A method and system for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN based on machine learning atomic potential.

By constructing a machine learning atomic potential model covering different material compositions and defect local environments, the high cost problem of large-scale, long-term lattice thermal conductivity calculation was solved, achieving efficient and accurate thermal conductivity prediction, and providing a computational basis for the thermal management of GaN, AlN, and AlGaN materials.

CN123120106APending Publication Date: 2026-11-17NANJING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611620233.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for calculating lattice thermal conductivity are costly in terms of computational resources and time in large-scale, long-term simulations, and traditional empirical atomic potentials are difficult to accurately describe the local atomic environment of different material compositions and defects.

Method used

A machine learning-based atomic potential approach was adopted. By constructing a first-principles dataset covering different material compositions and defect local environments, a neural evolution potential model was trained to predict the atomic defect thermal conductivity of GaN, AlN, and AlGaN. Large supercell simulations were then performed using a graphics processor-based molecular dynamics program.

Benefits of technology

It significantly reduces computational resource and time requirements while maintaining high accuracy, and can accurately predict the impact of different material compositions and defect types on thermal conductivity, providing a basis for thermal transport mechanism analysis and thermal management design.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120106A_ABST
    Figure CN123120106A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN based on machine learning atomic potentials. The method involves constructing atomic structures with different atomic arrangements and defect types, acquiring energy, atomic force, and stress data of the structures, training the machine learning atomic potential based on the data, and verifying its reliability. The atomic structures are adjusted according to the atomic defect configurations to establish simulation systems containing different atomic defects, setting different atomic defect types, defect concentrations, and temperatures. Molecular dynamics simulations are performed using atomic potentials to calculate the lattice thermal conductivity of each system and compare it with the corresponding defect-free system. Finally, the quantitative relationship between defect type, concentration, alloy composition, and thermal conductivity is output. This invention can uniformly model multiple atomic defects in GaN, AlN, and AlGaN, maintain high prediction accuracy, reduce computational costs, and provide a computational method for analyzing the influence of atomic defects on lattice thermal conductivity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method and system for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN based on machine learning atomic potential, belonging to the field of material thermal transport calculation and machine learning atomic potential technology. Background Technology

[0002] GaN, AlN, and AlGaN are typical group III nitride semiconductor materials, possessing wide bandgap, high breakdown field strength, and good thermal stability, and are widely used in high-power electronic devices, high-frequency devices, and optoelectronic devices. As device power density continues to increase, the heat generated during operation leads to an increase in device temperature, which in turn affects device performance, stability, and lifespan. Therefore, accurately obtaining the lattice thermal conductivity of GaN, AlN, and AlGaN with different Al compositions is of great significance for analyzing the material's thermal transport mechanism and for device thermal management.

[0003] The lattice thermal conductivity of GaN, AlN, and AlGaN is primarily determined by phonon transport and is influenced by factors such as material composition, temperature, and crystal defects. Atomic defects cause changes in local mass distribution, lattice structure, and interatomic interactions, thus altering lattice thermal conductivity by affecting phonon scattering. The degree of influence of different types and concentrations of atomic defects on thermal conductivity varies, and the lattice thermal conductivity of AlGaN also varies with Al composition.

[0004] Currently, the calculation of lattice thermal conductivity mainly employs first-principles phonon transport calculations and molecular dynamics methods based on empirical atomic potentials. First-principles methods can accurately obtain atomic structure, system energy, and atomic forces, but their computational cost increases significantly with the number of atoms in the system and the simulation time, making them difficult to directly apply to large-scale, long-term thermal transport simulations involving tens of thousands of atoms. Especially when studying various material compositions, defect types, and defect concentrations, it is necessary to establish numerous defect configurations and perform repeated calculations, resulting in high computational resource and time costs. While traditional molecular dynamics methods can expand the simulation system and extend the simulation time, their calculation results are highly dependent on empirical atomic potentials. Existing empirical atomic potentials are often insufficient to simultaneously and accurately describe the complex local atomic environments formed by Ga-N bonds, Al-N bonds, and different atomic defects.

[0005] In recent years, machine learning-based atomic potentials have been able to utilize system energy, atomic forces, and stress data obtained from first-principles calculations to establish a mapping relationship between the local atomic environment and the system's potential energy, improving the efficiency of molecular dynamics simulations while approaching the accuracy of first-principles calculations. The range of local atomic environments covered by the training data directly affects the reliability of atomic potential predictions. For GaN, AlN, and AlGaN with different Al compositions, Ga-N bonds, Al-N bonds, and Al / Ga mixtures occupy different nearest-neighbor atomic environments; atomic defects alter the coordination state, bond lengths, and local structural symmetry of atoms near the defects, forming local atomic environments distinct from defect-free crystals.

[0006] Therefore, it is necessary to construct a first-principles dataset covering GaN, AlN, AlGaN with different Al compositions, and various atomic defect local environments. The prediction accuracy of machine learning atomic potentials under different material compositions and defect configurations should be verified. For material compositions or defect local environments with insufficient prediction accuracy, first-principles data should be supplemented and retrained to obtain machine learning atomic potentials that can be used for thermal transport simulation of large-scale defect systems. Summary of the Invention

[0007] This invention proposes a method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN based on machine learning atomic potentials. This method addresses the challenges of first-principles methods in directly conducting large-scale, long-term thermal transport simulations, as well as the difficulty of using traditional empirical atomic potentials to uniformly and accurately describe the local atomic environment of different material compositions and defects.

[0008] The technical solution adopted in this invention is as follows: A method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN based on machine learning atomic potential includes the following steps: (1) Construct small supercells containing n atoms for GaN, AlN and AlGaN with different Al compositions respectively. Set random Al or Ga atom arrangements in the small supercells and introduce randomly distributed atomic defects to obtain atomic configurations for first-principles calculations. The n is a multiple of 4 and is not less than 32 and not greater than 512. (2) First-principles calculation optimization is performed on the small supercell atomic configurations obtained in step (1). The optimized structures are used as the initial configurations. First-principles molecular dynamics calculations are performed at different temperature points within the range of 200-500 K with fixed temperature intervals. Each run is 800-1200 steps, and an atomic configuration is selected every 10-30 steps. The configurations selected in the structure optimization process and molecular dynamics trajectory are summarized. The unit cell parameters, atomic element types and atomic coordinates of each configuration are extracted, and the corresponding system energy, atomic force and stress tensor are used as annotation data. The atomic potential function is trained using the training dataset, and the prediction error of the model is evaluated using the validation dataset. When the prediction accuracy does not meet the preset requirements, atomic configurations are added and first-principles annotations are performed. The new samples are added to the training dataset, and training and validation are performed again until the prediction accuracy meets the preset requirements. (3) The Neuroevolution Potential (NEP) algorithm is used in the Graphics Processor Molecular Dynamics (GPUMD) program to generate local environment descriptors containing radial and angular components based on the element types of each atom in the training dataset and the relative positions of atoms in their neighborhoods. The mapping relationship between the descriptors and atomic energy is established, and the model parameters are optimized using system energy, atomic force and stress data to establish the NEP model. The prediction accuracy of the trained NEP model is tested using the validation dataset. If the preset requirements are not met, first-principles data of different material compositions or defect local environments are added and retrained. If the preset requirements are met, proceed to step (4). (4) Based on a wurtzite cell containing 4 atoms, expand the cell along the three lattice vector directions x, y, and z to construct a defect-free supercell of GaN, AlN, or AlGaN. The expansion factor is adjusted according to the calculation requirements. The number of lattice sites in the defect-free supercell is greater than the number of atoms in the small supercell. When studying AlGaN alloys, Al atoms and Ga atoms are randomly arranged at the cation sites according to the target Al composition to construct an AlGaN random alloy model. Atomic defects are randomly introduced into the defect-free supercell. The number of defects is determined according to the predetermined defect concentration and the corresponding number of reference lattice sites to obtain the target supercell containing defects. The defect-free supercell before the introduction of defects is used as the reference supercell. (5) The validated NEP model is used for the target supercell and the reference supercell constructed in step (4), respectively; for each supercell, the initial temperature is set to the target temperature T, and the initial velocity setting command "velocity Tseed" is used in GPUMD. The initial atomic velocities are generated by 's', where T is in Kelvin and 's' is a random seed number. Five independent simulations are assigned different seed numbers, such as 1, 2, 3, 4, and 5. Under the same initial atomic configuration and temperature, the program generates five different atomic velocity configurations by changing the random seed. Each configuration contains velocity components of all atoms along the x, y, and z directions. During velocity initialization, the program automatically sets the total bus momentum and total angular momentum of the system to zero to eliminate overall translation and rotation. Then, temperature equilibration is performed starting from each velocity configuration, followed by thermal transport simulations to reduce statistical fluctuations in a single simulation. Temperature equilibration and uniform nonequilibrium molecular dynamics simulations are performed in GPUMD to obtain the lattice thermal conductivity of the large supercell at the target temperature through five independent calculations. The arithmetic mean of these calculations is taken as the lattice thermal conductivity of the large supercell at the target temperature. (6) Under the same temperature and thermal transport calculation conditions, compare the lattice thermal conductivity of the target supercell with defects and the supercell without defects with the same material composition and initial Al and Ga atomic arrangement to obtain the influence of defect type and defect concentration on lattice thermal conductivity; under defect-free conditions, compare the lattice thermal conductivity of AlGaN supercells with different Al compositions and random Al and Ga atomic arrangement to obtain the variation law of lattice thermal conductivity under Al composition and random atomic arrangement conditions.

[0009] Preferably, step (1) specifically includes: Construct GaN, AlN and Al with hexagonal wurtzite structure x Ga 1-x N small supercells, of which Al x Ga 1-x The N-type supercell is configured with Al and Ga atoms at cation sites according to the target Al composition x. The number of atoms in the supercell is selected from 32, 48, 64, 72, 80, 96, 108, 112, 120, 128, 144, 160, 168, 180, 192, 200, 216, 224, 240, 252, 256, and 512. In the supercell, the types, numbers, or spatial positions of atoms are adjusted according to a predetermined atomic defect configuration to set atomic defects. Multiple supercell atomic configurations with different local atomic environments are constructed for different material compositions, atomic defect types, defect positions, or defect concentrations. The atomic defects include one or more of Ga vacancies, Al vacancies, N vacancies, NN splitting gap defects, carbon-substituted nitrogen defects (C_N), oxygen-substituted nitrogen defects (O_N), and silicon-substituted cation defects (Si_Al or Si_Ga).

[0010] Preferably, step (2) specifically includes: First-principles structure optimization was performed on all the small supercell atomic configurations constructed in step (1), with each atomic configuration treated as an independent computational system; subsequently, using the optimized structures as the initial configurations, optimization was performed at 10 kJ / kL within the range of 200–500 K. K represents 31 temperature points at intervals. First-principles molecular dynamics calculations are performed at each temperature point with a time step of 1 femtosecond. Each run consists of 1000 steps, corresponding to a simulation duration of 1 picosecond. Trajectories are recorded frame by frame, and an atomic configuration is extracted every 20 frames, i.e., sampling occurs every 20 femtoseconds. 50 atomic configurations are extracted from each trajectory. The above atomic configurations are combined with those selected during structure optimization to form an atomic configuration set for subsequent first-principles annotation. For each selected atomic configuration, the unit cell parameters, element type and atomic coordinates of each atom, as well as the total energy of the system corresponding to that configuration, the forces on each atom, and the stress tensor are recorded to form a one-to-one corresponding data sample. The data samples are divided into non-repeating training datasets and validation datasets according to the complete atomic configurations, and then organized into extended XYZ format files readable by the NEP training program.

[0011] Preferably, in step (2), a single atomic configuration and its corresponding cell parameters, atomic element types, atomic coordinates, system energy, atomic forces, and stress tensors are used as a complete data sample. After performing a repeat configuration check on the initial sample, a portion of the samples are randomly selected to form the validation dataset, and the remaining samples form the training dataset. The same atomic configuration is not repeatedly assigned to the training dataset and the validation dataset. The first-principles samples added during subsequent training and validation processes are added to the training dataset after a repeat configuration check, while the validation dataset remains unchanged. The training dataset is used to optimize the model parameters. The validation dataset is used to evaluate the prediction error of the model, and the samples in the validation dataset do not participate in the optimization of the model parameters.

[0012] Preferably, step (3) specifically includes: Input the training dataset into the NEP training program in GPUMD. Based on the element type of each atom and the relative position of the atoms in its neighborhood, generate a local environment descriptor containing radial and angular components. Establish the mapping relationship between the descriptor and the atomic energy through a neural network. Summate the atomic energies to obtain the system energy. Calculate the atomic force and stress tensor based on the system energy. Use the system energy, atomic force, and stress tensor obtained from first-principles calculations as labeled data to optimize the NEP model parameters. Calculate the prediction errors of the NEP model for system energy, atomic force, and stress tensor for different material compositions and defect configurations using the validation dataset. If the prediction error does not meet the preset accuracy requirements, supplement the first-principles calculation data for the material composition or defect local environment corresponding to the error, update the training dataset, and retrain. If the prediction error meets the preset accuracy requirements, use the obtained NEP model for the large supercell simulation in step (4). The preset requirements are: the root mean square error of the energy per atom predicted by the model relative to the first-principles calculation result shall not exceed 1 millielectronvolt / atom, and the root mean square error of the force on the atom in the x, y, and z directions shall not exceed 50 millielectronvolts per angstrom.

[0013] Preferably, step (4) specifically includes: Based on a hexagonal wurtzite unit cell containing 4 atoms, a GaN, AlN, or AlGaN supercell with 64,000 to 768,000 lattice sites is established by expanding the cell along three lattice directions by (20~40)×(20~40)×(40~120). For the AlGaN supercell, Al and Ga atoms are randomly arranged at the cation lattice sites according to the target Al composition. In the supercell, the types, numbers, or spatial positions of atoms are adjusted according to the predetermined atomic defect configuration to set the target atomic defects. Target supercells with different defect concentrations are established according to the predetermined number of defects. For each defective target supercell, the corresponding defect-free supercell before the introduction of atomic defects is used as a reference supercell. The atomic defects include one or more of Ga vacancies, Al vacancies, N vacancies, NN splitting gap defects, carbon-substituted nitrogen defects (C_N), oxygen-substituted nitrogen defects (O_N), and silicon-substituted cation defects (Si_Al or Si_Ga).

[0014] This invention utilizes examples to test the supercell size convergence, confirming that a large supercell with 64,000 lattice sites meets the size convergence requirements for thermal conductivity calculations. Therefore, this size is chosen for subsequent calculations to balance the characterization of long-wavelength phonon modes with computational costs. With sufficient computing power, even larger-scale cell expansion can be performed, extending the supercell size to millions or more lattice sites. In this case, the model achieves higher accuracy, but the computational requirements are also significantly higher.

[0015] Preferably, step (5) specifically includes: The validated NEP model was used for the target supercell and the reference supercell. Structural relaxation and temperature equilibrium at the target temperature were performed on each supercell. An external field was applied along the predetermined heat transport direction, and uniform nonequilibrium molecular dynamics simulations were performed under temperature control. The initial unsteady state interval was discarded, and the lattice thermal conductivity was calculated based on the heat flow response of the subsequent statistical interval. Five independent simulations were performed for each supercell using different initial velocities. Under the same temperature, heat transport direction, external field strength, and statistical interval, the arithmetic mean of the lattice thermal conductivity obtained from the five simulations was taken as the lattice thermal conductivity result of the supercell.

[0016] Preferably, step (6) specifically includes: For the analysis of the impact of atomic defects, the average lattice thermal conductivity of the target supercell with defects and its corresponding defect-free reference supercell were obtained through five independent simulations. The difference in thermal conductivity, normalized lattice thermal conductivity, and relative rate of change between the two were calculated, and the relationship between the above results and the defect type and defect concentration was output. For the analysis of the impact on AlGaN alloys, under the same temperature, heat transport direction, and simulation conditions, the average lattice thermal conductivity of defect-free randomly arranged supercells with different Al compositions were obtained through five independent simulations, and the results of the lattice thermal conductivity changing with Al composition were output.

[0017] This invention also discloses a machine learning-based atomic potential-based system for predicting the thermal conductivity of GaN, AlN, and AlGaN atomic defects, used to implement the aforementioned machine learning-based atomic potential-based method for predicting the thermal conductivity of GaN, AlN, and AlGaN atomic defects, comprising: The small supercell construction module is used to construct small supercells of GaN, AlN and AlGaN with different Al compositions. Al atoms and Ga atoms are randomly arranged at the cation lattice sites of the AlGaN small supercell according to the target Al composition, and atomic defects are set in the small supercell. The first-principles data module is used to obtain the cell parameters, element types, atomic coordinates, system energy, atomic forces and stress tensors of the atomic configuration of the small supercell, and to form training datasets and validation datasets. The atomic potential training and validation module is used to train the NEP model using the training dataset and to verify the prediction accuracy using the validation dataset. When the prediction accuracy does not meet the preset requirements, first-principles data of the corresponding components or defect configurations are added and the model is retrained. The large supercell construction module is used to build a target large supercell with defects and a corresponding defect-free reference large supercell, and to build defect-free AlGaN large supercells with different Al compositions and random arrangement of Al and Ga atoms. The thermal transport calculation module is used to perform temperature equilibrium and uniform nonequilibrium molecular dynamics simulations on large supercells using the validated NEP model to obtain lattice thermal conductivity. The results analysis module is used to compare the lattice thermal conductivity of the target supercell and the reference supercell, and to compare the lattice thermal conductivity of defect-free AlGaN supercells with different Al compositions. It outputs the defect type, defect concentration, and the corresponding thermal conductivity changes for the Al composition.

[0018] Preferably, the heat transport calculation module includes: Temperature balancing unit, used for structural relaxation of large supercells and temperature balancing at target temperature; The simulation unit is used to apply an external field along a predetermined heat transport direction, perform uniform non-equilibrium molecular dynamics simulations under temperature-controlled conditions, and calculate the lattice thermal conductivity based on the heat flow response in the stable statistical interval. The statistical unit is used to perform five independent simulations for each large supercell using different initial velocities and to calculate the arithmetic mean of the lattice thermal conductivity of the five simulations.

[0019] This invention trains neural evolution potentials using first-principles data containing different material compositions and atomic defect local environments, and verifies the prediction accuracy of the obtained neural evolution potentials under different material compositions and defect configurations; for material compositions or defect local environments with insufficient prediction accuracy, first-principles data are supplemented and retrained; the verified neural evolution potentials are used for large supercell molecular dynamics thermal transport simulations to obtain lattice thermal conductivity under different atomic defect types, defect concentrations and Al composition conditions.

[0020] The beneficial effects of this invention are as follows: (1) This invention establishes a first-principles dataset for GaN, AlN, AlGaN with different Al compositions and different atomic defect configurations, so that the trained NEP model can cover different material compositions and defect local atomic environments, providing an atomic potential basis for the unified simulation of multiple material compositions and defect systems.

[0021] (2) This invention verifies the prediction accuracy of the NEP model for system energy, atomic force and stress tensor under different material compositions and defect configurations. Based on the material composition or defect local environment where the prediction error is insufficient, the applicability and prediction reliability of the NEP model are improved by supplementing first-principles data and retraining.

[0022] (3) This invention utilizes the validated NEP model to conduct molecular dynamics thermal transport simulations on large supercells and long time scales. Compared with directly using first-principles methods to simulate atomic motion on the same scale, it can significantly reduce the computational resource and computation time requirements while maintaining the prediction accuracy constrained by first-principles data.

[0023] (4) The present invention establishes a corresponding defect-free reference supercell for each defective target supercell and performs calculations under the same temperature, heat transport direction, external field intensity and statistical interval, thereby improving the comparability of thermal conductivity results for different defective systems.

[0024] (5) The present invention sets five different initial velocities for each large supercell and performs five independent simulations. By calculating the arithmetic mean of multiple results, the influence of random fluctuations in a single molecular dynamics simulation on the lattice thermal conductivity results is reduced.

[0025] (6) This invention can output the defect type, defect concentration and the relationship between Al composition and lattice thermal conductivity, providing a calculation basis for the analysis of thermal transport mechanism and thermal management design of GaN, AlN and AlGaN materials. Attached Figure Description

[0026] Figure 1 A comparison chart of the system energy predicted by the NEP model and the energy calculated using first-principles calculations.

[0027] Figure 2 A comparison diagram showing the forces on atoms in the x-direction predicted by the NEP model and the results of first-principles calculations.

[0028] Figure 3 A comparison diagram showing the forces on atoms in the y-direction predicted by the NEP model and the results of first-principles calculations.

[0029] Figure 4 A comparison diagram of the forces on atoms in the z-direction predicted by the NEP model and the results of first-principles calculations.

[0030] Figure 5 For Al x Ga 1-x The result of N lattice parameter a varying with Al composition x is shown in the figure.

[0031] Figure 6 For Al x Ga 1-x The result of N lattice parameter c as a function of Al composition x is shown in the figure.

[0032] Figure 7 This is a comparison chart of NEP calculations and experimental results for the thermal conductivity of GaN lattice as a function of temperature.

[0033] Figure 8 This is a comparison graph showing the NEP calculation results and experimental results of the thermal conductivity of AlN lattice as a function of temperature.

[0034] Figure 9 Al at 300 K x Ga 1-xThe graph shows the result of the lattice thermal conductivity of the N random alloy varying with the Al composition x.

[0035] Figure 10 The graph shows the normalized lattice thermal conductivity of GaN as a function of defect concentration under different example atomic defect conditions at 300 K.

[0036] Figure 11 The graph shows the normalized lattice thermal conductivity of AlN as a function of defect concentration under different example atomic defect conditions at 300 K.

[0037] Figure 12 This is a schematic diagram of the defect-free local atomic configuration of a hexagonal wurtzite structure, using GaN as an example.

[0038] Figure 13 This is a schematic diagram illustrating the local atomic configuration of a Ga vacancy, using GaN as an example.

[0039] Figure 14 This is a schematic diagram illustrating the local atomic configuration of an N-vacancy, using GaN as an example.

[0040] Figure 15 This is a schematic diagram illustrating the local configuration of N split interstitial atoms, using GaN as an example.

[0041] Figure 16 This is a flowchart illustrating the method of the present invention.

[0042] Figure 17 The graph shows the normalized lattice thermal conductivity of GaN at 300 K as a function of defect concentration for different substitutional defects.

[0043] Figure 18 The graph shows the normalized lattice thermal conductivity of AlN at 300 K as a function of defect concentration for different substitutional defects.

[0044] Figure 19 This is a schematic diagram of the local atomic configuration of a silicon-substituted gallium defect (Si_Ga) using GaN as an example.

[0045] Figure 20 This is a schematic diagram of the local atomic configuration of a carbon-substituted nitrogen defect (C_N) using GaN as an example.

[0046] Figure 21 This is a schematic diagram of the local atomic configuration of oxygen-substituted nitrogen defects (O_N) using GaN as an example.

[0047] Figure 22 This is a schematic diagram of the (0001) face atomic configuration of a GaN supercell containing 768,000 atoms after structural relaxation using the NEP model, as shown in Embodiment 5 of the present invention. Detailed Implementation

[0048] The present invention is further illustrated below through embodiments, but these are not intended to limit the invention. The parameters used in the embodiments of the present invention are provided below. However, it should be understood that these are merely exemplary and are not intended to limit the invention. Adjustments made by those skilled in the art to material composition, atomic defect configuration, calculation parameters, and simulation conditions without departing from the technical concept of the present invention are all within the scope of protection of the present invention. Unless otherwise specified, the technical means used in the following embodiments are conventional means well known to those skilled in the art.

[0049] The atomic defects actually constructed and calculated in the following embodiments include Ga vacancies, Al vacancies, N vacancies, NN split-gap defects, carbon-substituted nitrogen defects (C_N), oxygen-substituted nitrogen defects (O_N), silicon-substituted gallium defects (Si_Ga), and silicon-substituted aluminum defects (Si_Al). Specifically, GaN systems are studied for Ga vacancies, N vacancies, NN split-gap defects, C_N, O_N, and Si_Ga; AlN systems are studied for Al vacancies, N vacancies, NN split-gap defects, C_N, O_N, and Si_Al. Corresponding defect-containing supercells are constructed for each type of defect, with a defect-free supercell of the same material used as a reference.

[0050] Shared model building and validation methods GaN, AlN, and Al with different Al compositions were constructed respectively. x Ga 1-xN-sized supercells, with defect-free supercells containing 32, 48, 64, 72, 80, 96, 108, 112, 120, 128, 144, 160, 168, 180, 192, 200, 216, 224, 240, 252, 256, and 512 atoms. Based on the number of cation sites in each supercell, different combinations of Al and Ga atom numbers are randomly selected, and their positions are randomly configured on the cation sites to construct various atomic configurations covering low, medium, and high Al compositions. The Al composition of each configuration before the introduction of defects is determined by the ratio of the number of Al atoms to the total number of cation sites. By combining supercells of different sizes and various combinations of atomic numbers and arrangements, a set of atomic configurations broadly covering the 0–1 Al composition range is formed, including GaN and AlN as endpoint components. Based on the above configurations, one or more of the following are randomly introduced according to the elements and corresponding lattice sites present in the material: Ga vacancies, Al vacancies, N vacancies, NN split interstitial defects, carbon-substituted nitrogen defects (C_N), oxygen-substituted nitrogen defects (O_N), and silicon-substituted cation defects (Si_Al or Si_Ga). The number and spatial distribution of these defects are varied to obtain small supercell atomic configurations with different material compositions, defect types, defect concentrations, and defect local environments. These configurations, along with the defect-free configuration, are used for subsequent first-principles calculations and model training. First-principles calculations are used to optimize the structure of each small supercell atomic configuration, with each configuration treated as an independent computational system. The structural optimization ends when the system energy and atomic forces meet the convergence conditions set in the input file, yielding relaxed atomic configurations under different material compositions and atomic defect conditions.

[0051] Using the optimized structures as initial configurations, 31 temperature points were set at 10 K intervals within the range of 200–500 K. First-principles molecular dynamics calculations were performed at each temperature point. The time step was set to 1 femtosecond, with each run consisting of 1000 steps, corresponding to a simulation duration of 1 picosecond, to collect the local atomic environment generated by thermal vibrations, local lattice distortions, and atomic motions near defects at different temperatures. Molecular dynamics trajectories were recorded one frame per step. Starting from step 20, one atomic configuration was extracted every 20 steps, i.e., sampling every 20 femtoseconds, with 50 atomic configurations extracted from each trajectory to reduce the similarity between adjacent samples. The above configurations, along with those selected during the structure optimization process, were combined to form the atomic configuration set for subsequent first-principles annotation. The above parameters were explicitly set by the user in the calculation input file.

[0052] For each selected atomic configuration, extract and record the following data: (1) Cell parameters; (2) The elemental type of each atom; (3) Spatial coordinates of each atom; (4) The total energy of the system corresponding to the atomic configuration; (5) The forces acting on each atom in three spatial directions; (6) Stress tensor corresponding to atomic configuration.

[0053] The unit cell parameters, atomic element types, and atomic coordinates of the same atomic configuration are used as input data, and the corresponding total system energy, atomic forces, and stress tensors are used as annotation data to form a one-to-one first-principles data sample.

[0054] A single atomic configuration, along with its corresponding cell parameters, element types, atomic coordinates, system energy, atomic forces, and stress tensor, is considered a complete data sample. After performing a repetition configuration check on the initial sample, 1000 samples are randomly selected to form the validation dataset, and the remaining samples form the training dataset. The same atomic configuration is not repeatedly assigned to either the training or validation datasets. First-principles samples added during subsequent training and validation processes are added to the training dataset after a repetition configuration check, while the validation dataset remains unchanged. A total of 32562 samples are ultimately obtained, of which 31562 samples form the training dataset for optimizing model parameters, and 1000 samples form the validation dataset for evaluating the model's prediction error. The validation samples are not used for model parameter optimization.

[0055] The Neuroevolution Potential (NEP) training program is invoked within the Graphics Processing Units Molecular Dynamics (GPUMD) program. For each atom in the training dataset, a local environment descriptor containing radial and angular components is generated based on the element type of the central atom and the element types and relative positions of its neighboring atoms within its cutoff radius. A mapping relationship between the local environment descriptor and atomic energy is established using a neural network.

[0056] The total energy of a system with atomic configuration is obtained by summing the atomic energies of each atom:

[0057] in, The total energy of the system, representing the atomic configuration. This represents the atomic energy of the i-th atom.

[0058] The predicted forces on each atom are obtained based on the negative gradient of the total energy of the system relative to the atomic coordinates:

[0059] in, This represents the predicted force on the i-th atom. This represents the spatial coordinates of the i-th atom.

[0060] During training, existing NEP training programs are used. Following the loss function calculation method provided by the program, a loss function is constructed based on the errors between the model's predicted system energy, atomic forces, and stress tensors and the first-principles labeled data. The model parameters are then iteratively optimized using the program's built-in neural evolution algorithm. The loss function calculation method and neural evolution algorithm utilize methods provided by existing tools.

[0061] The NEP model was tested using a validation dataset on GaN, AlN, and Al with different Al compositions. x Ga 1-x The prediction accuracy of N and different atomic defect local environments is assessed. The preset accuracy requirements for this embodiment are: the root mean square error of energy per atom does not exceed 1 millielectron volt / atom, and the root mean square error of the force on atoms in the x, y, and z directions does not exceed 35 millielectron volts / angstrom. When the prediction error on the validation dataset does not meet the preset accuracy requirements, more atomic configurations covering different material compositions, atomic arrangements, and defect local environments are added, and first-principles annotations are performed. The new samples are then added to the training dataset to expand the coverage of the training data. Subsequently, the NEP model is retrained, and the prediction accuracy is checked using the unchanged validation dataset. This process is repeated until the preset accuracy requirements are met.

[0062] Figure 1 This graph compares the system energy predicted by the NEP model with the energy calculated using first-principles calculations to verify the data. The data points are distributed near the diagonal line where both are equal, and the root mean square error of the energy prediction is 0.5 mEV / atom. Figure 2 , Figure 3 and Figure 4 The figures show a comparison between the predicted forces in the x, y, and z directions of atoms in the validation dataset and the first-principles calculations, with corresponding root mean square errors of 30.48, 30.49, and 30.06 millievolts / angtze Å, respectively. These results demonstrate that the NEP model has prediction accuracy close to that of first-principles calculations for the system energy and atomic forces of the atomic configurations in the validation dataset.

[0063] To further test the model's ability to describe crystal structures, the NEP model was used to analyze Al crystals with different Al compositions. x Ga 1- x The N configuration is used for structural relaxation, and the lattice parameters a and c are extracted. Figure 5 and Figure 6 The results for a and c as a function of Al composition x are presented. As the Al composition increases, the lattice parameters a and c predicted by the NEP model decrease approximately linearly, which conforms to Vigarde's law.

[0064] In terms of computational efficiency, taking a configuration containing 128 atoms as an example, with a time step of 1 femtosecond and a simulation time of 1 picosecond, the actual runtime of first-principles molecular dynamics calculations performed in parallel using 96 CPU cores was 41,278 seconds from task start to completion. Using a single RTX 3050 GPU, the same number of atoms and simulation time were completed in GPUMD using the NEP model, with an actual runtime of approximately 3 seconds. Compared to these two actual runtimes of 41,278 seconds and approximately 3 seconds, under the above test conditions, the time required for NEP molecular dynamics to complete this task was reduced by approximately 14,000 times. Since the system with 128 atoms is relatively small, it is difficult to fully utilize the parallel performance of the GPU. Therefore, calculations were further performed on a larger system: a system containing 64,000 atoms was simulated in GPUMD using the NEP model with a 1-nanosecond simulation, with a runtime of approximately 8 hours. Figure 1-6 The verification results show that the model can be used to calculate the lattice thermal conductivity on a timescale of tens of thousands of atoms and nanoseconds, while maintaining the reliability of energy, force and lattice parameter predictions.

[0065] Example 1: Variation of lattice thermal conductivity of GaN and AlN with temperature This embodiment uses the verified NEP model to predict the lattice thermal conductivity of defect-free GaN and AlN in the range of 200 K to 500 K, respectively. The specific process is as follows.

[0066] Step 1: Establish a thermal transport calculation model. Based on hexagonal wurtzite GaN and AlN unit cells containing 4 atoms each, expand the cells along three lattice directions at 20×20×40 to obtain defect-free large supercells containing 64,000 atoms each. Perform structural relaxation on the two large supercells to obtain the initial configurations for subsequent temperature-varying calculations.

[0067] Step 2: Set the target temperatures and bring the system to equilibrium. Seven target temperatures of 200 K, 250 K, 300 K, 350 K, 400 K, 450 K, and 500 K were set for the GaN and AlN supercells, respectively. At each target temperature, initial velocities were set for the atoms, and temperature equilibrium was achieved to bring the system to a stable state at the corresponding temperature. The same thermal transport calculation procedure was used at each temperature point to compare the change in lattice thermal conductivity with temperature. The above conditions are the calculation settings selected in this embodiment and are not fixed values ​​uniformly specified by the software.

[0068] Step 3: Calculate the lattice thermal conductivity at each temperature. For a large supercell that has reached temperature equilibrium, apply an external field along a predetermined heat transport direction and perform a uniform nonequilibrium molecular dynamics simulation under temperature-controlled conditions. Record the heat flow response along the stated direction during the simulation, discard the unsteady-state data from the initial stage, and calculate the lattice thermal conductivity of this simulation based on the heat flow response within the subsequent statistical interval.

[0069] Step 4: Perform independent repeated calculations. For the large supercells of GaN and AlN at each target temperature, five sets of initial atomic velocities are generated using random seeds 1, 2, 3, 4, and 5, respectively. For example, at 300 K, five independent simulations are initialized using “velocity 300 seed 1”, “velocity 300 seed 2”, “velocity 300 seed 3”, “velocity 300 seed 4”, and “velocity 300 seed 5”, respectively. The program automatically sets the system's total momentum and total angular momentum to zero, and then performs temperature equilibrium and thermal transport simulations. The time step for each simulation is set to 1 femtosecond, with a 200-picosecond temperature equilibrium first, followed by an application of an intensity of 5 × 10⁻⁶ along the z-direction corresponding to the

[0001] crystal orientation. -5 Angstrom -1 A uniform nonequilibrium molecular dynamics simulation was performed for 10 nanoseconds using a driving field. Thermal conductivity data was output every 1000 steps, discarding the data from the first 2 nanoseconds, and using 2–10 nanoseconds as the statistical interval. Five independent simulations were performed under the same driving field strength, thermal transport direction, and statistical interval. The lattice thermal conductivity of each simulation was calculated, and the arithmetic mean was taken as the lattice thermal conductivity of the corresponding material at the target temperature. Thus, the lattice thermal conductivity data of GaN and AlN at seven temperature points were obtained. In this embodiment, the target temperature, random seed number, time step, temperature equilibrium duration, driving field strength and direction, simulation duration, data output interval, and statistical interval were all explicitly set in the GPUMD input file according to the calculation requirements, and were not fixed parameters uniformly specified by the software. Velocity initialization, temperature equilibrium, and uniform nonequilibrium molecular dynamics calculations adopted existing methods in GPUMD.

[0070] Step 5: Compare the calculation results with the experimental results. Figure 7 The calculated results and experimental results of GaN lattice thermal conductivity as a function of temperature are presented. As the temperature increases from 200 K to 500 K, the calculated GaN lattice thermal conductivity gradually decreases, and its trend is consistent with the experimental results shown in the figure. Figure 8The corresponding results for AlN are presented. The calculated lattice thermal conductivity of AlN also decreases with increasing temperature, showing a similar trend to the experimental results illustrated. These results demonstrate that the validated NEP model can be used to calculate the lattice thermal conductivity of defect-free GaN and AlN at different temperatures, providing a model basis for subsequent thermal transport calculations of random AlGaN alloys and defective systems.

[0071] Example 2: Variation of lattice thermal conductivity of AlGaN random alloys with Al composition This embodiment uses the validated NEP model described above to calculate the Al content of different Al components at 300 K. x Ga 1-x The lattice thermal conductivity of N random alloys is determined through the following process.

[0072] Step 1: Constructing a large supercell for random alloys with different Al compositions. Based on a hexagonal wurtzite unit cell containing 4 atoms, the cell is expanded along three lattice directions at a 20×20×40 grid, resulting in a large supercell containing 64,000 lattice sites. Of these, 32,000 are cation lattice sites. x is set to 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0, respectively. For each composition, the number of Al and Ga atoms is determined according to the target Al composition and randomly distributed on the cation lattice sites. x of 0 and 1.0 correspond to defect-free GaN and AlN, respectively.

[0073] Step 2 involves structural relaxation of the large supercells with different compositions to obtain the equilibrium atomic configurations and unit cell parameters for each composition. The results of the lattice parameters a and c obtained from the NEP model as a function of Al composition are shown in [the table / examples]. Figure 5 and Figure 6 Its change conforms to Vigarde's law. The relaxed configuration is used for lattice thermal conductivity calculation.

[0074] Step 3: Set the target temperature of the supercell for each component to 300 K and perform temperature equilibration. After equilibration, apply an external field along the same heat transport direction as in Example 1, and conduct a uniform nonequilibrium molecular dynamics simulation under the same temperature control method, external field intensity, and heat flow statistics. Discard the unsteady-state data at the beginning of the simulation, and calculate the lattice thermal conductivity of each component based on the heat flow response in the subsequent statistical interval.

[0075] Step 4: For each Al component's large supercell, set five different initial velocities and perform five independent simulations. Take the arithmetic mean of the lattice thermal conductivity obtained from the five simulations as the lattice thermal conductivity of that component at 300 K. Summarize the calculation results according to the Al component from low to high to obtain the relationship between lattice thermal conductivity and composition.

[0076] Figure 9 Al at 300 K is given x Ga 1-x The results show the thermal conductivity of the AlGaN random alloy lattice as a function of Al composition x. The vertical axis of the thermal conductivity in the figure is logarithmic. As the Al composition increases from the GaN end, the lattice thermal conductivity initially decreases significantly, reaching its lowest value among the compositions shown in the figure when x is approximately 0.6; subsequently, it increases with further increases in Al composition, recovering to a relatively high level at the AlN end. Previous experimental studies have observed that the thermal conductivity of AlGaN alloys is lower than that of pure-phase materials, mainly attributed to alloy scattering caused by the mixing of Al and Ga atoms. The decreasing trend of thermal conductivity predicted in this embodiment is consistent with this experimental observation, indicating that the method of the present invention can predict the non-monotonic relationship between the thermal conductivity of random alloy lattice and Al composition.

[0077] This embodiment employs existing structural relaxation, temperature equilibrium, and homogeneous nonequilibrium molecular dynamics calculation methods in GPUMD, without designing a separate calculation algorithm. The parameters explicitly set in this embodiment include: a cell expansion factor of 20×20×40, a supercell size of 64,000 lattice sites, Al component values ​​of x = 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0, and a target temperature of 300 K. The number of Al atoms in each component is set to 32,000x, and the number of Ga atoms is set to 32,000(1-x), randomly distributed at cation sites. The initial velocity seed, simulation time step, equilibrium duration, driving field strength and direction, simulation duration, and statistical interval are explicitly specified in the input file in this embodiment, using the same settings as in Embodiment 1, and are not fixed values ​​uniformly specified by the software.

[0078] Example 3: The effect of vacant atoms and N-split interstitial atomic defects in GaN and AlN on lattice thermal conductivity In this embodiment, the validated NEP model is used to calculate the lattice thermal conductivity of GaN and AlN at 300 K for different atomic defect types and defect concentrations. The specific process is as follows.

[0079] Step 1: Constructing Defect-Free Reference Supercells. Based on the hexagonal wurtzite unit cells of GaN and AlN, respectively, the cells are expanded along three lattice directions at a 20×20×40 scale to obtain defect-free supercells with 64,000 lattice sites each. The supercells are then subjected to structural relaxation and used as defect-free reference supercells for their respective materials.

[0080] Step 2: Construct a large supercell containing the defective target. Taking GaN as an example, Figure 12 The local atomic configuration of a defect-free crystal is shown. Removing Ga atoms from the corresponding Ga lattice sites yields... Figure 13The Ga vacancy configuration shown; by removing N atoms at the corresponding N lattice sites, the following is obtained: Figure 14 The N vacancy configuration shown; by adjusting the number and position of local N atoms, the following can be obtained. Figure 15 The N-split interstitial atom configuration is shown. For AlN, construct the Al vacancy, N vacancy, and N-split interstitial atom configurations accordingly. Figure 12-15 The local configuration is shown using GaN as an example only, and the defect types described are merely examples in this embodiment.

[0081] Step 3: Set different defect concentrations. For each type of defect, target supercells with defect concentrations of 0.2%, 0.4%, 0.6%, 0.8%, and 1.0% are established, with the corresponding defect-free supercell at 0% defect concentration used as a reference. The defect concentration is determined according to the ratio of the set number of defects to the 64,000 lattice sites of the defect-free supercell. For example, a defect concentration of 0.2% corresponds to setting 128 of the aforementioned defects. For each target supercell containing defects, its reference supercell adopts a defect-free configuration of the same material as before the introduction of defects.

[0082] Step 4: Calculate and compare lattice thermal conductivity. The target supercell containing defects and its defect-free reference supercell are subjected to structural relaxation and temperature equilibrium at 300 K. Subsequently, uniform nonequilibrium molecular dynamics simulations are performed under the same heat transport direction, external field intensity, and statistical range. Each supercell is calculated five times independently using five different initial velocities, and the arithmetic mean of the resulting lattice thermal conductivity is taken. The average lattice thermal conductivity of the target supercell containing defects is denoted as κdef, and the average lattice thermal conductivity of the corresponding defect-free reference supercell is denoted as κref. The normalized lattice thermal conductivity is calculated according to κdef / κref. For the same material, all defect types and concentrations are compared against the defect-free results for that material.

[0083] Figure 10 The results show the variation of normalized lattice thermal conductivity in GaN with defect concentration. As defect concentration increases, the normalized lattice thermal conductivity corresponding to Ga vacancies, N vacancies, and N split interstitial atoms gradually decreases. Within the concentration range shown in the figure, at the same defect concentration, the thermal conductivity corresponding to N split interstitial atoms decreases the most, followed by N vacancies, while Ga vacancies show a relatively smaller decrease. When the defect concentration is 1.0%, the normalized lattice thermal conductivity of the three is approximately 0.41, 0.30, and 0.25, respectively.

[0084] Figure 11The corresponding results for AlN are given. As the defect concentration increases, the normalized lattice thermal conductivity corresponding to Al vacancies, N vacancies, and N split interstitial atoms also gradually decreases. Within the concentration range shown in the figure, the decrease is greatest for N split interstitial atoms, followed by N vacancies, with Al vacancies showing a relatively smaller decrease. When the defect concentration is 1.0%, the normalized lattice thermal conductivity of the three is approximately 0.29, 0.22, and 0.18, respectively.

[0085] Depend on Figure 10 and Figure 11 It is evident that in the GaN and AlN systems examined, both the type and concentration of atomic defects affect the lattice thermal conductivity. The method of this invention can predict the changes in thermal conductivity caused by different examples of atomic defects, using a corresponding defect-free system as a benchmark, and compare the responses of the two materials to different defects.

[0086] This embodiment employs existing structural relaxation, temperature equilibrium, and uniform nonequilibrium molecular dynamics calculation methods in GPUMD, without designing a separate algorithm. The conditions explicitly set in this embodiment include: a cell expansion factor of 20×20×40, a reference supercell size of 64,000 lattice sites, Ga vacancies, N vacancies, and NN splitting interstitial defects in GaN, and Al vacancies, N vacancies, and NN splitting interstitial defects in AlN. Defect concentrations are set to 0.2%, 0.4%, 0.6%, 0.8%, and 1.0%, uniformly using the 64,000 lattice sites of a defect-free reference supercell as the calculation benchmark, corresponding to 128, 256, 384, 512, and 640 defects respectively, with randomized defect locations. The calculation temperature is set to 300 K, and the initial velocity random seed, time step, equilibrium duration, driving field strength and direction, simulation duration, and statistical interval remain consistent with Example 1. The above parameters and defect construction conditions are explicitly set in this embodiment and are not fixed values ​​uniformly specified by the software.

[0087] Example 4: The effect of substitutional impurity defects in GaN and AlN on lattice thermal conductivity This embodiment uses the verified NEP model to calculate the effects of carbon-substituted nitrogen defects (C_N), oxygen-substituted nitrogen defects (O_N), and silicon-substituted cation defects (Si_Ga or Si_Al) on the lattice thermal conductivity of GaN and AlN at 300K, and compares the changes in thermal conductivity caused by different substitutional impurities at the same defect concentration. The specific process is as follows.

[0088] Step 1: Constructing a defect-free reference supercell. Based on the primitive cells of GaN and AlN with hexagonal wurtzite structures containing 4 atoms, respectively, the cells are expanded along three lattice directions at a 20×20×40 grid, resulting in supercells containing 64,000 lattice sites each, including 32,000 cation sites and 32,000 nitrogen sites. These defect-free supercells are used as reference supercells for GaN and AlN, respectively.

[0089] Step 2: Construct target supercells containing substitutional impurity defects. For GaN, randomly select non-repeating sites at nitrogen sites and replace the corresponding N atoms with C or O atoms to form C_N or O_N defects, respectively; randomly select non-repeating sites at Ga sites and replace the corresponding Ga atoms with Si atoms to form Si_Ga defects. For AlN, introduce C_N or O_N defects at nitrogen sites in the same manner, and replace the corresponding Al atoms at Al sites with Si atoms to form Si_Al defects. Construct target supercells containing each of the above single-type substitutional defects. By changing the number of substitutional atoms and the selected lattice sites, atomic configurations with different defect concentrations and random spatial distributions are obtained. After the substitutional operation, the total number of atoms in each target supercell remains 64,000.

[0090] Step 3: Set the defect concentration. In this embodiment, the defect concentration is defined as the ratio of the number of substitutional impurity atoms to the total number of lattice sites in a defect-free reference supercell. The concentrations are set to 0.2%, 0.4%, 0.6%, 0.8%, and 1.0%, corresponding to 128, 256, 384, 512, and 640 substitutional impurity atoms, respectively. A defect-free supercell with a 0% defect concentration is used as a reference. Under the same material and the same defect concentration, the same number of defects is used for each type of substitutional defect to compare the impact of defect type on lattice thermal conductivity.

[0091] Step 4: Calculate lattice thermal conductivity. Load the trained NEP potential function file into GPUMD, and use the NEP model to calculate the system energy and atomic forces. Perform structural relaxation on each defective target supercell and its corresponding defect-free reference supercell. Set the target temperature to 300 K, and use the same time step, temperature control method, heat transport direction, external field intensity, and heat flow statistics interval as in Example 1 to perform temperature equilibrium and uniform nonequilibrium molecular dynamics simulations. For each supercell, five sets of initial atomic velocities are generated using five different random seeds. Each set of initial velocities is used for one temperature equilibrium and subsequent heat transport simulation, resulting in five independently calculated lattice thermal conductivity values, which are then taken as their arithmetic mean. In this example, the cell expansion factor, defect type, defect concentration, random occupancy, target temperature, and number of independent simulations are all calculation conditions set in this example.

[0092] Step 5: Calculate the normalized lattice thermal conductivity and compare the results. The average lattice thermal conductivity of the defective target supercell is denoted as κdef, and the average lattice thermal conductivity of the corresponding defect-free reference supercell is denoted as κref. The normalized lattice thermal conductivity is calculated according to κdef / κref. For GaN and AlN, their respective defect-free results are used as the normalization benchmark, and the calculation results for different defect types and concentrations are summarized.

[0093] Figure 17 The variation of normalized lattice thermal conductivity in GaN with substitutional defect concentration is presented. With increasing defect concentration, the normalized lattice thermal conductivity for C_N, O_N, and Si_Ga gradually decreases. Within the concentration range shown in the figure, at the same defect concentration, Si_Ga exhibits the lowest normalized lattice thermal conductivity; the results for C_N and O_N are relatively close, with C_N's value being slightly lower than O_N's. When the defect concentration is 1.0%, the normalized lattice thermal conductivity for Si_Ga, C_N, and O_N are approximately 0.37, 0.54, and 0.56, respectively.

[0094] Figure 18 The corresponding results for AlN are presented. With increasing defect concentration, the normalized lattice thermal conductivity for C_N, O_N, and Si_Al gradually decreases. Within the concentration range shown in the figure, at the same defect concentration, the normalized lattice thermal conductivity for C_N and O_N is lower than that for Si_Al, and the results for C_N and O_N are quite similar. When the defect concentration is 1.0%, the normalized lattice thermal conductivity for Si_Al, C_N, and O_N is approximately 0.71, 0.55, and 0.56, respectively.

[0095] The above results indicate that within the concentration range examined in this embodiment, Si-Ga has a more significant inhibitory effect on the lattice thermal conductivity of GaN than C-N and O-N, while the inhibitory effects of C-N and O-N on AlN are more significant than those of Si-Al. The method of this invention can predict the changes in lattice thermal conductivity caused by different substitutional impurity defects based on a corresponding defect-free system, and compare their differences in influence in GaN and AlN.

[0096] The computational conditions in this embodiment were set according to the research object and comparison purpose, specifically including: a cell expansion factor of 20×20×40 and the corresponding 64,000 lattice sites; defect types of C_N, O_N, and Si_Ga or Si_Al; defect concentrations of 0.2%, 0.4%, 0.6%, 0.8%, and 1.0% and the corresponding 128, 256, 384, 512, and 640 substitutional impurity atoms; the spatial distribution of randomly selected substitutional sites on the corresponding sublattices; a target temperature of 300 K; and a repetitive calculation scheme using five different random seeds to generate initial velocities and performing five independent simulations. The parameters for heat transport direction, external field intensity, and heat flow statistics interval are the same as those set in Example 1. Structural relaxation, velocity initialization, and uniform nonequilibrium molecular dynamics simulations were performed using existing functions of GPUMD, and the interatomic interactions in the simulations were calculated using the NEP model trained and verified by the present invention.

[0097] Example 5: Structural relaxation of large-scale GaN supercells This embodiment uses the verified NEP model to perform structural relaxation on a GaN supercell containing 768,000 atoms, demonstrating the model's ability to handle large-scale atomic systems. The specific process is as follows.

[0098] Step 1: Construct a large-scale atomic model. Based on a GaN unit cell containing 4 atoms of hexagonal wurtzite structure, the cell is expanded along three lattice directions at a size of 40×40×120 to obtain a defect-free supercell containing 768,000 atoms, of which 384,000 are Ga atoms and 384,000 are N atoms. The number of atoms in this system is 12 times that of the aforementioned 64,000-atom supercell, and periodic boundary conditions are adopted along the three lattice directions.

[0099] Step 2: Load the NEP potential function. Organize the cell parameters, element types, and atomic coordinates of the large supercell into an atomic configuration file that GPUMD can read, and load the NEP potential function obtained during training into GPUMD. During structural relaxation, the total energy of the system is calculated using the NEP model, and the atomic forces are obtained from the negative gradient of the total energy with respect to each atomic coordinate, providing energy and force information for structural updates.

[0100] Step 3: Perform structural relaxation and check convergence. Utilize the existing energy minimization function in GPUMD to iteratively optimize the atomic configuration with the goal of reducing the total system energy. Set the force convergence threshold to 10. -2 Electron volts per angstrom (eV / Å) refers to the maximum absolute value of the force components in the x, y, and z directions for all atoms, not exceeding 0.01 eV / Å. Simultaneously, the absolute value of the total energy change of the system between two adjacent structural updates does not exceed 10. -4Electron volts, specifically 0.0001 electron volts, are used as an additional energy convergence check condition. The optimization results are checked to obtain the relaxed atomic configuration that simultaneously satisfies the above energy and force conditions.

[0101] Step 4: Record the time consumed by the structural relaxation calculation. Under the computing equipment and parallel resource configuration used in this embodiment, the structural relaxation of a GaN supercell containing 768,000 atoms was performed using the aforementioned NEP model. The calculation time was recorded when the absolute value of the total energy change of the system between two adjacent structural updates did not exceed 10. -4 Electron volts, and the maximum absolute value of the force components on all atoms does not exceed 10. -2 At electron volts per angstrom (eV / Å), the relaxed configuration satisfying the convergence condition is obtained. According to the actual operation record of this task, the calculation time to complete the above structural relaxation is 33 minutes. This result demonstrates the ability of the NEP model trained in this invention to handle systems with 768,000 atoms and complete large-scale structural relaxation.

[0102] The material type, cell expansion factor of 40×40×120, system scale of 768,000 atoms, and 10 in this embodiment are described. -4 Electron-volt energy change threshold and 10 -2 Electron volts / angstrom and the force threshold are calculation conditions set by the present embodiment. Structural relaxation uses existing functions of GPUMD, and interatomic interactions are calculated by the NEP model trained and verified by the present invention.

[0103] This embodiment demonstrates the ability to perform large-scale system energy, atomic force calculations, and structural optimization using NEP models trained with first-principles data from small supercells by performing structural relaxation on a 768,000-atom GaN supercell.

Claims

1. A method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN based on machine learning atomic potential, characterized in that, Includes the following steps: (1) Construct small supercells containing n atoms for GaN, AlN and AlGaN with different Al compositions respectively. Set random Al or Ga atom arrangements in the small supercells and introduce randomly distributed atomic defects to obtain atomic configurations for first-principles calculations. The n is a multiple of 4 and is not less than 32 and not greater than 512. (2) First-principles calculations were performed to optimize the small supercell atomic configurations obtained in step (1). The optimized structures were used as the initial configurations. First-principles molecular dynamics calculations were performed at different temperature points within the range of 200–500 K at fixed temperature intervals. Each run was 800–1200 steps, and an atomic configuration was selected every 10–30 steps. The configurations selected in the structure optimization process and molecular dynamics trajectory were summarized. The unit cell parameters, atomic element types and atomic coordinates of each configuration were extracted, and the corresponding system energy, atomic forces and stress tensors were used as annotation data. The atomic potential function is trained using the training dataset, and the prediction error of the model is evaluated using the validation dataset. When the prediction accuracy does not meet the preset requirements, the atomic configuration is supplemented and first-principles annotation is performed. The new samples are added to the training dataset, and the training and validation are repeated until the prediction accuracy meets the preset requirements. (3) The neural evolution potential algorithm is used in the molecular dynamics program of the graphics processor. Based on the element type of each atom in the training dataset and the relative position of the atoms in its neighborhood, a local environment descriptor containing radial and angular components is generated. The mapping relationship between the descriptor and the atomic energy is established. The model parameters are optimized using the system energy, atomic force and stress data to establish the NEP model. The prediction accuracy of the trained NEP model is verified using the validation dataset. If the preset requirements are not met, first-principles data of different material components or defect local environments are added and retrained. If the preset requirements are met, proceed to step (4). (4) Based on a wurtzite structure cell containing 4 atoms, expand the cell along the three lattice vector directions x, y, z to construct a defect-free supercell of GaN, AlN, or AlGaN. The cell expansion factor is adjusted according to the calculation requirements. The number of lattice sites in the defect-free supercell is greater than the number of atoms in the small supercell. When studying AlGaN alloys, Al atoms and Ga atoms are randomly configured at cation sites according to the target Al composition to construct an AlGaN random alloy model. Atomic defects are randomly introduced into the defect-free supercell. The number of defects is determined according to the predetermined defect concentration and the corresponding number of reference lattice sites to obtain the target supercell containing defects. The defect-free supercell before the introduction of defects is used as the reference supercell. (5) The validated NEP model is used for the target supercell and the reference supercell constructed in step (4), respectively; For each large supercell, the initial temperature is set to the target temperature T. The initial atomic velocities are generated using the GPUMD initial velocity setting command "velocity T seeds", where s is a random seed number. Five independent simulations are each assigned a different seed number. Under the same initial atomic configuration and temperature, by changing the random seed, the program generates five different atomic velocity configurations. Each configuration contains the velocity components of all atoms along the x, y, and z directions. During velocity initialization, the program automatically sets the total bus momentum and total angular momentum of the system to zero to eliminate overall translation and rotation. Then, temperature equilibrium is performed starting from each velocity configuration, followed by thermal transport simulations to reduce statistical fluctuations in a single simulation. Temperature equilibrium and uniform nonequilibrium molecular dynamics simulations are performed in GPUMD to obtain the lattice thermal conductivity calculated five times independently. The arithmetic mean of these five calculations is taken as the lattice thermal conductivity of the large supercell at the target temperature. (6) Under the same temperature and thermal transport calculation conditions, compare the lattice thermal conductivity of the target supercell with defects and the supercell without defects with the same material composition and initial Al and Ga atomic arrangement to obtain the influence of defect type and defect concentration on lattice thermal conductivity; under defect-free conditions, compare the lattice thermal conductivity of AlGaN supercells with different Al compositions and random Al and Ga atomic arrangement to obtain the variation law of lattice thermal conductivity under Al composition and random atomic arrangement conditions.

2. The method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN according to claim 1, characterized in that, Step (1) specifically includes: Construct GaN, AlN and Al with hexagonal wurtzite structure x Ga 1-x N small supercells, of which Al x Ga 1-x The N-type supercell is configured with Al and Ga atoms at cation sites according to the target Al component x; the number of atoms in the supercell is selected from 32, 48, 64, 72, 80, 96, 108, 112, 120, 128, 144, 160, 168, 180, 192, 200, 216, 224, 240, 252, 256, and 512; in the supercell, the types, numbers, or spatial positions of atoms are adjusted according to a predetermined atomic defect configuration to set atomic defects; multiple supercell atomic configurations with different local atomic environments are constructed for different material components, atomic defect types, defect positions, or defect concentrations; the atomic defects include one or more of Ga vacancies, Al vacancies, N vacancies, NN splitting gap defects, carbon-substituted nitrogen defects, oxygen-substituted nitrogen defects, and silicon-substituted cation defects.

3. The method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN according to claim 2, characterized in that, Step (2) is as follows: First-principles structure optimization was performed on all the small supercell atomic configurations constructed in step (1), with each atomic configuration as an independent computational system. Subsequently, with the optimized structure as the initial configuration, 31 temperature points were set in the range of 200-500 K at 10 K intervals. First-principles molecular dynamics calculations were performed at each temperature point with a time step of 1 femtosecond and 1000 steps per run, corresponding to a simulation duration of 1 picosecond. The trajectory was recorded frame by frame per step, and an atomic configuration was extracted every 20 frames, i.e., every 20 femtoseconds. 50 atomic configurations were extracted from each trajectory. The atomic configurations above and the atomic configurations selected during the structure optimization process were summarized to form a set of atomic configurations for subsequent first-principles annotation. For each selected atomic configuration, the cell parameters, element types and atomic coordinates of each atom, as well as the total energy of the system corresponding to the configuration, the forces on each atom and the stress tensor were recorded to form a one-to-one corresponding data sample. The data samples are divided into non-repeating training and validation datasets according to the complete atomic configuration, and then organized into extended XYZ format files that can be read by the NEP training program.

4. The method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN according to claim 3, characterized in that, In step (2), a single atomic configuration and its corresponding cell parameters, atomic element types, atomic coordinates, system energy, atomic forces, and stress tensors are used as a complete data sample. After checking the initial sample for repeated configurations, a portion of the samples are randomly selected to form the validation dataset, and the remaining samples form the training dataset. The same atomic configuration is not repeatedly assigned to the training dataset and the validation dataset. The first-principles samples added during subsequent training and validation processes are added to the training dataset after repeated configuration checks, while the validation dataset remains unchanged. The training dataset is used to optimize the model parameters. The validation dataset is used to evaluate the prediction error of the model, and the samples in the validation dataset do not participate in the optimization of model parameters.

5. The method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN according to claim 3, characterized in that, Step (3) is as follows: The training dataset is input into the NEP training program in GPUMD. Based on the element type of each atom and the relative positions of atoms in its neighborhood, a local environment descriptor containing radial and angular components is generated. The mapping relationship between the descriptor and the atomic energy is established through a neural network. The system energy is obtained by summing the energy of each atom, and the atomic force and stress tensor are calculated based on the system energy. The NEP model parameters were optimized using the system energy, atomic forces, and stress tensor calculated using first-principles calculations as labeled data. The prediction errors of the NEP model for system energy, atomic forces, and stress tensor for different material compositions and defect configurations were statistically analyzed using the validation dataset. If the prediction error does not meet the preset accuracy requirement, first-principles calculation data is added for the material composition or defect local environment corresponding to the error, the training dataset is updated and retrained; if the prediction error meets the preset accuracy requirement, the obtained NEP model is used for the large supercell simulation in step (4). The preset requirements are: the root mean square error of the energy per atom predicted by the model relative to the first-principles calculation result shall not exceed 1 millielectronvolt / atom, and the root mean square error of the force on the atom in the x, y, and z directions shall not exceed 50 millielectronvolts per angstrom.

6. The method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN according to claim 5, characterized in that, Step (4) is as follows: Based on a hexagonal wurtzite unit cell containing 4 atoms, a GaN, AlN, or AlGaN supercell with 64,000 to 768,000 lattice sites is established by expanding the cell along three lattice directions by (20~40)×(20~40)×(40~120). For the AlGaN supercell, Al atoms and Ga atoms are randomly arranged at the cation lattice sites according to the target Al composition. In the supercell, the types, numbers, or spatial positions of atoms are adjusted according to a predetermined atomic defect configuration to set the target atomic defects. Target supercells with different defect concentrations are established according to the predetermined number of defects. For each defective target supercell, the corresponding defect-free supercell before the introduction of atomic defects is used as the reference supercell. The atomic defects include one or more of Ga vacancies, Al vacancies, N vacancies, NN splitting gap defects, carbon-substituted nitrogen defects, oxygen-substituted nitrogen defects, and silicon-substituted cation defects.

7. The method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN according to claim 6, characterized in that, Step (5) is as follows: The validated NEP model was used for the target supercell and the reference supercell. Structural relaxation and temperature equilibrium at the target temperature were performed on each supercell. An external field was applied along the predetermined heat transport direction, and uniform nonequilibrium molecular dynamics simulations were performed under temperature control. The initial unsteady state interval was discarded, and the lattice thermal conductivity was calculated based on the heat flow response of the subsequent statistical interval. Five independent simulations were performed for each supercell using different initial velocities. Under the same temperature, heat transport direction, external field strength, and statistical interval, the arithmetic mean of the lattice thermal conductivity obtained from the five simulations was taken as the lattice thermal conductivity result of the supercell.

8. The method for predicting the thermal conductivity of atomic defects in GaN, AlN, and AlGaN according to claim 7, characterized in that, Step (6) specifically involves: For the analysis of the impact of atomic defects, the average lattice thermal conductivity of the target supercell with defects and its corresponding defect-free reference supercell were obtained through five independent simulations. The difference in thermal conductivity, normalized lattice thermal conductivity, and relative rate of change between the two were calculated, and the relationship between the above results and the defect type and defect concentration was output. For the analysis of the impact on AlGaN alloys, under the same temperature, heat transport direction, and simulation conditions, the average lattice thermal conductivity of defect-free randomly arranged supercells with different Al compositions were obtained through five independent simulations, and the results of the lattice thermal conductivity changing with Al composition were output.

9. A system for predicting the thermal conductivity of GaN, AlN, and AlGaN atomic defects based on machine learning atomic potential, used to implement the method for predicting the thermal conductivity of GaN, AlN, and AlGaN atomic defects based on machine learning atomic potential as described in any one of claims 1-8, characterized in that, include: The small supercell construction module is used to construct small supercells of GaN, AlN and AlGaN with different Al compositions. Al atoms and Ga atoms are randomly arranged at the cation lattice sites of the AlGaN small supercell according to the target Al composition, and atomic defects are set in the small supercell. The first-principles data module is used to obtain the cell parameters, element types, atomic coordinates, system energy, atomic forces and stress tensors of the atomic configuration of the small supercell, and to form training datasets and validation datasets. The atomic potential training and verification module is used to train the NEP model using the training dataset and to verify the prediction accuracy using the verification dataset. When the prediction accuracy does not meet the preset requirements, first-principles data of the corresponding component or defect configuration is added and the model is retrained. The large supercell construction module is used to build a target large supercell with defects and a corresponding defect-free reference large supercell, and to build defect-free AlGaN large supercells with different Al compositions and random arrangements of Al and Ga atoms. The thermal transport calculation module is used to perform temperature equilibrium and uniform nonequilibrium molecular dynamics simulations on large supercells using the validated NEP model to obtain lattice thermal conductivity. The results analysis module is used to compare the lattice thermal conductivity of the target supercell and the reference supercell, and to compare the lattice thermal conductivity of defect-free AlGaN supercells with different Al compositions. It outputs the defect type, defect concentration, and the corresponding thermal conductivity changes for the Al composition.

10. The GaN, AlN, and AlGaN atomic defect thermal conductivity prediction system according to claim 9, characterized in that, The heat transport calculation module includes: Temperature balancing unit, used for structural relaxation of large supercells and temperature balancing at target temperature; The simulation unit is used to apply an external field along a predetermined heat transport direction, perform uniform non-equilibrium molecular dynamics simulations under temperature-controlled conditions, and calculate the lattice thermal conductivity based on the heat flow response in the stable statistical interval. The statistical unit is used to perform five independent simulations for each large supercell using different initial velocities and to calculate the arithmetic mean of the lattice thermal conductivity of the five simulations.