substrate modeling;

[0045] If the determination result indicates that the element system described by the potential function file does not contain a stacked material, each substrate is modeled respectively;

[0046] After the model is constructed, the number of calculation and iteration rounds is set, and the screening condition is determined;

[0047] running molecular dynamics simulation, the etching dynamics process is deduced in the molecular dynamics simulation module to obtain the interaction between the incident particles and the material interface, the change of the etching rate and the reaction behavior of the material;

[0048] The statistical etching yield and the reaction etching selectivity are estimated, comprising:

[0049] If the modeled structure is a laminated material, different laminates are numbered and grouped; the dynamic simulation is performed, the interaction of each incident particle with the material interface is recorded, and the etching yield of each laminated material or single material is counted;

[0050] The reaction etching selectivity is calculated according to the etching yield of each laminated material or single material.

[0051] Compared with the prior art, the present application provides a method for estimating the reaction etching selectivity of a laminated material. The basic data set of the laminated material is obtained, which at least includes the key reaction configuration, the energy barrier and the activation parameter; the basic data set is integrated and screened to obtain an initial sample; based on the initial sample, the iterative optimization mechanism of the machine learning dynamic potential function is used to optimize the basic data set by active learning iterative sampling, to obtain a final potential function; molecular dynamics simulation is performed based on the final potential function, and the reaction etching selectivity of the laminated material is estimated by combining the adaptive time step algorithm. The present application enhances the potential function potential energy surface sampling by active learning method to form a high-precision machine learning potential function and complete the dynamic model of the etching process. A customized potential function is established for different material systems, and molecular dynamics is used to deduce the microprocess in a long time span, to quantitatively calculate the reaction rate difference of each layer of material, and finally to deduce the etching selectivity. Through atomic-level simulation, the visualization tracking of the etching interface behavior is realized, the automatic training simulation process supports server cluster parallel operation, and the atomic-level process simulation cycle of the semiconductor manufacturing end can be shortened.

[0052] In the second aspect, the present application provides a device for estimating the reaction etching selectivity of a laminated material. The device is applied to the method for estimating the reaction etching selectivity of a laminated material, and the device comprises:

[0053] A basic data set acquisition module is used to acquire the basic data set of the laminated material; the basic data set at least includes the key reaction configuration, the energy barrier and the activation parameter;

[0054] An initial sample determination module is used to integrate and screen the basic data set to obtain an initial sample;

[0055] A final potential function determination module is configured to optimize the basic data set based on the initial sample and an iterative optimization mechanism of a machine learning dynamic potential function of active learning iterative sampling to obtain a final potential function.

[0056] A reaction etching selectivity determination module is configured to perform molecular dynamics simulation based on the final potential function and estimate the reaction etching selectivity of the laminated material in combination with an adaptive time step algorithm.

[0057] In a third aspect, the present application provides a device for estimating the reaction etching selectivity of a laminated material, and the device comprises:

[0058] A memory, a processor, and a communication interface coupled to the processor; the memory has a computer program stored thereon, which can be executed by the processor; when the processor executes the computer program, the method for estimating the reaction etching selectivity of a laminated material is executed.

[0059] The device provided in the second aspect and the device provided in the third aspect have the same technical effects as the method provided in the first aspect, and thus will not be described here. BRIEF DESCRIPTION OF DRAWINGS

[0060] The accompanying drawings, which are included to provide a further understanding of the present application, form a part of the present application and illustrate the illustrative embodiments of the present application and the description thereof, and do not limit the present application. In the drawings:

[0061] Figure 1 A method flowchart for estimating the reaction etching selectivity of a laminated material is provided in the present application;

[0062] Figure 2 An interactive implementation flowchart of the method for estimating the reaction etching selectivity of a laminated material is provided in the present application;

[0063] Figure 3 An interface reaction initial-final state diagram for etching is provided in the present application;

[0064] Figure 4 A diagram for searching the transition state of the initial-final state by the NEB method is provided in the present application;

[0065] Figure 5 A diagram for constructing a machine learning potential function according to an existing data set is provided in the present application;

[0066] Figure 6 A diagram of a monomer structure and a laminated structure is provided in the present application;

[0067] Figure 7 A diagram for dynamically counting etching yields in molecular dynamics simulation is provided in the present application;

[0068] Figure 8 A device structure diagram for estimating a selectivity of reactive etching of a laminated material is provided in the present application.

[0069] Figure 9 A device structure diagram for estimating a selectivity of reactive etching of a laminated material is provided in the present application. DETAILED DESCRIPTION

[0070] In order to clearly describe the technical solutions of the embodiments of the present application, in the embodiments of the present application, the terms "first", "second", etc. are used to distinguish the same or similar items with basically the same function and role. For example, the first threshold and the second threshold are only used to distinguish different thresholds, and do not limit the order. Those skilled in the art can understand that the terms "first", "second", etc. do not limit the number and execution order, and the terms "first", "second", etc. also do not mean that they must be different.

[0071] It should be noted that in the present application, the words "exemplary" or "for example" are used to indicate an example, illustration or description. Any embodiment or design scheme described as "exemplary" or "for example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the use of the words "exemplary" or "for example" is intended to present the relevant concept in a specific manner.

[0072] In the present application, "at least one" means one or more, and "multiple" means two or more. The association relationship of the associated objects is described, which means that there can be three relationships, for example, A and / or B, which can represent the following cases: A exists alone, A and B exist together, B exists alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the front and rear associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can represent: a, b, c, the combination of a and b, the combination of a and c, the combination of b and c, or the combination of a, b and c, where a, b and c can be single or multiple.

[0073] In semiconductor manufacturing, with the continuous improvement of the integration of integrated circuits (IC), the functional density is significantly increased, and the selectivity and accuracy of the etching process are rigorously required. Especially for laminated material systems (such as high-k dielectric / metal gate structure, polysilicon / dielectric composite layer, etc.), the complex interface characteristics and heterogeneous material combinations lead to major technical challenges in predicting and controlling etching selectivity. Although existing etching techniques perform well in some applications, they still face the following challenges when dealing with complex laminated materials:

[0074] (1) Bottleneck problem of machine learning interatomic potential training

[0075] Although machine learning interatomic potentials can achieve near-quantum mechanical (QM) accuracy in simulation length and time scales, their training process still faces challenges:

[0076] Training data sampling bias: Existing methods usually rely on random sampling, which cannot effectively capture key configurations in complex etching processes. In particular, there are key regions such as material interfaces and grain boundaries in layered materials, and traditional sampling methods may miss these high-information configurations. In addition, random sampling requires a large amount of unlabeled initial data, and obtaining high-quality labeled data is computationally expensive.

[0077] To address the first deficiency of the prior art, the technical solution proposed by the present application constructs an active evaluation-iterative optimization framework, systematically evaluates the prediction confidence of the current model on unlabeled data, and intelligently selects the most informative samples for labeling and training. This strategy significantly reduces the training data requirements, while improving the modeling capability of complex interface systems, achieving efficient use of computing resources.

[0078] (2) Precise modeling of interface characteristics of layered materials

[0079] In the etching process, layered materials (such as metal / dielectric, semiconductor / dielectric structures) have significant differences in material properties, resulting in complex and unpredictable etching behavior in the interface region. The main difficulties include:

[0080] Dynamic evolution of interface energy barrier and diffusion potential well: Traditional potential functions have difficulty accurately describing the bonding and breaking processes in the interface region, making it difficult to accurately predict the removal rate and selectivity of atoms near the interface layer.

[0081] Variety of anisotropic etching effects: Layered materials exhibit significant anisotropic etching behavior, especially in complex structures such as trenches, vias, and other micro-nano structures. Traditional potential function models cannot capture local etching differences due to insufficient description of space.

[0082] To address the second deficiency of the prior art, the technical solution proposed by the present application uses machine learning potential functions with active learning iterative sampling to dynamically focus on special configurations and transition states in the interface region, capturing key physical and chemical mechanisms in the bonding and breaking processes of the interface layer. Combined with high-fidelity molecular dynamics (MD) and efficient sampling of active learning, the atomic-level behavior in the interface region is accurately characterized, thereby improving the prediction ability of anisotropic etching behavior in complex layered materials.

[0083] (3) Balance of computing resources and prediction efficiency

[0084] High-performance molecular dynamics simulations, especially those applied to etching processes, are computationally expensive in terms of resources and time. Current models face challenges in handling large-scale, long-time-scale etching processes:

[0085] Limited simulation scale: Traditional MD simulations are limited by computational resources and can only cover small spatial scales and short time scales, making it difficult to fully reflect the dynamic evolution of real etching processes.

[0086] Time-consuming data processing and model evaluation: Random sampling-trained models are computationally expensive and inefficient when making predictions, which cannot meet the needs of industrial-level rapid optimization.

[0087] To address the third deficiency of the prior art, the technical solution proposed by the present application uses an active learning strategy to iteratively optimize the sampling strategy, gradually focusing on high-value samples, significantly reducing the initial training data size and subsequent iterative training resource consumption. At the same time, combined with the physical information embedded machine learning potential function, not only improves the computational efficiency of the model, but also ensures the prediction accuracy in large-scale simulation, achieving a reasonable balance between computational resources and prediction accuracy.

[0088] (4) Lack of combination of data-driven models and physical constraints

[0089] Many existing machine learning potential function models mainly rely on data-driven methods, ignoring prior knowledge of physical laws. This data-driven approach may result in poor model performance in untrained data domains, especially in extreme or rare configurations.

[0090] To address the fourth deficiency of the prior art, the technical solution proposed by the present application embeds a physical constraint module in the active learning framework, ensuring that the machine learning potential function captures data-driven features while complying with known physical laws (such as energy conservation, force symmetry, etc.). This hybrid strategy not only enhances the robustness of the model but also improves its prediction reliability under different material systems and process conditions.

[0091] (5) Disconnection between simulation prediction and actual process

[0092] In semiconductor manufacturing, the optimization of etching processes highly depends on accurate simulation prediction tools. However, existing simulation tools often have large deviations when predicting actual process parameters (such as etching rate and selectivity), making it difficult to directly guide process development.

[0093] To address the fifth drawback of the prior art, the present invention proposes a technical solution that enhances the model's ability to predict actual process parameters by optimizing a machine learning potential function using an active learning strategy, combined with a meticulous parameter calibration and verification process. The resulting simulation framework outputs physical quantities that are highly correlated with the actual etching process, providing reliable data support for process optimization.

[0094] This paper introduces an enhanced sampling method based on active learning to effectively address the training challenges of machine-learning interatomic potentials in etch selectivity prediction. This method constructs a high-precision, high-reliability machine-learning potential function with low computational resource consumption, enabling accurate prediction and process optimization of isotropic etch selectivities in complex laminated materials. Compared to traditional simulation methods, this paper significantly improves computational efficiency and prediction accuracy, providing a new technical approach for micro-nanostructure processing in integrated circuit manufacturing.

[0095] This paper utilizes a machine learning potential function model to establish an etching process model for laminated materials. Compared to conventional simulation methods, the present invention utilizes a molecular dynamics simulation etching deduction calculation method based on a machine learning potential function to estimate etching selectivity. This method relies on an accurate and robust machine learning force field to predict the morphological evolution of reactants over a certain time scale and estimates the etching selectivity through precise statistical analysis of etching yields, providing predictive guidance for etching process optimization. The machine learning potential function generates a new scenario space traversed by an interface reaction dataset and an active learning scheme to approximate real-world reaction scenarios.

[0096] Next, the solutions provided in the embodiments of this specification are described with reference to the accompanying drawings:

[0097] like Figure 1 As shown, the process may include the following steps:

[0098] Step 110: Obtain a basic data set of the laminate material; the basic data set at least includes key reaction configurations, energy barriers, and activation parameters.

[0099] Laminated materials refer to two or more different properties, different functions of materials, combined together by a certain process, to form a composite material with unique performance and structure. For example: fiber reinforced laminated materials: with fiber as the reinforcing phase, such as carbon fiber, glass fiber, etc., and the matrix material (such as resin, metal, etc.) is laminated and combined. For example, carbon fiber reinforced composite materials have high strength, high modulus, low density and other characteristics. For example, layered composite materials: composed of alternating thin layers of different materials, such as metal and non-metal laminates, such as titanium / steel, copper / aluminum, etc. Laminated composite materials have high strength and toughness of metals and corrosion resistance, wear resistance and other properties of non-metals. For example, Si / SiO2 laminated materials, Si / SiO2 laminated materials are an important class of materials composed of alternating silicon (Si) and silicon dioxide (SiO2) stacks, widely used in microelectronics, optoelectronics, energy and other fields.

[0100] Step 120: performing integrated screening on the basic data set to obtain an initial sample.

[0101] The integrated screening of the basic data set can more intuitively understand the key configuration, energy change and activation parameter in the reaction process.

[0102] Step 130: based on the initial sample, using the iterative optimization mechanism of the machine learning dynamic potential function of the active learning iterative sampling to optimize the basic data set, to obtain the final potential function.

[0103] The initial model can be constructed based on the initial sample obtained before, using machine learning method to construct an initial dynamic potential function model. This model can predict the energy barrier and activation parameter and other properties according to the input of molecular configuration and other characteristic information.

[0104] When performing iterative sampling, the sampling criteria can be determined by first defining a measure of sample uncertainty, such as variance based on model predictions, confidence intervals, etc. Samples with higher uncertainty often contain more information that the model has not fully learned. According to the sampling criteria, key samples with representative and high uncertainty are selected from the basic data set. These samples may be located in areas of the reaction path that have not been fully explored, or configurations with unstable model prediction results, etc. Further high-precision calculations or experimental measurements are performed on the selected key samples to obtain more accurate energy barriers, activation parameters, etc. to enrich and perfect the basic data set. The newly obtained data is integrated into the original basic data set, and the machine learning model is retrained to obtain an updated dynamic potential function. The new data helps to correct the previous incorrect predictions of the model and improves the prediction ability of the reaction path and related properties. By comparing with known experimental data or high-precision theoretical calculation results, it is verified whether the prediction accuracy of the updated model for the reaction path has been improved. The final accurate dynamic potential function is applied to more extensive reaction simulation and prediction. It can be used to explore unknown reaction paths, predict reaction mechanisms, design new catalysts, etc., providing strong support for the research and application of chemical reactions.

[0105] Step 140: Perform molecular dynamics simulation based on the final potential function to estimate the etching selectivity of the layered material.

[0106] When performing molecular dynamics simulation based on the final potential function, the simulation system can be constructed. According to the structural characteristics of the layered material, an initial simulation system containing reactants, etchants and atoms of the layered material is constructed. During the construction process, the types, positions and quantities of atoms are accurately determined to ensure the accuracy and representativeness of the simulation. Simulation parameters are set, and according to the research purpose and material characteristics, appropriate simulation temperature, pressure and other thermodynamic parameters, as well as simulation time step and other dynamic parameters are selected. Reasonable setting of these parameters can ensure the stability and reliability of the simulation process, and also improve the simulation efficiency. Define the interaction between atoms, and use the final potential function obtained before to define the interaction potential between different atoms. This potential function can accurately describe the energy relationship and force between atoms, providing a basis for molecular dynamics simulation. During the simulation process, the force between atoms is calculated according to the potential function, and then the motion trajectory of the atom and the dynamic behavior of the material are predicted. Initialize the simulation system, input the constructed simulation system and defined interaction potential and other information into the molecular dynamics simulation software for initialization operation. This usually includes assigning initial velocity, energy, etc. to make the simulation system in a reasonable initial state, and prepare for the subsequent simulation process.

[0107] The core idea of the adaptive time step algorithm is to dynamically adjust the simulation time step according to the speed of atomic motion and the change of system energy, etc. When the atomic motion is slow and the energy change is small, the time step can be appropriately increased to improve the simulation efficiency; when the atomic motion is intense and the energy change is large, the time step is reduced to ensure the accuracy and stability of the simulation.

[0108] In the process of molecular dynamics simulation, the state of the system is monitored in real time, such as atomic velocity, acceleration, energy, etc. According to these information, the simulation time step is automatically adjusted according to the rules of the adaptive time step algorithm set in advance. This requires implementing the corresponding algorithm module in the simulation software or reasonably using the existing related functions of the software.

[0109] Figure 1 The method in the application comprises the following steps: obtaining a basic data set of the laminated material, the basic data set comprising at least a key reaction configuration, an energy barrier and an activation parameter; integrating and screening the basic data set to obtain an initial sample; based on the initial sample, using an iterative optimization mechanism of a machine learning dynamic potential function based on active learning iterative sampling to optimize the basic data set to obtain a final potential function; and based on the final potential function, performing molecular dynamics simulation and estimating the etching selectivity of the laminated material by using an adaptive time step algorithm. The application enhances the potential function potential energy surface sampling by using an active learning method to form a machine learning potential function with high precision and complete the dynamic model of the etching process. Customized potential functions are established for different material systems, molecular dynamics is used to deduce the microprocess in a long time span, the reaction rate difference of each layer of material is quantitatively calculated, and finally the etching selectivity is deduced. The visualization tracking of the etching interface behavior is realized through atomic level simulation, the automatic training simulation process supports server cluster parallel operation, and the atomic level process simulation cycle of the semiconductor manufacturing end can be shortened.

[0110] Further, the scheme of the present application can solve the problem of difficult efficient acquisition of microscale etching selectivity in the prior art by enhancing sampling of machine learning molecular dynamics simulation of semiconductor device stack material etching selectivity simulation based on an active learning method. The method enhances potential function potential energy surface sampling by an active learning method to form a higher precision machine learning potential function and complete a dynamic model of the etching process. In specific implementation, a customized potential function is established for different material systems, molecular dynamics is used to deduce microscale processes for a long time span, the reaction rate difference of each layer of material is quantitatively calculated, and finally the etching selectivity is derived. Compared with the traditional experimental method, the present scheme realizes the visualization tracking of the etching interface behavior through atomic level simulation, the automatic training simulation process supports server cluster parallel operation, and the atomic level process simulation cycle of the semiconductor manufacturing end can be shortened. The technology is particularly suitable for the etching process development of high-precision stacked devices in semiconductor manufacturing, which significantly reduces the research and development cost while ensuring the prediction accuracy, and provides an efficient and reliable digital twin platform for high-throughput etching process quantification evaluation.

[0111] Based on the method of Figure 1 , the embodiments of the present specification also provide some specific implementations of the method, which are described below.

[0112] In the scheme of the present application, in specific interaction, the interaction process is as shown in Figure 2 , first collect the data set and initialize, then introduce temperature gradient disturbance, pressure gradient change and random noise (local perturbation, etc.) and multi-dimensional ensemble conditions such as artificial defects and doping introduction, generate an extrapolation configuration set, perform small model reasoning, then perform differential judgment, and screen out abnormal configurations deviating from the ground state distribution. The abnormal configuration is subjected to DFT accuracy verification, the energy convergence result is fused with the original data set to form an enhanced training set, a potential function training module is used on the enhanced training set, and the final potential function is obtained through multiple rounds of potential function iteration, and then the large system model is extrapolated, and data analysis is performed.

[0113] Further, the technical scheme provided by the present application is a cross-scale etching selectivity modeling, which is composed of a multi-scale coupling modeling framework and a dynamic iterative optimization process. The specific implementation process is as follows:

[0114] In specific implementation, step 110 and step 120 perform transition state path analysis and data set initialization. Specifically, based on the atomic scale dynamics characteristics research of material interface reaction, first, the Nudged Elastic Band (NEB) method and its improved algorithm CINEB (Constrained Interpolation Nudged Elastic Band) method are used to construct a basic data set Dataset00 This dataset can realize three-dimensional visualization of reaction paths through a specific platform, providing high-confidence initial samples for potential function training.

[0115] Step 130 may specifically include:

[0116] An initial neural evolution potential function is trained based on the initial sample, and an extrapolated configuration set is generated by introducing a multidimensional ensemble condition; the multidimensional ensemble condition includes at least temperature gradient perturbation, pressure fluctuation and random noise injection; a structural differentiation evaluation function is established, and an atomic displacement matrix comparison is performed on the extrapolated configuration set and the initial sample to screen out abnormal configurations that deviate from the ground state distribution; DFT accuracy verification is performed on the abnormal configurations, and the energy convergence results are fused with the original data set to obtain an enhanced training set; a potential function training module is used to perform multiple rounds of potential function iteration on the enhanced training set, and the convergence condition is triggered when the differentiation rate of the extrapolated configuration set is lower than the first preset threshold for three consecutive times, and the final potential function is output.

[0117] Specifically, the dynamic potential function iterative optimization mechanism may include the following steps:

[0118] (1) Based on Dataset 00 Training the initial neural evolution potential function (Nep 00 ), the Calorine module can be used to implement small-scale extrapolation experiments: by introducing multi-dimensional ensemble conditions such as temperature gradient perturbations (300-1500K), pressure fluctuations (0.1-10GPa) and random noise injection, a set of extrapolated configurations can be generated.

[0119] (2) Establish a structural differentiation evaluation function to extrapolate the configuration set and the Dataset 00 Perform atomic displacement matrix alignment ( As the benchmark threshold), abnormal configurations that deviate from the ground state distribution are screened out.

[0120] (3) Perform DFT accuracy verification on abnormal configurations and fuse the energy convergence results with the original data set to form an enhanced training set Detaset 01 This process uses automated scripts to implement parameter transfer and data format conversion, ensuring seamless cross-platform computing.

[0121] (4) Using the potential function training module to form Nep for the enhanced training set 01 , similarly, for Dataset i Repeat steps (1) to (3) to form a Dataset i+1 At the same time, the potential function training module is used to form Nep i With Nep i+1 (5) Through multiple rounds of potential function iteration (Nep00 →Nep i ), when the extrapolation configuration set difference rate is lower than 5% for 3 times in succession, the convergence condition is triggered, and the final potential function Nep with cross-scale adaptability is output Final .

[0122] In a specific implementation, step 140 can include:

[0123] Constructing interface reaction path information of different stacks; constructing machine learning potential functions of different stack systems according to the interface reaction path information and combining a classical potential function strategy; generating a dedicated potential function by using an active learning method; deducing by using a stack machine learning potential function; and estimating a reaction etching selectivity by counting etching yields.

[0124] Further, the specific implementation process of the above steps is as follows:

[0125] Based on NEP-Final, large-scale molecular dynamics simulation (>10^6 atomic system) is carried out, and a self-adaptive time step algorithm is used to track the spatio-temporal evolution process of the etching front. The specific steps are as follows:

[0126] Step one, constructing interface reaction path information of different stacks.

[0127] Taking Si / SiO2 stack material as an example, as shown in Figure 3 , first, a substrate model of Si and SiGe needs to be established. This modeling process includes defining the initial state and the final state in the absence of reactants and in the presence of reactants, respectively. It is particularly important to note that the initial and final states of the reaction must strictly consider the state of each transition state molecule or compound to ensure that each state group before and after meets the definition of the elementary reaction.

[0128] To this end, as shown in Figure 4 , the NEB (Nudged Elastic Band) or CI-NEB (Climbing Image Nudged Elastic Band) method can be used to calculate the saddle points on the reaction path. These methods can effectively find the lowest energy path of the reaction path and identify the key transition states. During the calculation process, each saddle point on the reaction path needs to be evaluated to determine whether it meets the convergence criteria.

[0129] After the convergence criteria are met, the system will output the atomic information of the reaction path. These information not only provides structural information during the reaction process, but also provides basic data for subsequent reaction rate estimation and selectivity analysis.

[0130] Step two, constructing machine learning potential functions of different stack systems according to interface reaction path information and combining a classical potential function strategy.

[0131] In this step, as shown in Figure 5 , first, it is necessary to pre-collect the alloy potential function training dataset files of the substrate and reaction products. These datasets should contain the calculation results of energy and force related to different stack materials, providing the basis for subsequent machine learning potential function construction. At the same time, it is also necessary to collect the potential function training dataset files of the gas kinetic reaction to cover the case of gas molecule and solid surface interaction.

[0132] Next, using the interface reaction path information obtained in the first step, the interpolation energy information obtained by CINEB is traversed by sampling method (for example, the experience value is set to 100 fs / frame). This process can obtain the energy distribution of different states in the reaction path.

[0133] Through the density functional theory (DFT) self-consistent field (SCF) calculation method, single-point energy calculation is performed on the atomic structure obtained by traversal. This calculation will provide energy and force data for each atomic configuration. Subsequently, it is necessary to screen out non-converged points and results with excessively large force and energy calculation values to ensure the accuracy and effectiveness of the training data.

[0134] Finally, using the Neuro-Evolution Potentials (NEP) module, the empirical cutoff radius of the radial and angular descriptors is set to Combined with the structure data of the alloy and gas system and the structure information file generated by the NEP method, the alloy potential function suitable for various stack systems is finally obtained. This potential function can provide high-precision force field description for subsequent molecular dynamics simulation, thereby simulating the complex scenarios that may be involved in the etching process in advance, such as the dynamic evolution process of surface passivation, crystalline defects (doped / undoped), and etchant interaction. Among them, angstrom is the unit of length, 1 angstrom is equal to 10-10 meters This unit is often used in atomic physics, structural chemistry, and other fields to represent the size of atoms or molecules.

[0135] Step three, generate a dedicated potential function using an active learning method.

[0136] In the initial stage, based on the Dataset 00 trained neural evolution potential Nep 00 will enter the small-scale extrapolation experiment stage. This stage uses a multi-physics field coupling perturbation strategy to construct extreme ensemble conditions in the temperature gradient of 300-1500K and the pressure fluctuation range of 0.1-10GPa through the Calorine module, while introducing a vibration amplitude of no more than Gaussian white noise, systematically excite the material structural response in non-equilibrium state. This multi-modal perturbation method can effectively expand the generalization boundary of potential function, and through the high-temperature quenching dynamics simulation of the software platform of molecular dynamics simulation (such as: LAMMPS platform), the extrapolated configuration set containing metastable state configuration is collected.

[0137] In the configuration screening link, a double evaluation system based on geometric structure and electronic state characteristics is established. First, the root mean square deviation (RMSD) between the extrapolated configuration set and the basic data set is calculated, and when the deviation exceeds the threshold, the configuration is marked; at the same time, kernel principal component analysis is used to reduce the dimension of the charge density distribution, and the Mahalanobis distance is used to identify the abnormal points of the electronic structure. This double criterion mechanism can accurately capture abnormal configurations deviating from the ground state distribution, and the selected candidate configurations will be automatically pushed to the DFT verification process through the message queue.

[0138] After entering the high-precision verification stage, a three-way verification protocol is executed: including plane wave cutoff energy convergence test, K-point grid density scanning and force convergence threshold verification. The configurations verified will extract key parameters such as energy / force matrix and Hessian matrix, and convert them into standardized data blocks through an automated data pipeline. A specially developed format conversion adapter can realize lossless conversion from VASP output to TFRecord format, and through a consistent hashing algorithm, seamless integration of new and old data sets is realized, and finally an enhanced training set Dataset 01 is formed.

[0139] The entire iteration process adopts an intelligent convergence control strategy, including dynamic learning rate adjustment, early stopping mechanism and coupling geometric similarity evaluation method. When the RMSD of the verification set fluctuates by less than 5% for 3 consecutive epochs, it is determined that the potential function has reached the convergence standard. This optimization mechanism makes the final obtained Nep Fianl potential function not only maintains the accuracy of first principles, but also greatly expands the effective prediction range, providing a reliable potential energy surface description for cross-scale etching selection modeling.

[0140] Step four, deduce through the stacked machine learning potential function.

[0141] In this stage, first of all, according to the distribution state of the incident particles (such as F2 or CF4, etc.) at the interface, including the number, incident angle and velocity distribution, the input parameters of the etching dynamics model of the stacked structure are converted. In this step, since the temperature in the experiment is the cavity temperature, the interface can be set to be slightly higher than the cavity temperature.

[0142] Next, the system is judged: if the element system described by the potential function file contains stacked materials, the different stacks can be combined for unified substrate modeling; if not, each substrate needs to be modeled separately.

[0143] After the model is built, the calculation and iteration rounds need to be set, and the screening conditions are determined. For example, the condition can be set that the particles are deleted after reaching a certain interface height, and the experience value can be set to be greater than 20 angstroms. This condition helps to avoid invalid particle collisions and reduce waste of computing resources.

[0144] Subsequently, the molecular dynamics simulation is run, and the etching dynamics process is implemented in the molecular dynamics simulation module, such as using LAMMPS, GROMACS, etc. These simulation tools can efficiently handle complex many-body interactions, thereby deeply analyzing the microscopic mechanisms in the etching process. During the simulation, the interaction of incident particles with the material interface, the change of etching rate, and the reaction behavior of the material can be observed.

[0145] Step five, statistics etching yield, estimate the reaction etching selectivity.

[0146] In this step, as shown in Figure 6 and Figure 7 If the modeled structure is a stacked material, the different stacks need to be numbered and grouped in advance to facilitate effective screening and statistics after the dynamics simulation. This grouping strategy helps to clearly understand the performance of each stack in the etching process, thereby better understanding the influence of different materials on the overall etching yield. If the modeled system is different single-body materials, numbering and grouping are not required, and statistics can be directly performed on different systems.

[0147] Next, the dynamics simulation is performed, and the interaction of each incident particle with the material interface is recorded. In this process, the etching yield of each stacked material or single-body material needs to be carefully counted, i.e., the amount of material successfully etched away per unit time. These data will provide a basis for subsequent selectivity estimation.

[0148] After the etching yield is counted, the reaction etching selectivity can be calculated. The selectivity is usually defined as the ratio of the etching rate of a certain material to the etching rate of another material. By analyzing the etching behavior of different materials under the same conditions, the relative selectivity between materials can be evaluated, providing guidance for optimizing the etching process.

[0149] Furthermore, in the comprehensive characterization of the training and verification results of the machine learning potential function, the predictive performance of the machine learning potential function (MLFF) was verified through a four-dimensional indicator system. Its training dynamics and verification accuracy, the model can not only accurately capture the mean characteristics of the Virial tensor, but also effectively characterize its statistical fluctuation characteristics, providing a reliable mechanical mapping relationship for describing the microscopic evolution of the volume work of the material during the stress process.

[0150] The technical solutions provided by the present invention can achieve at least the following technical effects:

[0151] The present invention has achieved a breakthrough innovation in atomic-level simulation in the field of semiconductor device etching process simulation by integrating first-principles calculations and machine learning technology. Compared with traditional experimental methods that rely on trial and error, this simulation system has achieved the first visual tracking of the dynamic evolution of atomic-level etching interfaces in large systems. The machine learning potential function model error constructed by the active learning method ensures that the atomic force and system energy errors are lower than the empirical threshold while enhancing the accuracy and robustness of its scene simulation. The NEP using GPU inference can improve the efficiency of microscopic process simulation by several orders of magnitude. Through the collaborative optimization of the elastic band method and iterative screening of data sets, the effective transition state path in the interface reaction can be accurately captured, significantly avoiding the risk of missed transition state detection in traditional MD methods. Practical applications have shown that this method has the ability to predict the etching selectivity of typical systems such as SiGe / Si and metal / dielectric stacks, with low relative error, and its automated training simulation process supports cloud-based distributed computing. The single simulation process verification cycle is shortened from several months in traditional molecular dynamics experiments to several weeks. The cross-scale correlation model established by this method not only outputs a quantitative index of selectivity, but also simultaneously generates microscopic details including defect distribution in the atomic-level system and atomic migration process, providing a multi-dimensional decision-making basis for defect control in the etching process.

[0152] Based on the same idea, the present invention also provides a device for estimating the reactive etching selectivity of stacked materials, such as Figure 8 As shown, the device may include:

[0153] A basic data set acquisition module 810 is used to acquire a basic data set of the laminate material; the basic data set includes at least key reaction configurations, energy barriers, and activation parameters;

[0154] An initial sample determination module 820 is configured to integrate and screen the basic data set to obtain an initial sample;

[0155] A final potential function determination module 830 is configured to optimize the basic data set based on the initial samples using an iterative optimization mechanism of a machine learning dynamic potential function using active learning iterative sampling to obtain a final potential function;

[0156] The reactive etching selectivity determination module 840 is configured to perform molecular dynamics simulation based on the final potential function and estimate the reactive etching selectivity of the stacked material in combination with an adaptive time step algorithm.

[0157] based on Figure 8 The device may further include some specific implementation units:

[0158] Optionally, the final potential function determination module 830 may specifically include:

[0159] An extrapolated configuration set generating unit, configured to train an initial neural evolution potential function based on the initial sample and generate an extrapolated configuration set by introducing a multidimensional ensemble condition; the multidimensional ensemble condition at least includes temperature gradient perturbation, pressure fluctuation, and random noise injection;

[0160] An abnormal configuration determination unit is used to establish a structural differentiation evaluation function, perform atomic displacement matrix comparison between the extrapolated configuration set and the initial sample, and screen out abnormal configurations that deviate from the ground state distribution;

[0161] An enhanced training set determination unit is used to perform DFT accuracy verification on the abnormal configuration and fuse the energy convergence result with the original data set to obtain an enhanced training set;

[0162] The final potential function output unit is used to perform multiple rounds of potential function iteration on the enhanced training set using the potential function training module, trigger the convergence condition when the extrapolated configuration set differentiation rate is lower than the first preset threshold for three consecutive times, and output the final potential function.

[0163] Optionally, the reactive etching selectivity ratio determining module 840 may specifically include:

[0164] An interface reaction path information construction unit, used to construct interface reaction path information of different stacked layers;

[0165] A potential function construction unit, used to construct machine learning potential functions for different stacked systems based on the interface reaction path information and classical potential function strategies;

[0166] A dedicated potential function generating unit, used for generating a dedicated potential function by adopting an active learning method;

[0167] A deduction unit, used for performing deduction by stacking machine learning potential functions;

[0168] The reactive etching selectivity ratio estimation unit is used to count the etching output and estimate the reactive etching selectivity ratio.

[0169] Optionally, the initial sample determination module 820 may be configured to:

[0170] The CINEB method is used to construct a basic data set including key reaction configurations, energy barriers and activation parameters through first-principle calculation; the basic data set is used to realize three-dimensional visualization of the reaction path through a preset platform, and high-confidence initial samples for potential function training are obtained.

[0171] Optionally, the interface reaction path information construction unit can be used for:

[0172] The initial state and the final state are defined in the absence of reactants and in the presence of reactants, respectively, and a substrate model is established;

[0173] The NEB method is used to calculate the saddle points on the reaction path, and each saddle point on the reaction path is evaluated to determine whether it meets the convergence criteria; the lowest energy path of the reaction path is found, and the key transition state is identified; atomic information of the reaction path is output; the atomic information includes the initial and final states in the reaction process and the structure information of the transition state saddle points existing.

[0174] Optionally, the potential function construction unit can be used for:

[0175] Obtain the alloy potential function training data set file of the substrate and the reaction product and the potential function training data set file of the gas dynamics reaction; the alloy potential function training data set file includes the calculation results of energy and force related to different laminated materials;

[0176] Based on the interface reaction path information, the energy information of the interpolation points obtained by CINEB is traversed by sampling to obtain the energy distribution of different states in the reaction path;

[0177] The atomic structure obtained by traversal is calculated by the density functional theory self-consistent field calculation method to obtain the energy and force data under each atomic configuration;

[0178] The neural evolution potential function is used to train the structure data of the alloy and gas system and the structure information file generated by the NEP method to obtain an alloy potential function suitable for various laminated systems.

[0179] Optionally, the dedicated potential function generation unit can be used for:

[0180] In the initial stage, the neural evolution potential function trained based on the basic data set enters the small-scale extrapolation experiment stage; the small-scale extrapolation experiment stage uses a multi-physics field coupling perturbation strategy to construct extreme ensemble conditions within a preset temperature gradient and a preset pressure fluctuation range, and introduces a preset amplitude of Gaussian white noise to excite the structural response of the material in a non-equilibrium state;

[0181] Through high-temperature quenching dynamics simulation of the preset platform, an extrapolation configuration set containing metastable configurations is collected;

[0182] In the configuration screening link, a dual evaluation system based on geometric structure and electronic state characteristics is established, and the atomic root mean square deviation between the extrapolated configuration set and the basic data set is calculated;

[0183] When the atomic root mean square deviation exceeds the second preset threshold, the configuration label is triggered; the charge density distribution is reduced by kernel principal component analysis, and the Mahalanobis distance is used to identify the electronic structure abnormal point to obtain the abnormal configuration deviating from the ground state distribution;

[0184] Conduct plane wave cutoff energy convergence test, K point grid density scanning and force convergence threshold verification to obtain the verified configuration;

[0185] The energy matrix, force matrix and Hessian matrix of the verified configuration are extracted and converted into standardized data blocks;

[0186] The standardized data blocks are losslessly converted, and the old and new data sets are fused by a consistent hash algorithm to obtain an enhanced training set;

[0187] An active learning method is used to generate a dedicated potential function.

[0188] Optionally, the deduction unit can be used for:

[0189] Determine the input parameters of the etching dynamics model of the laminated structure according to the distribution state of the incident particles at the interface; the distribution state at least includes the number, the incident angle and the velocity distribution;

[0190] Determine whether the element system described by the potential function file contains laminated materials, and obtain a determination result;

[0191] If the determination result indicates that the element system described by the potential function file contains laminated materials, the different laminated substrates are selected to be merged and modeled uniformly;

[0192] If the determination result indicates that the element system described by the potential function file does not contain laminated materials, each substrate is modeled respectively;

[0193] After the model is constructed, the calculation and iteration rounds are set, and the screening conditions are determined;

[0194] Run the molecular dynamics simulation, deduce the etching dynamics process in the molecular dynamics simulation module, and obtain the interaction between the incident particles and the material interface, the change of the etching rate and the reaction behavior of the material;

[0195] The statistical etching yield and the estimated reactive etching selectivity include:

[0196] If the modeled structure is a stack of materials, then the different stacks are numbered and grouped; a dynamic simulation is performed, the interaction of each incident particle with the material interface is recorded, and the etching yield of each stack material or monomer material is counted;

[0197] The reactive etching selectivity of each stack material or monomer material is calculated according to the etching yield.

[0198] Based on the same idea, the embodiments of the present specification also provide a device for estimating the reactive etching selectivity of a stack material. As shown in the figure, Figure 9 the device comprises:

[0199] a memory, a processor, and a communication interface coupled to the processor; the memory stores a computer program executable by the processor; and the processor executes the computer program to perform the method for estimating the reactive etching selectivity of a stack material.

[0200] As shown in the figure, Figure 9 the processor can be a general central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of programs of the present application. The communication interface can be one or more. The communication interface can use any transceiver device to communicate with other devices or communication networks.

[0201] As shown in the figure, Figure 9 the terminal device can further include a communication line. The communication line can include a path for transmitting information between the components. Optionally, as shown in the figure, Figure 9 the terminal device can further include a memory. The memory stores a computer program executable by the processor; and the processor executes the computer program to implement the method provided by the embodiments of the present application.

[0202] Optionally, the computer execution instructions in the embodiments of the present application can also be referred to as application program codes, which are not specifically limited by the embodiments of the present application.

[0203] In a specific implementation, as an embodiment, as shown in the figure, Figure 9 the processor can include one or more CPUs, such as CPU0 and CPU1 in the figure. Figure 9

[0204] In a specific implementation, as an embodiment, as shown in the figure, Figure 9 the terminal device can include multiple processors, such as CPU0 and CPU1 in the figure. ​ ​Each of these processors can be a single core processor or a multiple core processor.

[0205] The above mainly introduces the scheme provided by the embodiments of the application from the perspective of interaction between the modules. It can be understood that each module comprises a hardware structure and / or a software unit for executing the corresponding functions in order to achieve the above functions. Those skilled in the art should easily realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present text can be realized in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed in the form of hardware or computer software driven hardware depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0206] The embodiments of the application can divide the functional modules according to the above method examples. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing module. The above integrated module can be realized in the form of hardware or software functional module. It should be noted that the division of the modules in the embodiments of the application is illustrative, and is only a logical function division. There can be another division manner in actual implementation.

[0207] The processor in the present specification can also have the function of a memory. The memory is used to store computer execution instructions for executing the scheme of the present application, and is controlled by the processor to execute. The processor is used to execute the computer execution instructions stored in the memory, thereby realizing the method provided by the embodiments of the application.

[0208] The memory may be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, an optical disc storage (including a compact disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory may exist independently and be connected to the processor via a communication line. The memory may also be integrated with the processor.

[0209] Although the present invention has been described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0210] Although the present invention has been described with reference to specific features and embodiments thereof, it will be apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention as defined by the appended claims and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the invention. It will be apparent that various modifications and variations may be made to the present invention by those skilled in the art without departing from the spirit and scope of the invention. Thus, the present invention is intended to include such modifications and variations as fall within the scope of the claims of the present invention and their equivalents.

Claims

1. A method for estimating the reactive etching selectivity of a stacked material, characterized in that: Methods include: Obtain basic data sets of laminated materials; The basic data set includes at least key reaction configurations, energy barriers and activation parameters; Integrating and screening the basic data set to obtain an initial sample; Based on the initial samples, the basic data set is optimized using an iterative optimization mechanism of a machine learning dynamic potential function with active learning iterative sampling to obtain a final potential function; Molecular dynamics simulation is performed based on the final potential function, and the reactive etching selectivity of the stacked material is estimated in combination with an adaptive time step algorithm.

2. The method for estimating the reactive etching selectivity of a stacked material according to claim 1, wherein: Based on the initial samples, the basic data set is optimized using an iterative optimization mechanism of a machine learning dynamic potential function with active learning iterative sampling to obtain a final potential function, including: Training an initial neural evolution potential function based on the initial sample, and generating an extrapolated configuration set by introducing a multidimensional ensemble condition; the multidimensional ensemble condition at least includes temperature gradient perturbation, pressure fluctuation, and random noise injection; Establishing a structural differentiation evaluation function, performing atomic displacement matrix comparison between the extrapolated configuration set and the initial sample, and screening out abnormal configurations that deviate from the ground state distribution; Perform DFT accuracy verification on the abnormal configuration, fuse the energy convergence result with the original data set to obtain an enhanced training set; The enhanced training set is subjected to multiple rounds of potential function iteration using a potential function training module. When the differentiation rate of the extrapolated configuration set is lower than a first preset threshold for three consecutive times, a convergence condition is triggered and a final potential function is output.

3. The method for estimating the reactive etching selectivity of a stacked material according to claim 1, wherein: Molecular dynamics simulation is performed based on the final potential function, and an adaptive time step algorithm is used to estimate the reactive etching selectivity of the stacked material, including: Constructing interface reaction path information of different stacks; According to the interface reaction path information and the classical potential function strategy, machine learning potential functions of different stacked systems are constructed; Active learning method is used to generate exclusive potential function; Deduction is performed through stacked machine learning potential functions; Calculate etching output and estimate reactive etching selectivity.

4. The method for estimating the reactive etching selectivity of a stacked material according to claim 1, wherein: The basic data set is integrated and screened to obtain an initial sample, including: Using the CINEB method, a basic dataset containing key reaction configurations, energy barriers, and activation parameters was constructed through first-principles calculations. This basic dataset enabled three-dimensional visualization of the reaction pathway through a preset platform, yielding high-confidence initial samples for potential function training.

5. The method for estimating the reactive etching selectivity of a stacked material according to claim 3, wherein: Construct interface reaction path information for different stacks, including: Define the initial state and final state in the absence and presence of reactants, respectively, and establish a substrate model; The NEB method is used to calculate saddle points on the reaction path. Each saddle point on the reaction path is evaluated to determine whether it meets the convergence criteria. The lowest energy path of the reaction path is found and the key transition state is identified. The atomic information of the reaction path is output; the atomic information includes the initial and final states of the reaction process and the existing saddle point structures of multiple transition states.

6. The method for estimating the reactive etching selectivity of a stacked material according to claim 3, wherein: Based on the interface reaction path information and the classical potential function strategy, machine learning potential functions for different stacked systems are constructed, including: Obtaining alloy potential function training data set files of substrates and reaction products and potential function training data set files of gas kinetic reactions; the alloy potential function training data set files include energy and force calculation results associated with different laminated materials; Based on the interface reaction path information, the energy distribution of different states in the reaction path is obtained by traversing the insertion point energy information obtained by CINEB through sampling; Through the density functional theory self-consistent field calculation method, the single-point energy calculation of the traversed atomic structure is performed to obtain the energy and force data of each atomic configuration; The neural evolution potential function is used to train the structural data of alloy and gas systems and the structural information file generated by the NEP method to obtain the alloy potential function suitable for a variety of stacked systems.

7. The method for estimating the reactive etching selectivity of a stacked material according to claim 3, wherein: Generate a dedicated potential function using active learning methods, including: In the initial stage, based on the neural evolution potential function obtained from training of the basic data set, a small-scale extrapolation experiment phase is entered. The small-scale extrapolation experiment phase adopts a multi-physics field coupling perturbation strategy to construct extreme ensemble conditions within a preset temperature gradient and a preset pressure fluctuation range, and introduces Gaussian white noise of a preset amplitude to stimulate the structural response of the material in a non-equilibrium state. Through high-temperature quenching dynamics simulation of the preset platform, a collection of extrapolated configurations including metastable configurations is collected; In the configuration screening phase, a dual evaluation system based on geometric structure and electronic state characteristics is established to calculate the atomic root mean square deviation between the extrapolated configuration set and the basic data set; When the atomic root mean square deviation exceeds a second preset threshold, the configuration flag is triggered; the charge density distribution is reduced in dimension using nuclear principal component analysis, and the electronic structure anomalies are identified through Mahalanobis distance to obtain abnormal configurations that deviate from the ground state distribution; Conduct plane wave cutoff energy convergence test, K-point grid density scan and force convergence threshold verification to obtain a verified configuration; Extract the energy matrix, force matrix and Hessian matrix of the verified configuration and convert them into standardized data blocks; Performing lossless conversion on the standardized data blocks and fusing the new and old data sets through a consistent hashing algorithm to obtain an enhanced training set; Generate exclusive potential functions using active learning methods.

8. The method for estimating the reactive etching selectivity of a stacked material according to claim 3, wherein: Deduction is performed through stacked machine learning potential functions, including: Determining input parameters of an etching kinetics model of a stacked structure based on a distribution state of incident particles at an interface; the distribution state includes at least quantity, incident angle, and velocity distribution; Determine whether the element system described by the potential function file contains a laminated material and obtain a determination result; If the judgment result indicates that the element system described by the potential function file includes a stacked material, different stacks are merged to perform a unified substrate modeling; If the judgment result indicates that the element system described by the potential function file does not contain a laminated material, each substrate is modeled separately; After the model is built, set the number of calculations and iterations, and determine the screening conditions; Run molecular dynamics simulation to deduce the etching dynamics process in the molecular dynamics simulation module to obtain the interaction between the incident particles and the material interface, the change in etching rate, and the reaction behavior of the material; The statistical etching yield and the estimation of the reactive etching selectivity ratio include: If the structure being modeled is a laminated material, the different layers are numbered and grouped; a dynamic simulation is performed, the interaction between each incident particle and the material interface is recorded, and the etching yield of each laminated material or monomer material is calculated; The reactive etching selectivity is calculated based on the etching yield of each stacked material or single material.

9. A device for estimating the reactive etching selectivity of a stacked material, characterized in that: The device is applied to the method for estimating the reactive etching selectivity ratio of a stacked material according to any one of claims 1 to 8, and the device comprises: A basic data set acquisition module is used to acquire a basic data set of the laminated material; the basic data set includes at least key reaction configurations, energy barriers, and activation parameters; An initial sample determination module is used to integrate and screen the basic data set to obtain an initial sample; A final potential function determination module is used to optimize the basic data set based on the initial samples using an iterative optimization mechanism of a machine learning dynamic potential function with active learning iterative sampling to obtain a final potential function; The reactive etching selectivity determination module is used to perform molecular dynamics simulation based on the final potential function and estimate the reactive etching selectivity of the stacked material in combination with an adaptive time step algorithm.

10. A device for estimating the selectivity ratio of reactive etching of stacked materials, characterized in that the device include: a memory, a processor, and a communication interface coupled to the processor; The memory stores a computer program executable by the processor; When the processor runs the computer program, it executes the method for estimating the reactive etching selectivity ratio of stacked materials according to any one of claims 1 to 8.

Images