A method for developing a machine learning force field for semiconductor materials

Abstract

The present invention provides a method for developing a machine learning force field for semiconductor materials, including using VASP software to perform first principle molecular dynamics simulation of semiconductor materials, obtaining the potential energy and atomic force information of each frame conformation in MD; using an empirical force field to perform MD simulation containing PKA, obtaining an MD simulation trajectory containing PKA; using VASP software to perform single point energy calculation on the conformation in the trajectory, obtaining the potential energy and atomic force information of the MD simulation conformation containing PKA; using the conformation potential energy and atomic force information in the MD simulation and the MD simulation containing PKA as a database, and training to obtain a machine learning force field. The present invention uses an empirical force field to perform MD simulation containing PKA, and then uses the first principle to calculate the potential energy and atomic force of the conformation in the trajectory, and obtains a database close to the actual working conditions of the irradiation defect evolution, thereby realizing the establishment of a material machine learning force field that accurately characterizes the irradiation defect evolution process in the material.

Description

Technical Field

[0001] The present invention relates to the field of force field development technology, and in particular to a method for developing a machine learning force field for semiconductor materials. Background Art

[0002] When semiconductor electronic devices are in a space environment for a long time, the semiconductor materials in the devices will be damaged by radiation from charged particles in space. Defects caused by radiation damage are mainly generated by ionization damage and displacement damage. In order to clearly understand the generation and formation mechanism of displacement defects, it is necessary to study the whole process of primary knockout particle (PKA) cascade collision, high temperature annealing, and finally stabilization in the defect evolution. The time scale of the cascade process caused by PKA occurs at the femtosecond to picosecond level, which is difficult to observe experimentally. This time scale is within the range that can be handled by molecular dynamics simulation (MD), and molecular dynamics simulation can become an important technology for handling this process. The displacement effect of radiation in the material is usually the direct generation and cascade of PKA. In the early stage of PKA transport, its collision mean free path is large due to its high energy. With the development of the cascade collision process, the terminal PKA can collide with more lattice atoms in a short distance until its energy is insufficient to displace the lattice atoms. In addition, diffusion and annealing caused by thermal effects will also affect the defect structure. The entire PKA cascade process starts from the initial fierce collision, to the subsequent high temperature annealing, and finally tends to equilibrium. PKA cascade damage is still at a microscopic scale. Therefore, in order to accurately describe the microscopic dynamic evolution of the PKA cascade, including PKA cascade transport, system energy balance, and relaxation processes, molecular dynamics methods including defect identification can be used as a key technology to solve this process. However, the accuracy of molecular dynamics simulation depends on the accuracy of the force field of molecular simulation. Currently, there is no force field that can accurately describe the evolution of irradiation displacement defects in semiconductor silicon materials. Therefore, it is urgent to develop a high-precision force field for molecular simulation that can characterize the evolution of irradiation defects in semiconductor materials. Summary of the invention

[0003] In order to solve the problem of poor accuracy of the molecular dynamics force field model used in the simulation of radiation defect evolution of semiconductor materials under extreme environments, the present invention provides a machine learning force field development method required for the molecular dynamics simulation of radiation defect evolution of semiconductor devices under extreme environments.

[0004] A method for developing a semiconductor material machine learning force field of the present invention comprises:

[0005] Step S1: Use VASP software to perform first principle molecular dynamics simulation of semiconductor materials to obtain each frame conformation in the molecular dynamics simulation and the potential energy and atomic force information of the conformation;

[0006] Step S2: using an empirical force field to perform molecular dynamics simulation on a semiconductor material containing PKA to obtain a molecular dynamics simulation trajectory containing PKA;

[0007] Step S3: using the VASP software to perform single-point energy calculation on the conformation in the molecular dynamics simulation trajectory containing PKA, and obtaining the potential energy and atomic force information of the molecular dynamics simulation conformation containing PKA;

[0008] Step S4: The potential energy and atomic force information of each frame conformation in the molecular dynamics simulation obtained in step S1 and the potential energy and atomic force information of the molecular dynamics simulation conformation containing PKA obtained in step S3 are used as a database to obtain a machine learning force field for semiconductor materials through training.

[0009] Preferably, the use of VASP software to perform first principles molecular dynamics simulation of semiconductor materials includes: using the VASP software to perform the first principles molecular dynamics simulation of the semiconductor material at multiple temperatures and multiple pressures.

[0010] Preferably, the ensemble of the first-principle molecular dynamics simulation is an NPT ensemble, the time step is 0.5 fs, and the cutoff energy is 400 eV.

[0011] Preferably, the use of an empirical force field to perform molecular dynamics simulation on a semiconductor material containing PKA includes: setting an initial conformation, a time step and an ensemble, performing a molecular dynamics simulation in LAMMPS software, and obtaining a molecular dynamics simulation trajectory of the PKA-containing material from the LAMMPS software.

[0012] Preferably, the single-point energy calculation of the conformation in the molecular dynamics simulation trajectory containing PKA using VASP software includes: setting the functional file, iteration accuracy and K point, and performing single-point energy calculation, wherein the cutoff energy of the single-point energy calculation is 400eV, the K point is selected according to the size of the calculated model, and the electronic convergence accuracy is 1E -3 .

[0013] Preferably, the obtaining of the semiconductor material machine learning force field through training includes: taking the conformation as a sample, taking the information of the conformation as a label, and using the Adam method to perform neural network training to obtain the semiconductor material machine learning force field.

[0014] Preferably, the neural network training includes 3 hidden layers, each of which includes 240 nodes, the starting learning rate of the training is 0.005, and the ending learning rate is 1.0E -8 -1.0E -7 , the training batch size is generally 1,000,000.

[0015] Preferably, the empirical force field includes a force field whose potential functions are SW and Tersoff.

[0016] The advantages of the present invention compared to the prior art are:

[0017] The present invention uses an empirical force field to perform molecular dynamics simulations containing PKA, and then uses the first principles to calculate the potential energy of the conformation and the atomic forces in the trajectory, thereby obtaining a database close to the actual working conditions of irradiation defect evolution, thereby achieving the establishment of a semiconductor material machine learning force field that accurately characterizes the irradiation defect evolution process in semiconductor materials.

[0018] The present invention provides a theoretical model and necessary basis for in-depth research on the evolution of radiation defects in semiconductor materials, and has obvious advantages and broad application prospects in the design and application of new semiconductor materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of a method for developing a machine learning force field for semiconductor materials in an embodiment of the present invention;

[0020] Figure 2 Comparison of reference values ​​and predicted values ​​of potential energy and atomic forces on a training set and a test set of a machine learning force field for semiconductor materials according to an embodiment of the present invention;

[0021] Figure 3 is a lattice constant and binding energy curve of single crystal silicon in an embodiment of the present invention;

[0022] Figure 4 1 is a curve showing the change of Frenkel defects generated by PKA collision with single crystal silicon with an energy of 1000 eV over time in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] Machine learning methods are an effective means to develop high-precision force fields for molecular simulations. However, there is currently a lack of methods for developing high-precision machine learning force fields for molecular simulations that can characterize the evolution of irradiated defects in semiconductor materials. The present invention develops a machine learning force field for semiconductor materials that can achieve accurate molecular dynamics simulations, thereby providing more reliable predictions of the physical and chemical properties of semiconductor materials, laying the foundation for improving the performance of semiconductor electronic components.

[0024] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0025] See also Figure 1 As shown, a method for developing a machine learning force field for semiconductor materials according to an embodiment of the present invention includes:

[0026] Step S1: Use VASP software to perform first principle molecular dynamics simulation of semiconductor materials to obtain each frame conformation in the molecular dynamics simulation and the potential energy and atomic force information of the conformation;

[0027] Step S2: using an empirical force field to perform molecular dynamics simulation on a semiconductor material containing PKA to obtain a molecular dynamics simulation trajectory containing PKA;

[0028] Step S3: using VASP software to perform single-point energy calculation on the conformation in the molecular dynamics simulation trajectory containing PKA, and obtaining the potential energy and atomic force information of the molecular dynamics simulation conformation containing PKA;

[0029] Step S4: The potential energy and atomic force information of each frame conformation in the molecular dynamics simulation obtained in step S1 and the potential energy and atomic force information of the molecular dynamics simulation conformation containing PKA obtained in step S3 are used as a database to obtain a machine learning force field for semiconductor materials through training.

[0030] This embodiment uses an empirical force field to perform molecular dynamics simulations containing PKA, and then uses the first principle to calculate the potential energy of the conformation and the atomic forces in the trajectory, thereby obtaining a database close to the actual working conditions of the irradiation defect evolution, thereby achieving the establishment of a machine learning force field that accurately characterizes the irradiation defect evolution process in semiconductor materials. This embodiment provides a theoretical model and necessary foundation for in-depth research on the irradiation defect evolution of semiconductor materials, and has obvious advantages and broad application prospects in the design and application of new semiconductor materials.

[0031] According to the present invention, step S1 is first performed, using VASP software to perform first principle molecular dynamics simulations of semiconductor materials at multiple temperatures and multiple pressures. It should be noted that obtaining a data set from the conformational space composed of temperature and pressure is a conventional practice for developing a molecular dynamics force field model. In other implementations, other parameters such as density, material state, reaction coordinates, etc. may also be used.

[0032] Specifically, first set the parameters such as functional file, iteration accuracy, K point, initial conformation, time step and ensemble, then start the first principle molecular dynamics simulation, and finally obtain each frame conformation and its potential energy and atomic force information in the molecular dynamics simulation from the VASP software. Among them, the first principle molecular dynamics simulation used the NPT ensemble, the molecular dynamics simulation time step was 0.5fs, the cutoff energy of the first principle calculation was 400eV, the K point was selected according to the size of the calculated model, ensuring that the size of the K point and the three sides of the box was greater than 20 angstroms, and the electronic convergence accuracy was 1E -3 .

[0033] Then, step S2 is performed to perform a molecular dynamics simulation containing PKA using an empirical force field, wherein, when performing the molecular dynamics simulation, firstly, parameters such as the initial conformation, time step, and ensemble are set, and then the molecular dynamics simulation is started, and finally, the molecular dynamics simulation trajectory containing PKA is obtained from the LAMMPS software.

[0034] In step S3, after using the molecular dynamics simulation trajectory containing PKA obtained in step S2, VASP software is used to perform single-point energy calculations on the conformations in the trajectory. First, the functional file, iteration accuracy and K point are set, and then the single-point energy calculation is started to calculate the potential energy and atomic force information of the conformations in the molecular dynamics simulation trajectory containing PKA. Among them, the calculated cutoff energy is 400eV, and the K point is selected according to the size of the calculated model to ensure that the size of the K point and the three sides of the box is greater than 20 angstroms, and the electronic convergence accuracy is 1E -3 .

[0035] In step S4, the potential energy and atomic force information of the conformation in the first principle molecular dynamics simulation trajectory obtained in step S1 and the potential energy and atomic force information of the molecular dynamics simulation of PKA obtained in step S3 are used as a database, the conformation is used as a sample, the potential energy and atomic force information of the conformation is used as a label, and the Adam method is used to train the neural network to train the final machine learning force field. The neural network used generally has 3 hidden layers, each with 240 nodes, the training start learning rate is generally 0.005, and the end learning rate is generally 1.0E -8 and 1.0E -7 The training batch size is usually 1,000,000.

[0036] The final machine learning force field developed in this embodiment is used as the force field for molecular dynamics simulation of semiconductor materials. Calculations show that:

[0037] First, the semiconductor material machine learning force field developed by the present invention can well repeat the reference values ​​of potential energy and atomic force on the training set and the test set, such as Figure 2 As shown, Figure 2 (a) is the comparison between the reference value and the predicted value of potential energy on the training set. Figure 2 (b) is the comparison between the reference value and the predicted value of potential energy on the test set. Figure 2 (c) is the comparison between the reference value and the predicted value of the atomic force in the training set. Figure 2 (d) is the comparison between the reference value and the predicted value of the atomic force on the test set. Figure 2 (e) is the distribution of the error between the potential energy reference value and the predicted value on the training set and the test set, Figure 2 (f) Distribution of the error between the reference and predicted values ​​of atomic forces on the training and test sets.

[0038] Second, the lattice constant (a, in units) of single crystal silicon in semiconductor materials is obtained by using the machine learning force field of semiconductor materials developed in this embodiment, the existing empirical force field such as the force field with potential function SW and Tersoff, and the first principle calculation. (angstrom), ), binding energy (E c , in eV), elastic constant (C 11 , C 12 , C 44 , in Gpa), the results are shown in Table 1 and Figure 3 As shown. Among them, Figure 3 Here, DFT represents the first principle calculation result, ML represents the semiconductor material machine learning force field calculation result developed in this embodiment, and SW and tersoff represent the calculation results of the existing empirical force field.

[0039] Table 1

[0040]

[0041] It can be seen from Table 1 that the lattice constant, binding energy and elastic constant of single crystal silicon calculated by the machine learning force field of the semiconductor material in this embodiment are basically consistent with those calculated by the first principle.

[0042] Depend on Figure 3 It can be seen that the machine learning force field results (ML) of the semiconductor material developed in this embodiment and the first principle calculation results (DFT) are almost the same, which shows that the force field trained by this embodiment has an accuracy close to that of the first principle calculation.

[0043] It can be seen that the semiconductor material machine learning force field model developed by the present invention can accurately describe the static properties of semiconductor materials such as lattice constant, binding energy, elastic constant, etc., and is superior to existing empirical force fields.

[0044] Third, the machine learning force field for semiconductor materials developed by the present invention can describe the defect healing phenomenon during the evolution of irradiated defects in semiconductor materials, and the trend is similar to that of the existing empirical force field, such as Figure 4 As shown, Figure 4 This is a curve showing the change of Frenkel defects produced by PKA collision with single crystal silicon with energy of 1000eV over time.

[0045] Depend on Figure 4 It can be seen that the defect-time variation curve trend obtained using the semiconductor material machine learning force field developed in this embodiment is similar to that of SW and Tersoff. This shows that the force field developed in this embodiment can accurately describe the defect healing phenomenon during the evolution of semiconductor material irradiation defects.

[0046] Although the disclosure is disclosed as above, the protection scope of the disclosure is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the disclosure, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A method for developing a machine learning force field for semiconductor materials, characterized in that: include: Step S1: Use VASP software to perform first principle molecular dynamics simulation of semiconductor materials to obtain each frame conformation in the molecular dynamics simulation and the potential energy and atomic force information of the conformation; Step S2: using an empirical force field to perform molecular dynamics simulation on a semiconductor material containing PKA to obtain a molecular dynamics simulation trajectory containing PKA; Step S3: using the VASP software to perform single-point energy calculation on the conformation in the molecular dynamics simulation trajectory containing PKA, and obtaining the potential energy and atomic force information of the molecular dynamics simulation conformation containing PKA; Step S4: The potential energy and atomic force information of each frame conformation in the molecular dynamics simulation obtained in step S1 and the potential energy and atomic force information of the molecular dynamics simulation conformation containing PKA obtained in step S3 are used as a database to obtain a machine learning force field for semiconductor materials through training.

2. The method for developing a machine learning force field for semiconductor materials according to claim 1, characterized in that: The use of VASP software to perform first principles molecular dynamics simulation of semiconductor materials includes: using the VASP software to perform the first principles molecular dynamics simulation of the semiconductor material at multiple temperatures and multiple pressures.

3. The method for developing a machine learning force field for semiconductor materials according to claim 2, characterized in that: The ensemble of the first-principles molecular dynamics simulation is the NPT ensemble, with a time step of 0.5 fs and a cutoff energy of 400 eV.

4. The method for developing a machine learning force field for semiconductor materials according to claim 1, characterized in that: The molecular dynamics simulation of the semiconductor material containing PKA using the empirical force field includes: setting the initial conformation, time step and ensemble, performing molecular dynamics simulation in LAMMPS software, and obtaining the molecular dynamics simulation trajectory containing PKA from the LAMMPS software.

5. The method for developing a machine learning force field for semiconductor materials according to claim 1, characterized in that: The single-point energy calculation of the conformation in the molecular dynamics simulation trajectory containing PKA using VASP software includes: setting the functional file, iteration accuracy and K point, and performing single-point energy calculation, wherein the cutoff energy of the single-point energy calculation is 400eV, the K point is selected according to the size of the calculated model, and the electronic convergence accuracy is 1E -3 .

6. The method for developing a machine learning force field for semiconductor materials according to claim 1, characterized in that: The method of obtaining the semiconductor material machine learning force field through training includes: taking the conformation as a sample, taking the conformation information as a label, and using the Adam method to perform neural network training to obtain the semiconductor material machine learning force field.

7. The method for developing a machine learning force field for semiconductor materials according to claim 6, characterized in that: The neural network training includes 3 hidden layers, each of which includes 240 nodes. The starting learning rate of the training is 0.005 and the ending learning rate is 1.0E -8 -1.0E -7 , the training batch size is 1000000.

8. The method for developing a machine learning force field for semiconductor materials according to claim 1, characterized in that: The empirical force fields include force fields whose potential functions are SW and Tersoff.

Images