Molecular dynamics simulation method, device and equipment for molecular beam epitaxy process

Abstract

The invention discloses a molecular dynamics simulation method, device and equipment oriented to a molecular beam epitaxy process, and relates to the technical field of deposition in microelectronic manufacturing, and the method comprises the steps: analyzing a transition state path, and constructing an initialized data set; training an initial neural evolution potential function based on the initialized data set, performing a small-scale extrapolation experiment, establishing an evaluation function, screening a differentiation configuration, and executing DFT precision verification to form an enhanced training set; after multi-round iteration, a final potential function is output until a convergence condition is met; large-scale molecular dynamics simulation is carried out, different data sets are constructed, a potential function is trained after feature fusion and cross validation, deduction is carried out, the deposition rate is counted, and the reaction deposition selection ratio is estimated. According to the method, the dynamic evolution rule of the heterogeneous interface defect can be accurately analyzed, the real technological process problem is flexibly locked, and microsecond-level real-time closed-loop accurate control of the atomic layer growth rate is realized.

Description

Technical Field

[0001] This invention relates to the field of deposition technology in microelectronics manufacturing, and in particular to molecular dynamics simulation methods, apparatus and equipment for molecular beam epitaxy processes. Background Technology

[0002] As integrated circuit feature sizes enter the sub-3 nanometer node, atomically precise thin-film deposition control has become a key bottleneck restricting device performance. Molecular beam epitaxy (MBE) technology, with its ultra-high vacuum environment (10⁻⁶), offers advantages over traditional methods. -10 The Torr-level beam current control and precise beam current control (±1% flow accuracy) demonstrate unique advantages in heteroepitaxy of advanced materials such as III-V compound semiconductors and topological insulators. However, traditional MBE process optimization still faces three major technical barriers:

[0003] 1. The dynamic mechanism of interface defects is unclear: Existing molecular dynamics (MD) simulations are limited by the accuracy of empirical potential functions (typical errors reach hundreds or even thousands of meV / atom), making it difficult to accurately predict the dislocation nucleation process at heterogeneous interfaces such as SiGe / Si.

[0004] 2. The dimensionality problem of process parameters: Multiple parameters such as substrate temperature (400-600℃), beam intensity ratio (Si / Ge element ratio: 10-1), and growth rate (1~10μm / h) are coupled to form a high-dimensional optimization space;

[0005] 3. Real-time process control lag: Online monitoring based on Reflection High Energy Electron Diffraction (RHEED) suffers from a 10% lag. 3 -10 4 The second-level response delay makes it impossible to detect the atomic layer growth rate in real time.

[0006] Therefore, there is an urgent need to provide a more reliable machine learning molecular dynamics simulation scheme for molecular beam epitaxy. Summary of the Invention

[0007] The purpose of this invention is to provide a molecular dynamics simulation method, apparatus and equipment for molecular beam epitaxy, which can solve the problems of unclear interface defect dynamics mechanism, process parameter spatial dimension disaster and real-time process control lag in the prior art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a molecular dynamics simulation method for molecular beam epitaxy, the method comprising:

[0010] Parse the transitional path and construct the initial dataset;

[0011] The initial neural evolution potential function is trained based on the initial dataset, and a small-scale extrapolation experiment is conducted. An evaluation function is established to screen for differential configurations, and DFT accuracy verification is performed to form an enhanced training set.

[0012] After multiple rounds of iteration, the final potential function is output until the convergence condition is met.

[0013] Large-scale molecular dynamics simulations were conducted to construct different datasets. After feature fusion and cross-validation, the potential function was trained, and the deposition rate was extrapolated, statistically analyzed, and the reaction-deposition selectivity was estimated.

[0014] Optionally, parse the transitional path and construct an initial dataset, including:

[0015] Based on the atomic-level dynamic characteristics of material interface reactions, the climbing method and elastic band method are used to analyze the transition state path in the deposition process;

[0016] An initial dataset containing key reaction configurations, energy barriers, and activation parameters is constructed using first-principles calculations; the initial dataset is then used to reconstruct a three-dimensional visualization of the reaction pathway through the ASE platform.

[0017] After multiple iterations, until the convergence condition is met, the final potential function is output, including:

[0018] Through multiple rounds of potential function iteration, when the extrapolation configuration differentiation rate is lower than the preset value for several consecutive times, the convergence condition is triggered, and the final potential function with cross-scale adaptability is output; the final potential function is a combination of descriptor operators and multi-layer artificial neural networks.

[0019] Optionally, an initial neural evolution potential function is trained based on the initialization dataset, and a small-scale extrapolation experiment is conducted to establish an evaluation function to screen for differential configurations. DFT accuracy verification is then performed to form an enhanced training set, including:

[0020] The initial neural evolution potential function is trained based on the initial dataset, and a small-scale extrapolation experiment is carried out. By introducing multi-dimensional ensemble conditions, an extrapolation configuration set is generated. The multi-dimensional ensemble conditions include at least temperature gradient perturbation, pressure fluctuation and random noise injection.

[0021] A structural differentiation evaluation function is established, and the extrapolated configuration is compared with the atomic displacement matrix of the initial dataset to screen out the differential configurations that deviate from the ground state distribution.

[0022] Perform DFT accuracy verification on the differentiated configurations, and fuse the energy convergence results with the original dataset to form an enhanced training set.

[0023] Optionally, large-scale molecular dynamics simulations can be conducted to construct different datasets. After feature fusion and cross-validation, the potential function can be trained to extrapolate, statistically analyze deposition rates, and estimate the reaction-deposition selectivity ratio, including:

[0024] Construct atomic potential energy information files with different configurations;

[0025] Based on reaction path information and classical potential function strategies, machine learning potential functions for different stacked systems are constructed.

[0026] The accuracy of the dedicated potential function is evaluated, and the deposition rate is calculated and the reaction deposition selectivity is estimated by extrapolating the potential function through stacked machine learning.

[0027] Optionally, construct atomic potential energy information files with different configurations, including:

[0028] Obtain crystal structure data covering a variety of Si-Ge binary alloy systems;

[0029] Thermodynamic processes are simulated within a pre-defined wide temperature range using first-principles molecular dynamics methods.

[0030] The supercell expansion simulation scale is constructed by using lattice vibration and atomic diffusion behavior under supercell expansion temperature gradient. A pre-set lattice constant perturbation is applied to excite atomic displacement, and a pre-set vacancy concentration is randomly introduced to simulate defect evolution.

[0031] Molecular dynamics trajectory sampling is performed based on a preset time step, and the sampling frequency is dynamically adjusted according to the system energy gradient and atomic displacement variance. After high-precision single-point energy calculation, the sampled data is imported into a self-developed potential function dataset generation platform. After data cleaning and feature standardization, a high-quality crystal relaxation process initialization dataset is formed.

[0032] The three datasets were obtained by simulating the diffusion behavior of gaseous precursors through molecular dynamics simulation and by searching reaction paths based on transition state theory. After feature fusion and cross-validation, they constituted a complete training set for machine learning potential functions.

[0033] Optionally, machine learning potential functions for different stacked systems can be constructed based on reaction path information and classical potential function strategies, including:

[0034] Obtain key datasets; the key datasets include at least the alloy potential function training dataset of the substrate and reaction products, and the potential function training dataset of the gas dynamics reaction;

[0035] A single-point energy calculation of atomic configurations is performed using the density functional theory self-consistent field method to generate an initial dataset containing energy and force field information.

[0036] Perform data quality control to remove non-converged calculation results, filter out abnormal values ​​of force field energy, and verify the physical rationality of atomic configurations to obtain valid data;

[0037] Based on the neuroevolutionary potential function framework, the obtained structural data is trained at multiple scales and extrapolated for verification. Target interference factors are introduced to evaluate the generalization ability of the potential function. The target interference factors include at least temperature gradient changes, impurity concentration gradients, and defect configurations.

[0038] When the extrapolation configuration and the prediction error exceed the threshold, the extrapolation configuration is included in the dataset after being verified by DFT.

[0039] When the improvement rate of the potential function caused by the addition of new data in continuous iteration is less than the preset improvement rate, it is determined to be converged, and the final potential function is output.

[0040] Optionally, an accuracy assessment of the specific potential function is performed, and the deposition rate is calculated and the reaction-deposition selectivity is estimated through stacked machine learning of the potential function, including:

[0041] By combining potential energy surface sampling visualization with radial distribution function analysis, the potential energy surface sampling visualization presents the energy distribution characteristics of atoms in different spatial positions and configurations, identifies energy extreme regions and captures atomic transition paths;

[0042] Radial distribution function analysis provides statistical characteristics of interatomic distance distribution. By comparing RDF curves under different conditions, the accuracy of the potential function in describing interatomic interactions can be evaluated. The potential energy surface distribution characteristics between deposited atoms and substrate are introduced to characterize the energy barrier height during atomic diffusion and reconstruction, and to predict atomic diffusion rate and reconstruction mechanism.

[0043] The physical properties of the incident particles are analyzed, and the particle distribution at the interface is mapped to the input parameters of the cascade structure deposition dynamics model through the Boltzmann transport equation.

[0044] The interface temperature is dynamically corrected to a preset multiple of the cavity temperature. Periodic boundary conditions are used to simulate an infinitely large substrate, and a surface step structure with a specific crystal orientation is generated through atomic stacking.

[0045] The step height is set to a single atomic layer, and surface vacancy defects are introduced; the model size is optimized to a preset size based on computational resources.

[0046] A parallel molecular dynamics simulation platform was launched. The simulation used an ideal substrate model to obtain basic growth dynamics data. Random interface defects and beam disturbances were gradually introduced in each round. By comparing the deposition rate, surface roughness and interface defect density under different working conditions, the mapping relationship between process parameters and structural performance was established.

[0047] If the structure being modeled is a layered material, the different layers should be numbered and grouped in advance; if the system being modeled consists of different monolithic materials, then numbering and grouping are not necessary.

[0048] Perform dynamic simulations, record the interaction between each incident particle and the material interface, and statistically analyze the deposition rate of each stacked material or single material.

[0049] The reaction deposition selectivity is calculated. The selectivity is defined as the ratio of the deposition rate of a particular material to the deposition rate of another material. The larger the deposition selectivity, the higher the deposition selectivity.

[0050] Compared with existing technologies, this invention provides a molecular dynamics simulation method, apparatus, and equipment for molecular beam epitaxy (MBE). It constructs an initial dataset by analyzing the transition state path; trains an initial neural evolution potential function based on the initial dataset, conducts small-scale extrapolation experiments, establishes an evaluation function to screen differentiated configurations, and performs DFT accuracy verification to form an enhanced training set; after multiple iterations, the final potential function is output when the convergence condition is met; large-scale molecular dynamics simulations are conducted, different datasets are constructed, and the potential function is trained after feature fusion and cross-validation, performing extrapolation, statistical deposition rate analysis, and estimation of reaction-deposition selectivity. By constructing a fusion model of deep neural networks and molecular dynamics, it overcomes the bottlenecks of accuracy of traditional empirical potential functions and efficiency of first-principles molecular dynamics, accurately analyzing the dynamic evolution of defects at heterogeneous interfaces; the dynamic combination of process parameters and the microscopic simulation environment flexibly identifies problems in the real process; and by combining real-time monitoring data streams and dynamically updated dynamic models, it reduces process control response time, achieving microsecond-level real-time closed-loop precise control of atomic layer growth rates, opening a new path for the intelligent and high-precision development of MBE.

[0051] Secondly, the present invention provides a molecular dynamics simulation device for molecular beam epitaxy processes, the device comprising:

[0052] The initial dataset building module is used to parse the transition path and build the initial dataset;

[0053] The enhanced training set determination module is used to train the initial neural evolution potential function based on the initialization dataset, conduct small-scale extrapolation experiments, establish an evaluation function to screen differential configurations, and perform DFT accuracy verification to form an enhanced training set.

[0054] The final potential function determination module is used to output the final potential function after multiple rounds of iteration until the convergence condition is met.

[0055] The reaction-deposition selectivity determination module is used to conduct large-scale molecular dynamics simulations, construct different datasets, train the potential function after feature fusion and cross-validation, and perform deduction, statistical deposition rate calculation and estimate reaction-deposition selectivity.

[0056] Thirdly, the present invention provides a molecular dynamics simulation device for molecular beam epitaxy processes, the device comprising:

[0057] The system includes a memory, a processor, and a communication interface coupled to the processor; the memory stores a computer program that can be run by the processor; when the processor runs the computer program, it executes the aforementioned molecular dynamics simulation method for molecular beam epitaxy.

[0058] Fourthly, the present invention provides a computer storage medium storing instructions that, when executed, implement the aforementioned molecular dynamics simulation method for molecular beam epitaxy.

[0059] The technical effects achieved by the device-type solution provided in the second aspect, the equipment-type solution provided in the third aspect, and the computer storage medium solution provided in the fourth aspect are the same as those achieved by the method-type solution provided in the first aspect, and will not be repeated here. Attached Figure Description

[0060] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0061] Figure 1 A schematic diagram of the molecular dynamics simulation method for molecular beam epitaxy provided by this invention;

[0062] Figure 2 A schematic diagram of the overall framework steps of the molecular dynamics simulation method for molecular beam epitaxy provided by this invention;

[0063] Figure 3 This is a schematic diagram of the alloy dynamics process dataset;

[0064] Figure 4 This is a schematic diagram of a gas reaction dataset;

[0065] Figure 5 A schematic diagram illustrating the construction of a machine learning potential function based on an existing dataset;

[0066] Figure 6 This is a diagram comparing NEP and DFT RDF information;

[0067] Figure 7 A simplified schematic diagram of the interaction relationship of two-dimensional potential energy surfaces;

[0068] Figure 8 A schematic diagram illustrating the potential energy surface between deposited atoms and the substrate;

[0069] Figure 9 An image depicting the deposition of Si and Ge atoms on a single-crystal silicon surface to form a Si0.7Ge0.3 alloy using MLFF (Multi-Layer Fiber Optic Facility).

[0070] Figure 10 A schematic diagram of a molecular dynamics simulation device for molecular beam epitaxy.

[0071] Figure 11 This is a schematic diagram of a molecular dynamics simulation device for molecular beam epitaxy. Detailed Implementation

[0072] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0073] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0074] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0075] Addressing the three core technological bottlenecks faced by molecular beam epitaxy (MBE) in advanced material heteroepitaxial growth after the current integration process enters the sub-3 nanometer node—unclear interface defect dynamics, the curse of spatial dimensions in process parameters, and lag in real-time process control—this paper provides a high-precision and high-efficiency simulation and optimization method. Specifically:

[0076] Challenges in analyzing interface defects: Traditional molecular dynamics (MD) simulations rely on empirical potential functions, which are not accurate enough (typical errors reach hundreds or even thousands of meV / atom). They cannot accurately predict atomic-level processes such as dislocation nucleation and chemical bond reconstruction at heterogeneous interfaces such as SiGe / Si, resulting in unclear interface defect formation mechanisms and making it difficult to guide process optimization through simulation.

[0077] The dilemma of high-dimensional parameter optimization: In the MBE process, multiple parameters such as substrate temperature (400-600℃), (Si / Ge element ratio: 10-1), growth rate (1~10μm / h) are strongly coupled, forming a high-dimensional nonlinear optimization space. Traditional trial and error methods require hundreds of experimental iterations, which are time-consuming, labor-intensive, and prone to getting trapped in local optima.

[0078] Real-time control response lag: Online monitoring based on Reflection High Energy Electron Diffraction (RHEED) suffers from 10... 3 -10 4 The second-level latency cannot meet the real-time control requirements of atomic layer growth, resulting in large fluctuations in growth rate and uneven interface quality.

[0079] This invention deeply integrates machine learning and molecular dynamics to construct a high-precision simulation model coupled with multiple physics fields. This overcomes the accuracy and efficiency bottlenecks of traditional methods, achieving a complete process innovation from atomic-level mechanism analysis to real-time optimization of process parameters. It solves the key technical challenge restricting atomic-level control of MBE processes in advanced semiconductor material fabrication. The following description, in conjunction with the accompanying drawings, illustrates the solutions provided in the embodiments of this specification:

[0080] like Figure 1 As shown, the process may include the following steps:

[0081] Step 110: Parse the transitional path and construct the initial dataset.

[0082] The transition state path can be represented in transition state theory as the path through which reactants are transformed into products. There is a highest energy state, namely the transition state, on this path. The transition state path is the lowest energy path connecting reactants and products. Analyzing the transition state path helps to understand the detailed mechanism and key steps of the reaction.

[0083] An initial dataset can be a set of basic data collected and organized for subsequent calculations and analysis. It usually contains some initial information such as structure and energy, serving as the starting point for the entire process.

[0084] In this scheme, based on the atomic-level dynamic characteristics of material interface reactions, the Nudged ElasticBand and Elastic Band Methods are used to analyze the transition state path during the deposition process. An initial dataset, Dataset00, containing key reaction configurations, energy barriers, and activation parameters, is constructed through first-principles calculations, and the reaction path is reconstructed using the ASE (Atomic Simulation Environment) platform in three dimensions.

[0085] Step 120: Train the initial neural evolution potential function based on the initialization dataset, conduct small-scale extrapolation experiments, establish an evaluation function to screen for differential configurations, and perform DFT accuracy verification to form an enhanced training set.

[0086] The initial neural evolution potential function can be a potential function constructed based on neural network technology to describe the interaction between particles. The neural evolution potential function in the initial state is obtained through preliminary training using a certain method, and further optimization and improvement are needed.

[0087] Small-scale extrapolation experiments can be defined as extrapolation experiments conducted on a small range or with a small amount of data. They are used to initially verify the accuracy and reliability of the initial neural evolution potential function and to observe its predictive ability beyond known data.

[0088] Density functional theory (DFT) is a quantum mechanical calculation method used to study multi-electron systems. DFT accuracy verification involves comparing the calculation results with those obtained by the DFT method to verify the accuracy of the method or model used.

[0089] An augmented training set is a training set obtained by adding new data or information to the initial training set. This additional data helps to improve the performance and generalization ability of the training model.

[0090] Step 130: After multiple iterations, the final potential function is output until the convergence condition is met.

[0091] Multi-round iteration refers to a process of repeated calculations and optimizations. Each iteration adjusts and improves upon the results of the previous iteration to gradually enhance the model's performance and accuracy, ultimately reaching a better state. When certain metrics of the model (such as the loss function value, the error between the predicted and actual values, etc.) reach a certain level of stability or meet a set threshold requirement, the iteration process is considered to have converged, and iteration can be stopped. For example, through multi-round potential function iteration (NEP1→NEPn), when the extrapolation configuration differentiation rate is below 5% for three consecutive times, the convergence condition is triggered, and the final potential function NEP-Final with cross-scale adaptability is output.

[0092] Step 140: Conduct large-scale molecular dynamics simulations, construct different datasets, train the potential function after feature fusion and cross-validation, perform deduction, statistical deposition rate, and estimate reaction-deposition selectivity.

[0093] Feature fusion in large-scale molecular dynamics simulations can integrate and fuse different data features or information to form a more comprehensive and representative feature representation, which helps improve the model's ability to understand and analyze data.

[0094] Cross-validation is used to evaluate model performance and accuracy. The dataset is divided into several subsets, and one subset is used for training while the other subset is used for validation. The performance of the model is comprehensively evaluated through multiple cross-validations.

[0095] Deposition rate can refer to the number of deposited atoms contributed by a single reactive particle (such as active groups in plasma or sputtered target atoms) on the substrate surface during the deposition process. Statistical deposition rate involves collecting and statistically analyzing the yield data during the deposition process to understand the deposition effect and reaction efficiency.

[0096] The reaction-deposition selectivity ratio refers to the proportional relationship between the deposition rate and yield of different substances or different reaction pathways during the deposition process, reflecting the selectivity and specificity of the deposition process.

[0097] Figure 1The proposed method constructs an initial dataset by analyzing the transition state path; trains an initial neural evolution potential function based on the initial dataset, conducts small-scale extrapolation experiments, establishes an evaluation function to screen differentiated configurations, and performs DFT accuracy verification to form an enhanced training set; after multiple iterations, the final potential function is output when the convergence condition is met; large-scale molecular dynamics simulations are conducted, different datasets are constructed, and the potential function is trained after feature fusion and cross-validation, performing extrapolation, statistical deposition rate analysis, and estimation of reaction-deposition selectivity. By constructing a fusion model of deep neural networks and molecular dynamics, the method overcomes the bottlenecks of accuracy of traditional empirical potential functions and efficiency of first-principles molecular dynamics, accurately analyzing the dynamic evolution law of heterogeneous interface defects; the dynamic combination of process parameters and microscopic simulation environment flexibly identifies problems in real process processes; and by combining real-time monitoring data streams and dynamically updated dynamic models, the process control response time is reduced, achieving microsecond-level real-time closed-loop precise control of atomic layer growth rate, opening a new path for the development of molecular beam epitaxy towards intelligence and high precision.

[0098] based on Figure 1 In addition to the method described herein, this specification also provides some specific implementation methods of the method, which will be described below.

[0099] This invention proposes a machine learning-based molecular dynamics simulation method for molecular beam epitaxy (MBE) processes. The technical solution consists of a multi-scale coupled modeling framework and a dynamic iterative optimization process. The specific implementation process is as follows:

[0100] Transition State Path Analysis and Dataset Initialization: Based on the atomic-level dynamic characteristics of material interface reactions, the transition state path during the deposition process is first analyzed using the Nudged Elastic Band method and the Elastic Band method. Then, an initial dataset containing key reaction configurations, energy barriers, and activation parameters is constructed through first-principles calculations. 00 This dataset utilizes the ASE (Atomic Simulation Environment) platform to achieve 3D visualization and reconstruction of reaction paths, providing high-confidence initial samples for potential function training.

[0101] In step 120, the dynamic potential function iterative optimization mechanism may include:

[0102] The initial neural evolution potential function is trained based on the initial dataset, and a small-scale extrapolation experiment is carried out. By introducing multi-dimensional ensemble conditions, an extrapolation configuration set is generated. The multi-dimensional ensemble conditions include at least temperature gradient perturbation, pressure fluctuation and random noise injection.

[0103] A structural differentiation evaluation function is established, and the extrapolated configuration is compared with the atomic displacement matrix of the initial dataset to screen out the differential configurations that deviate from the ground state distribution.

[0104] Perform DFT accuracy verification on the differentiated configurations, and fuse the energy convergence results with the original dataset to form an enhanced training set.

[0105] For example: based on Dataset 00 An initial neural evolution potential function (NEP0) was trained, and a small-scale extrapolation experiment was conducted using the Calorine module: by introducing multi-dimensional ensemble conditions such as temperature gradient perturbation (300-1500K), pressure fluctuation (0.1-10GPa), and random noise injection, an extrapolation configuration set was generated; a structural differentiation evaluation function was established to evaluate the extrapolation configurations against the dataset. 00 Perform atomic displacement matrix comparison ( Using a baseline threshold, anomalous configurations deviating from the ground-state distribution are selected. DFT accuracy verification is performed on the differentiated configurations, and the energy convergence results are fused with the original dataset to form an enhanced training set, Dataset01. This process utilizes automated scripts for parameter passing and data format conversion, ensuring seamless cross-platform computation.

[0106] In step 130, the convergence criterion and the generation of the specific potential function are determined: through multiple rounds of potential function iteration (NEP1→NEPn), the convergence condition is triggered when the extrapolated configuration differentiation rate is below 5% for three consecutive times, outputting the final potential function NEP-Final, which has cross-scale adaptability. This potential function integrates the short-range action description of the classical EAM potential with the long-range correlation characteristics of the machine learning potential. By dynamically adjusting the potential energy contribution weights at different spatial scales through a gating network, the prediction accuracy of the interfacial response of the laminated material is significantly improved.

[0107] Multiphysics coupling solution for sedimentation dynamics: Large-scale molecular dynamics simulations (>10^6 atomic systems) are conducted based on NEP-Final, combined with an adaptive time-step algorithm to track the spatiotemporal evolution of the sedimentation front, such as... Figure 2 As shown, the initial response model is constructed, the training dataset generation module determines the training dataset, the DFT accuracy verification is performed by the DFT calculation module to form an enhanced training set, and the potential function is determined by the NEP potential function training module. After multiple rounds of iteration, the final potential function is output until the convergence condition is met. This can be used for large-scale system model extrapolation and further data analysis, or it can be used to train the initial neural evolution potential function based on the initial dataset and then perform small-scale inference, i.e., small-scale model extrapolation experiments. An evaluation function is established to screen for differential configurations and make differential judgments, and the steps are repeated.

[0108] The specific steps for multiphysics coupling solution of sedimentation dynamics are as follows:

[0109] First, atomic potential energy information files for different configurations were constructed: Regarding the construction of the crystal relaxation process dataset, authoritative materials databases such as the Materials Project were used to obtain data on various crystal structures (including zincblende, wurtzite, etc.) covering the Si-Ge binary alloy system. First-principles molecular dynamics (AIMD) was used to conduct in-depth simulations of thermodynamic processes over a wide temperature range of 100K to 1500K, employing lattice vibrations and atomic diffusion behavior under supercell expansion temperature gradients. To simulate non-equilibrium dynamic processes in actual manufacturing, strategies such as expansion, lattice perturbation, and vacancy defect introduction were employed: a 2x2x2 supercell was constructed to expand the simulation scale, ±5% lattice constant perturbation was applied to excite atomic displacements, and 0.5%–2% vacancy concentrations were randomly introduced to simulate defect evolution, thus realistically reproducing the dynamic evolution of the crystal structure during epitaxial growth. In the data sampling stage, molecular dynamic trajectory sampling was performed using a 10fs time step. The sampling frequency was dynamically adjusted based on the system energy gradient and atomic displacement variance to ensure the integrity of data for key relaxation stages. After high-precision single-point energy calculation (using HSE06 hybrid functionals combined with plane wave basis sets), the sampled data was imported into a self-developed potential function dataset generation platform. Through data cleaning and feature standardization, a high-quality crystal relaxation process initialization dataset was finally formed. The gas motion and chemical reaction kinetics datasets were obtained by simulating the diffusion behavior of gaseous precursors through molecular dynamics simulation and by searching reaction paths based on transition state theory, respectively. After feature fusion and cross-validation, the three datasets constitute a complete machine learning potential function training set, laying the data foundation for accurately describing the epitaxial growth process of Si / SiGe materials.

[0110] Secondly, based on reaction path information and classical potential function strategies, machine learning potential functions for different stacked systems are constructed: In this stage, by combining an initialization dataset with an active learning algorithm, a precise machine learning potential function (MLP) suitable for molecular beam epitaxy (MBE) processes is constructed. For example... Figure 5 As shown, the specific implementation process is as follows: In the data preparation stage, two types of key data need to be systematically collected: 1) a training dataset of the alloy potential function of the substrate and reaction products; 2) a training dataset of the potential function of the gas dynamics reaction, as shown below. Figure 3 and Figure 4 As shown. These datasets need to include energy / force field calculation results for different deposited film materials, as well as dynamic behavior data of gas molecules interacting with the solid surface, to establish a complete data foundation for subsequent potential function training.

[0111] Initialize Dataset Construction 002. Single-point energy calculations were performed on the atomic configuration using the density functional theory (DFT) self-consistent field (SCF) method to generate an initial dataset containing energy and force field information. During this process, strict data quality control was implemented: a) removing non-converged calculation results; b) filtering outliers in the force field energy; and c) verifying the physical rationality of the atomic configuration to ensure the accuracy of the training data.

[0112] The potential function iterative optimization process includes: potential function training (NEP) i Based on the Neural Evolutionary Potential Function (NEP) framework, a radial / angular descriptor with a 4 Å cutoff radius is set, and multi-scale training is performed by integrating structural data obtained from alloy systems, gas-phase reactions, and AIMD simulations. Particular emphasis is placed on optimizing the matching degree between descriptor parameters and material properties. Active learning validation: Extrapolation validation is implemented in a 1-2 nm scale MBE process model. The generalization ability of the potential function is evaluated by introducing interference factors such as: a) temperature gradient changes; b) impurity concentration gradients; and c) defect configurations. When the extrapolated configuration matches the dataset... i When the prediction error exceeds the threshold, the configuration is included in the dataset after DFT verification. i+1 Convergence Criteria and Output: When the improvement rate of the potential function due to newly added data in consecutive iterations is less than 5%, convergence is determined, and the final potential function NEP is output. Final This potential function will accurately describe key physical processes during MBE, including: a) gas-solid interface reactions; b) thin film growth kinetics; and c) defect evolution, providing high-fidelity force field support for molecular dynamics simulations.

[0113] Then, the accuracy of the specific potential function is evaluated: To achieve the above goal, a technical solution combining potential energy surface sampling and visualization with radial distribution function (RDF) analysis is first adopted. For example, Figure 6-8 As shown, potential energy surface sampling visualization can intuitively present the energy distribution characteristics of atoms in different spatial positions and configurations, providing key visualization basis for analyzing the complex mechanisms of interatomic interactions. By performing refined sampling and visualization processing on the potential energy surface, it is possible to clearly identify energy extreme value regions (including minimum and maximum values) and accurately capture the transition paths of atoms between these regions.

[0114] Meanwhile, radial distribution function (RDF) analysis provides statistical characteristics of the interatomic distance distribution. This method can quantitatively characterize the distribution probability of atoms at different spacings, thereby revealing the short-range and long-range ordered structural features of atomic systems. By comparing RDF curves under different conditions, the accuracy of the potential function in describing interatomic interactions and the stability level of atomic structures in different environments can be effectively evaluated.

[0115] Furthermore, this study introduces the potential energy surface distribution characteristics between deposited atoms and the substrate to deeply characterize the energy barrier height during atomic diffusion and reconstruction. This method can directly reflect the energy barrier encountered by atoms during diffusion and reconstruction on the substrate surface during molecular beam epitaxy (MBE). Through accurate calculation and systematic analysis of the aforementioned energy barrier height, effective prediction of atomic diffusion rates and reconstruction mechanisms can be achieved, thus providing key theoretical support for the optimization of molecular beam epitaxy (MBE) processes.

[0116] Next, the potential function is used for simulation through layered machine learning: after verifying the accuracy of the potential function, the process simulation stage begins. In this stage, a high-precision MBE process simulation model is constructed by coupling the physical properties of the incident particles with the dynamic behavior of the interface.

[0117] The physical properties of Ge (atomic / molecular beams) are analyzed. First, based on the particle distribution at the interface (including particle flux), the incident particles (such as Si, incident angle distribution, and velocity vector) are mapped to the input parameters of the stacked structure deposition kinetics model using the Boltzmann transport equation. Considering the difference between the interface temperature and the cavity temperature in actual processes, the interface temperature is dynamically corrected to 1.1–1.3 times the cavity temperature to more accurately reflect local thermal effects.

[0118] Secondly, such as Figure 9 As shown, a multi-scale initial substrate model is constructed: a periodic boundary condition is used to simulate an infinitely large substrate, and a surface step structure with a specific orientation (such as Si100) is generated by atomic layer stacking. The step height is set to a single atomic layer (~0.3 nm), and 0.5%~1% surface vacancy defects are introduced to simulate the non-ideal state of a real substrate. The model size is optimized to 20×20×10 nm (xyz direction) based on computational resources, so as to ensure computational efficiency while fully preserving the interface dynamics characteristics.

[0119] Subsequently, a parallel molecular dynamics simulation platform was launched. Based on the CUDA-accelerated GPUMD software (which offers tens of times greater computational efficiency than traditional CPUs), real MBE process parameters were loaded into the NVT ensemble: the temperature was set to the experimental chamber temperature (e.g., 500–800°C), and a constant temperature environment was maintained using a Nosé-Hoover thermal bath; the particle incident period was set to 0.1 µs, meaning each simulation run contained 100,000 time steps (1 fs). To avoid invalid calculations, particle trapping boundary conditions were defined: particles were automatically deleted when their deposition height exceeded 20 Å (approximately 2 nm) above the interface, ensuring the simulation focused on the effective growth region.

[0120] Finally, a multi-round iterative simulation strategy was implemented: the first round of simulation used an ideal substrate model (defect-free) to obtain basic growth kinetic data; subsequent rounds gradually introduced random interface defects (such as dislocation loops and grain boundaries) and beam perturbations (±5% flux fluctuation). By comparing deposition rates, surface roughness (root mean square error <0.5nm), and interface defect densities under different operating conditions, a mapping relationship between process parameters and structural performance was established. Under typical operating conditions, a single simulation at the scale of 100,000 atoms can be completed within hours (NVIDIA A100x4 cluster configuration), providing rapid feedback for process optimization.

[0121] Finally, the deposition rate is statistically analyzed, and the reaction-deposition selectivity ratio is estimated. In this step, if the structure being modeled is a layered material, the different layers need to be numbered and grouped in advance to facilitate effective screening and statistical analysis after kinetic simulation. This grouping strategy helps to clarify the performance of each layer during the deposition process, thereby better understanding the influence of different materials on the overall deposition rate. If the system being modeled consists of different monolithic materials, numbering and grouping are not required, and statistical analysis can be performed directly on different systems.

[0122] Next, dynamic simulations are performed to record the interaction between each incident particle and the material interface. During this process, it is necessary to statistically analyze the deposition rate of each layer or monomer material, i.e., the amount of material successfully deposited per unit time. This data will provide a basis for subsequent selectivity estimation.

[0123] After statistically analyzing the deposition rates, the reaction-deposition selectivity ratio can be calculated. The selectivity ratio is typically defined as the ratio of the deposition rate of one specific material to that of another material; a higher selectivity ratio indicates a better deposition process. By analyzing the deposition behavior of different materials under the same conditions, the relative selectivity between materials can be assessed, providing guidance for optimizing the deposition process.

[0124] The technical solution provided by this invention constructs an MLP (Machine Learning Potential) potential function model that integrates interfacial atomic interactions; accurately analyzes the deposition kinetics of heteroepitaxial systems such as III-V compound semiconductors / topological insulators; and breaks through the limitations of traditional empirical potential functions in describing complex material interfaces (bonding reconstruction, lattice mismatch, etc.).

[0125] Within picosecond to millisecond timescales, the morphological evolution of reactants on the substrate surface is predicted, including atomic-level phenomena such as step flow and two-dimensional island nucleation; by statistically analyzing the quantitative relationship between the deposition atomic distribution density and the deposition rate, the deposition coefficient of different material systems is estimated.

[0126] High-throughput computing is used to screen substrate-deposition source combinations (such as heterostructures like Si / SiGe and GaAs / Si) and generate process parameter-material property correlation maps. Based on these correlation maps, process optimization suggestions are provided, such as strategies for controlling atomic layer growth rates and schemes for interface defect formation.

[0127] To address the bottlenecks in traditional MBE processes, such as unclear interface defect mechanisms, difficulties in optimizing high-dimensional parameters, and lag in real-time control, this paper integrates multi-source datasets of crystal relaxation, gas motion, and chemical reaction kinetics. It utilizes active learning to iteratively optimize the neuroevolutionary potential function (NEP) and combines this with CUDA-accelerated GPUMD molecular dynamics simulations to achieve dynamic simulation of the deposition process with 0.1 nm spatial resolution and 1 ps temporal resolution. This method can accurately capture atomic-level phenomena such as step flow and two-dimensional island nucleation, providing an atomic-mechanism-based analytical solution for heteroepitaxy of advanced materials such as III-V semiconductors and topological insulators.

[0128] Based on the same idea, this invention also provides a molecular dynamics simulation device for molecular beam epitaxy processes, such as... Figure 10 As shown, the device may include:

[0129] Initialize dataset building module 1010, which is used to parse the transition path and build the initial dataset;

[0130] The enhanced training set determination module 1020 is used to train the initial neural evolution potential function based on the initialization dataset, conduct small-scale extrapolation experiments, establish an evaluation function to screen differential configurations, and perform DFT accuracy verification to form an enhanced training set.

[0131] The final potential function determination module 1030 is used to output the final potential function after multiple rounds of iteration until the convergence condition is met.

[0132] The reaction-deposition selectivity determination module 1040 is used to conduct large-scale molecular dynamics simulations, construct different datasets, train potential functions after feature fusion and cross-validation, and perform deduction, statistical deposition rate calculation and estimate reaction-deposition selectivity.

[0133] based on Figure 10 The device may also include specific implementation units:

[0134] Optionally, the initialization dataset building module 1010 may include:

[0135] The transition state path analysis unit is used to analyze the transition state path in the deposition process based on the atomic-level dynamic characteristics of material interface reactions, employing the climbing method and elastic band method.

[0136] An initial dataset determination unit is used to construct an initial dataset containing key reaction configurations, energy barriers, and activation parameters using first principles; the initial dataset is used to achieve three-dimensional visualization reconstruction of the reaction path through the ASE platform;

[0137] The final potential function determination module 1030 can be used for:

[0138] Through multiple rounds of potential function iteration, when the extrapolation configuration differentiation rate is lower than the preset value for several consecutive times, the convergence condition is triggered, and the final potential function with cross-scale adaptability is output; the final potential function is a combination of descriptor operators and multi-layer artificial neural networks.

[0139] Optionally, the enhanced training set determination module 1020 may include:

[0140] An extrapolation configuration set generation unit is used to train an initial neural evolution potential function based on the initialization dataset, implement a small-scale extrapolation experiment, and generate an extrapolation configuration set by introducing multi-dimensional ensemble conditions; the multi-dimensional ensemble conditions include at least temperature gradient perturbation, pressure fluctuation and random noise injection.

[0141] The differential configuration determination unit is used to establish a structural differential evaluation function, compare the extrapolated configuration with the initial dataset using atomic displacement matrices, and screen out differential configurations that deviate from the ground state distribution;

[0142] An enhanced training set determination unit is used to perform DFT accuracy verification on the differentiated configuration and fuse the energy convergence result with the original dataset to form an enhanced training set.

[0143] Optionally, the reaction deposition selectivity determination module 1040 may include:

[0144] Atomic potential energy information file construction unit, used to construct atomic potential energy information files with different configurations;

[0145] The machine learning potential function building unit is used to construct machine learning potential functions for different stacked systems based on reaction path information and classical potential function strategies.

[0146] The reaction-deposition selectivity estimation unit is used to evaluate the accuracy of the dedicated potential function and extrapolate the potential function through stacked machine learning to calculate the deposition rate and estimate the reaction-deposition selectivity.

[0147] Optionally, the atomic potential energy information file building block can be used for:

[0148] Obtain crystal structure data covering a variety of Si-Ge binary alloy systems;

[0149] Thermodynamic processes are simulated within a pre-defined wide temperature range using first-principles molecular dynamics methods.

[0150] The supercell expansion simulation scale is constructed by using lattice vibration and atomic diffusion behavior under supercell expansion temperature gradient. A pre-set lattice constant perturbation is applied to excite atomic displacement, and a pre-set vacancy concentration is randomly introduced to simulate defect evolution.

[0151] Molecular dynamics trajectory sampling is performed based on a preset time step, and the sampling frequency is dynamically adjusted according to the system energy gradient and atomic displacement variance. After high-precision single-point energy calculation, the sampled data is imported into a self-developed potential function dataset generation platform. After data cleaning and feature standardization, a high-quality crystal relaxation process initialization dataset is formed.

[0152] The three datasets were obtained by simulating the diffusion behavior of gaseous precursors through molecular dynamics simulation and by searching reaction paths based on transition state theory. After feature fusion and cross-validation, they constituted a complete training set for machine learning potential functions.

[0153] Optionally, machine learning potential function building blocks can be used for:

[0154] Obtain key datasets; the key datasets include at least the alloy potential function training dataset of the substrate and reaction products, and the potential function training dataset of the gas dynamics reaction;

[0155] A single-point energy calculation of atomic configurations is performed using the density functional theory self-consistent field method to generate an initial dataset containing energy and force field information.

[0156] Perform data quality control to remove non-converged calculation results, filter out abnormal values ​​of force field energy, and verify the physical rationality of atomic configurations to obtain valid data;

[0157] Based on the neuroevolutionary potential function framework, the obtained structural data is trained at multiple scales and extrapolated for verification. Target interference factors are introduced to evaluate the generalization ability of the potential function. The target interference factors include at least temperature gradient changes, impurity concentration gradients, and defect configurations.

[0158] When the extrapolation configuration and the prediction error exceed the threshold, the extrapolation configuration is included in the dataset after being verified by DFT.

[0159] When the improvement rate of the potential function caused by the addition of new data in continuous iteration is less than the preset improvement rate, it is determined to be converged, and the final potential function is output.

[0160] Optionally, the reactive deposition selectivity estimation unit can be used for:

[0161] By combining potential energy surface sampling visualization with radial distribution function analysis, the potential energy surface sampling visualization presents the energy distribution characteristics of atoms in different spatial positions and configurations, identifies energy extreme regions and captures atomic transition paths;

[0162] Radial distribution function analysis provides statistical characteristics of interatomic distance distribution. By comparing RDF curves under different conditions, the accuracy of the potential function in describing interatomic interactions can be evaluated. The potential energy surface distribution characteristics between deposited atoms and substrate are introduced to characterize the energy barrier height during atomic diffusion and reconstruction, and to predict atomic diffusion rate and reconstruction mechanism.

[0163] The physical properties of the incident particles are analyzed, and the particle distribution at the interface is mapped to the input parameters of the cascade structure deposition dynamics model through the Boltzmann transport equation.

[0164] The interface temperature is dynamically corrected to a preset multiple of the cavity temperature. Periodic boundary conditions are used to simulate an infinitely large substrate, and a surface step structure with a specific crystal orientation is generated through atomic stacking.

[0165] The step height is set to a single atomic layer, and surface vacancy defects are introduced; the model size is optimized to a preset size based on computational resources.

[0166] A parallel molecular dynamics simulation platform was launched. The simulation used an ideal substrate model to obtain basic growth dynamics data. Random interface defects and beam disturbances were gradually introduced in each round. By comparing the deposition rate, surface roughness and interface defect density under different working conditions, the mapping relationship between process parameters and structural performance was established.

[0167] If the structure being modeled is a layered material, the different layers should be numbered and grouped in advance; if the system being modeled consists of different monolithic materials, then numbering and grouping are not necessary.

[0168] Perform dynamic simulations, record the interaction between each incident particle and the material interface, and statistically analyze the deposition rate of each stacked material or single material.

[0169] The reaction deposition selectivity is calculated. The selectivity is defined as the ratio of the deposition rate of a particular material to the deposition rate of another material. The larger the deposition selectivity, the higher the deposition selectivity.

[0170] Based on the same approach, embodiments of this specification also provide molecular dynamics simulation equipment for molecular beam epitaxy processes. For example... Figure 11 As shown, the device includes:

[0171] The system includes a memory, a processor, and a communication interface coupled to the processor; the memory stores a computer program that can be run by the processor; when the processor runs the computer program, it executes the aforementioned molecular dynamics simulation method for molecular beam epitaxy.

[0172] like Figure 11 As shown, the processor described above can be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. The communication interface described above can be one or more. The communication interface can use any transceiver-like device for communicating with other devices or communication networks.

[0173] like Figure 11 As shown, the terminal device described above may also include a communication line. The communication line may include a path for transmitting information between the components described above.

[0174] Optional, such as Figure 11 As shown, the terminal device may further include a memory. The memory stores a computer program that can be executed by the processor; when the processor executes the computer program, it implements the method provided in the embodiments of the present invention.

[0175] like Figure 11As shown, the memory can be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed discs, laser discs, optical discs, digital universal discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited to these. The memory can exist independently and be connected to the processor via communication lines. The memory can also be integrated with the processor.

[0176] Optionally, the computer execution instructions in the embodiments of the present invention may also be referred to as application code, and the embodiments of the present invention do not specifically limit this.

[0177] In a specific implementation, as one example, such as Figure 11 As shown, a processor may include one or more CPUs, such as Figure 11 CPU0 and CPU1 in the CPU.

[0178] In a specific implementation, as one example, such as Figure 11 As shown, the terminal device may include multiple processors, such as Figure 11 The processors in the system. Each of these processors can be a single-core processor or a multi-core processor.

[0179] Based on the same idea, this specification also provides a computer storage medium corresponding to the above embodiments. The computer storage medium stores instructions that, when executed, implement the methods in the above embodiments.

[0180] The foregoing mainly describes the solutions provided by the embodiments of the present invention from the perspective of the interaction between various modules. It is understood that each module, in order to achieve the above functions, includes corresponding hardware structures and / or software units for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art 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 invention.

[0181] The embodiments of the present invention can divide functional modules according to the above method examples. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in the embodiments of the present invention is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0182] The processor described in this specification may also function as a memory. The memory stores computer execution instructions for carrying out the present invention, and its execution is controlled by the processor. The processor executes the computer execution instructions stored in the memory, thereby implementing the method provided in the embodiments of the present invention.

[0183] The memory can be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed discs, laser discs, optical discs, universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited to these. The memory can exist independently and be connected to the processor via communication lines. The memory can also be integrated with the processor.

[0184] Optionally, the computer execution instructions in the embodiments of the present invention may also be referred to as application code, and the embodiments of the present invention do not specifically limit this.

[0185] The methods disclosed in the above embodiments of the present invention can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an ASIC, a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of the present invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0186] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0187] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A molecular dynamics simulation method for molecular beam epitaxy, characterized in that, The methods include: Parse the transitional path and construct the initial dataset; The initial neural evolution potential function is trained based on the initial dataset, and a small-scale extrapolation experiment is conducted. An evaluation function is established to screen for differential configurations, and DFT accuracy verification is performed to form an enhanced training set. After multiple rounds of iteration, the final potential function is output until the convergence condition is met. Large-scale molecular dynamics simulations were conducted to construct different datasets. After feature fusion and cross-validation, the potential function was trained, and the deposition rate was extrapolated, statistically analyzed, and the reaction-deposition selectivity was estimated.

2. The molecular dynamics simulation method for molecular beam epitaxy as described in claim 1, characterized in that, Parse the transitional path and construct the initial dataset, including: Based on the atomic-level dynamic characteristics of material interface reactions, the climbing method and elastic band method are used to analyze the transition state path in the deposition process; An initial dataset containing key reaction configurations, energy barriers, and activation parameters is constructed using first-principles calculations; the initial dataset is then used to reconstruct a three-dimensional visualization of the reaction pathway through the ASE platform. After multiple iterations, until the convergence condition is met, the final potential function is output, including: Through multiple rounds of potential function iteration, when the extrapolation configuration differentiation rate is lower than the preset value for several consecutive times, the convergence condition is triggered, and the final potential function with cross-scale adaptability is output; the final potential function is a combination of descriptor operators and multi-layer artificial neural networks.

3. The molecular dynamics simulation method for molecular beam epitaxy as described in claim 1, characterized in that, An initial neural evolution potential function is trained based on an initial dataset, and a small-scale extrapolation experiment is conducted. An evaluation function is established to screen for differential configurations, and DFT accuracy verification is performed to form an enhanced training set, including: The initial neural evolution potential function is trained based on the initial dataset, and a small-scale extrapolation experiment is carried out. By introducing multi-dimensional ensemble conditions, an extrapolation configuration set is generated. The multi-dimensional ensemble conditions include at least temperature gradient perturbation, pressure fluctuation and random noise injection. A structural differentiation evaluation function is established, and the extrapolated configuration is compared with the atomic displacement matrix of the initial dataset to screen out the differential configurations that deviate from the ground state distribution. Perform DFT accuracy verification on the differentiated configurations, and fuse the energy convergence results with the original dataset to form an enhanced training set.

4. The molecular dynamics simulation method for molecular beam epitaxy as described in claim 1, characterized in that, Large-scale molecular dynamics simulations were conducted, different datasets were constructed, and potential functions were trained after feature fusion and cross-validation. Deposition rates were extrapolated, statistically analyzed, and reaction-deposition selectivity was estimated, including: Construct atomic potential energy information files with different configurations; Based on reaction path information and classical potential function strategies, machine learning potential functions for different stacked systems are constructed. The accuracy of the dedicated potential function is evaluated, and the deposition rate is calculated and the reaction deposition selectivity is estimated by extrapolating the potential function through stacked machine learning.

5. The molecular dynamics simulation method for molecular beam epitaxy according to claim 4, characterized in that, Construct atomic potential energy information files with different configurations, including: Obtain crystal structure data covering a variety of Si-Ge binary alloy systems; Thermodynamic processes are simulated within a pre-defined wide temperature range using first-principles molecular dynamics methods. The supercell expansion simulation scale is constructed by using lattice vibration and atomic diffusion behavior under supercell expansion temperature gradient. A pre-set lattice constant perturbation is applied to excite atomic displacement, and a pre-set vacancy concentration is randomly introduced to simulate defect evolution. Molecular dynamics trajectory sampling is performed based on a preset time step, and the sampling frequency is dynamically adjusted according to the system energy gradient and atomic displacement variance. After high-precision single-point energy calculation, the sampled data is imported into a self-developed potential function dataset generation platform. After data cleaning and feature standardization, a high-quality crystal relaxation process initialization dataset is formed. The three datasets were obtained by simulating the diffusion behavior of gaseous precursors through molecular dynamics simulation and by searching reaction paths based on transition state theory. After feature fusion and cross-validation, they constituted a complete training set for machine learning potential functions.

6. The molecular dynamics simulation method for molecular beam epitaxy according to claim 4, characterized in that, Based on reaction path information and classical potential function strategies, machine learning potential functions for different stacked systems are constructed, including: Obtain key datasets; the key datasets include at least the alloy potential function training dataset of the substrate and reaction products, and the potential function training dataset of the gas dynamics reaction; A single-point energy calculation of atomic configurations is performed using the density functional theory self-consistent field method to generate an initial dataset containing energy and force field information. Perform data quality control to remove non-converged calculation results, filter out abnormal values ​​of force field energy, and verify the physical rationality of atomic configurations to obtain valid data; Based on the neuroevolutionary potential function framework, the obtained structural data is trained at multiple scales and extrapolated for verification. Target interference factors are introduced to evaluate the generalization ability of the potential function. The target interference factors include at least temperature gradient changes, impurity concentration gradients, and defect configurations. When the extrapolation configuration and the prediction error exceed the threshold, the extrapolation configuration is included in the dataset after being verified by DFT. When the improvement rate of the potential function caused by the addition of new data in continuous iteration is less than the preset improvement rate, it is determined to be converged, and the final potential function is output.

7. The molecular dynamics simulation method for molecular beam epitaxy according to claim 1, characterized in that, The accuracy of the specific potential function is evaluated, and the deposition rate is calculated and estimated using stacked machine learning potential functions. This includes: By combining potential energy surface sampling visualization with radial distribution function analysis, the potential energy surface sampling visualization presents the energy distribution characteristics of atoms in different spatial positions and configurations, identifies energy extreme regions and captures atomic transition paths; Radial distribution function analysis provides statistical characteristics of interatomic distance distribution. By comparing RDF curves under different conditions, the accuracy of the potential function in describing interatomic interactions can be evaluated. The potential energy surface distribution characteristics between deposited atoms and substrate are introduced to characterize the energy barrier height during atomic diffusion and reconstruction, and to predict atomic diffusion rate and reconstruction mechanism. The physical properties of the incident particles are analyzed, and the particle distribution at the interface is mapped to the input parameters of the cascade structure deposition dynamics model through the Boltzmann transport equation. The interface temperature is dynamically corrected to a preset multiple of the cavity temperature. Periodic boundary conditions are used to simulate an infinitely large substrate, and a surface step structure with a specific crystal orientation is generated through atomic stacking. The step height is set to a single atomic layer, and surface vacancy defects are introduced; the model size is optimized to a preset size based on computational resources. A parallel molecular dynamics simulation platform was launched. The simulation used an ideal substrate model to obtain basic growth dynamics data. Random interface defects and beam disturbances were gradually introduced in each round. By comparing the deposition rate, surface roughness and interface defect density under different working conditions, the mapping relationship between process parameters and structural performance was established. If the structure being modeled is a layered material, the different layers should be numbered and grouped in advance; if the system being modeled consists of different monolithic materials, then numbering and grouping are not necessary. Perform dynamic simulations, record the interaction between each incident particle and the material interface, and statistically analyze the deposition rate of each stacked material or single material. The reaction deposition selectivity is calculated. The selectivity is defined as the ratio of the deposition rate of a particular material to the deposition rate of another material. The higher the deposition selectivity, the better.

8. A molecular dynamics simulation device for molecular beam epitaxy, characterized in that, The device includes: The initial dataset building module is used to parse the transition path and build the initial dataset; The enhanced training set determination module is used to train the initial neural evolution potential function based on the initialization dataset, conduct small-scale extrapolation experiments, establish an evaluation function to screen differential configurations, and perform DFT accuracy verification to form an enhanced training set. The final potential function determination module is used to output the final potential function after multiple rounds of iteration until the convergence condition is met. The reaction-deposition selectivity determination module is used to conduct large-scale molecular dynamics simulations, construct different datasets, train the potential function after feature fusion and cross-validation, and perform deduction, statistical deposition rate calculation and estimate reaction-deposition selectivity.

9. A molecular dynamics simulation device for molecular beam epitaxy, characterized in that the device... include: Memory, processor, and communication interface coupled to the processor; The memory stores computer programs that can be executed by the processor; When the processor runs the computer program, it executes the molecular dynamics simulation method for molecular beam epitaxy as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores instructions that, when executed by the processor, implement the molecular dynamics simulation method for molecular beam epitaxy as described in any one of claims 1 to 7.

Images