A machine learning force field development method

Abstract

The application provides a machine learning force field development method, comprising: using VASP software to perform first principle molecular dynamics simulation on semiconductor materials, obtaining MD simulation frame conformation information and potential energy and atomic force information thereof, obtaining multiple sets of machine learning force fields through training, performing molecular dynamics simulation on the calibration properties of the semiconductor materials, obtaining multiple simulation trajectories, calculating the information of each frame conformation in the simulation trajectory, comparing the conformation deviating from the average value to obtain a target conformation, performing single-point energy calculation on the target conformation, taking the frame conformation information and the target conformation information of the molecular dynamics simulation as a database, and training to obtain a final machine learning force field. The application effectively selects a target conformation close to the actual working condition by establishing multiple similar machine learning force field models and comparing the potential energy and atomic force of the semiconductor material molecular dynamics simulation conformation, so as to establish an accurate machine learning force field representing the semiconductor material.

Description

Technical Field

[0001] This invention relates to the field of simulation technology, and more specifically, to a method for developing machine learning force fields. Background Technology

[0002] Designing and developing superior semiconductor materials is central to improving the performance of semiconductor electronic components. Only by developing simulation models that can accurately describe semiconductor materials can we screen existing or potential semiconductor materials and provide theoretical predictions for the final design and development of superior semiconductor electronic components.

[0003] Currently, experimental methods can serve as the basis for theoretical calculations, but they cannot delve into the dynamic processes underlying the various physicochemical properties of semiconductor materials, nor can they reveal the relationship between these properties and the material's microstructure. Molecular dynamics (MD) simulations can handle timescales from femtoseconds to nanoseconds and system sizes from thousands to millions of atoms, making them a crucial technique for studying the physicochemical properties of semiconductor materials. However, the accuracy of MD simulations is highly dependent on the accuracy of the force field in the molecular simulation. Currently, there are very few force field models that can accurately describe semiconductor materials, especially for some novel semiconductor materials, where there are often no specific force field models to predict their physicochemical properties. Summary of the Invention

[0004] This invention addresses the problem of poor accuracy in molecular dynamics force field models currently used for semiconductor material simulation by providing a machine learning force field development method to accurately characterize molecular dynamics simulations of semiconductor materials.

[0005] The present invention provides a method for developing machine learning force fields, comprising:

[0006] Step S1: Use VASP software to perform first-principles molecular dynamics simulation of semiconductor materials to obtain each frame of the molecular dynamics simulation and the information of the conformation, including potential energy and atomic force information;

[0007] Step S2: Using the conformation and its information as a database, multiple sets of machine learning force fields are obtained through training;

[0008] Step S3: Using multiple sets of machine learning force fields, perform molecular dynamics simulations on the calibration properties of the semiconductor material to obtain multiple molecular dynamics simulation trajectories;

[0009] Step S4: Calculate the information of each frame conformation in the multiple molecular dynamics simulation trajectories, compare the information of each frame conformation under different machine learning force fields, and take the conformation whose information deviates from the average value as the target conformation;

[0010] Step S5: Use VASP software to perform single-point energy calculation on the target conformation to obtain information about the target conformation;

[0011] Step S6: Use the conformational information of each frame of the molecular dynamics simulation and the information of the target conformation as a database, and obtain the final machine learning force field through training.

[0012] Preferably, obtaining multiple sets of machine learning force fields through training includes: using the conformation as a sample, using the information of the conformation as a label, and using the Adam method to train a neural network to obtain multiple sets of machine learning force fields.

[0013] Preferably, the neural network training includes three hidden layers, each containing 240 nodes, with an initial learning rate of 0.005 and a final learning rate of 1.0E. -8 -1.0E -7 The training batch is typically 1,000,000.

[0014] Preferably, obtaining multiple sets of machine learning force fields through training further includes: using the same hyperparameters and different random numbers during training to obtain multiple different sets of said machine learning force fields.

[0015] Preferably, the first-principles molecular dynamics simulation of the semiconductor material using VASP software includes: performing first-principles molecular dynamics simulations of the semiconductor material at multiple temperatures and multiple pressures using the VASP software.

[0016] Preferably, the ensemble for the first principle molecular dynamics simulation is the NPT ensemble, with a time step of 0.5 fs and a cutoff energy of 400 eV.

[0017] Preferably, comparing the information of each frame's conformation under different machine learning force fields includes:

[0018] The average value of the information of the conformation in each frame under different machine learning force fields is obtained respectively;

[0019] Obtain the deviation of the information of the conformation in each frame from the average value under different machine learning force fields;

[0020] The deviation values ​​are compared, and the maximum value among the deviation values ​​is taken as the maximum deviation value between the information of the conformation in each frame under different machine learning force fields.

[0021] Preferably, the step of using the conformation whose information deviates from the average value as the target conformation includes:

[0022] The maximum deviation values ​​of each conformation are compared, and the conformation with the largest maximum deviation value is selected as the target conformation.

[0023] Preferably, the step of performing molecular dynamics simulations on the calibration properties of the semiconductor material includes: setting an initial conformation, time step, and ensemble; performing molecular dynamics simulations in LAMMPS software; and obtaining the molecular dynamics simulation trajectory from the LAMMPS software.

[0024] Preferably, the single-point energy calculation of the target conformation using VASP software includes: setting the functional file, iteration precision, and K-point, and performing the single-point energy calculation, wherein the cutoff energy of the single-point energy calculation is 400 eV, the K-point is selected according to the size of the calculated model, and the electronic convergence precision is 1 eV. -3 .

[0025] The advantages of this invention compared to existing technologies are:

[0026] This invention develops a high-precision molecular simulation force field that can characterize the physicochemical properties of semiconductor materials. It establishes multiple similar machine learning force field models and compares their potential energy and atomic forces in simulated molecular dynamics conformations of semiconductor materials. Conformations whose potential energy and atomic force information deviate from the average value are selected as target conformations. The final machine learning force field is then trained using these target conformations. This method effectively selects conformations that closely resemble actual working conditions, thereby establishing an accurate machine learning force field that characterizes semiconductor materials. This provides a theoretical model and necessary foundation for in-depth research on semiconductor materials, and has significant advantages and broad application prospects in the design and application of novel semiconductor materials. Attached Figure Description

[0027] Figure 1 This is a flowchart of the machine learning force field development method in an embodiment of the present invention;

[0028] Figure 2 The curves showing the changes in potential energy and atomic force error on the validation set with training batches during the final machine learning force field training process of this embodiment of the invention.

[0029] Figure 3 The binding energy of the Low-quartz phase of silicon dioxide under different lattice constants for different potential functions in embodiments of the present invention;

[0030] Figure 4 This represents the average radial distribution function of amorphous silica in the molecular dynamics simulation of the NVT ensemble at 300K in this embodiment of the invention. Detailed Implementation

[0031] Since the accuracy of the force field in molecular simulations directly affects the precision of molecular dynamics simulations, it is crucial to develop high-precision force fields for molecular simulations that can characterize the physicochemical properties of semiconductor materials.

[0032] Machine learning methods are an effective means of developing high-precision force fields for molecular simulations. Force fields for molecular dynamics simulations can be trained using first-principles calculations. The trained force fields achieve near-first-principles accuracy while handling molecular dynamics simulation-level models. Therefore, developing machine learning force fields for semiconductor materials can enable accurate molecular dynamics simulations, providing more reliable predictions of the physicochemical properties of semiconductor materials and laying the foundation for improving the performance of semiconductor electronic components.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Please see Figure 1 As shown, an embodiment of the present invention provides a method for developing machine learning force fields, comprising:

[0035] Step S1: Use VASP software to perform first-principles molecular dynamics simulation of semiconductor materials to obtain each frame of the molecular dynamics simulation and the information of the conformation, including potential energy and atomic force information;

[0036] Step S2: Using the conformation and its information as a database, multiple sets of machine learning force fields are obtained through training;

[0037] Step S3: Using multiple sets of machine learning force fields, perform molecular dynamics simulations on the calibration properties of the semiconductor material to obtain multiple molecular dynamics simulation trajectories;

[0038] Step S4: Calculate the information of each frame conformation in the multiple molecular dynamics simulation trajectories, compare the information of each frame conformation under different machine learning force fields, and take the conformation whose information deviates from the average value as the target conformation;

[0039] Step S5: Use VASP software to perform single-point energy calculation on the target conformation to obtain information about the target conformation;

[0040] Step S6: Use the conformational information of each frame of the molecular dynamics simulation and the information of the target conformation as a database, and obtain the final machine learning force field through training.

[0041] This embodiment establishes multiple similar machine learning force field models and compares their potential energy and atomic forces in the molecular dynamics simulation conformation of semiconductor materials. It effectively selects the target conformation that is close to the actual working conditions, thereby establishing a machine learning force field that accurately characterizes semiconductor materials. This provides a theoretical model and necessary foundation for in-depth research on semiconductor materials and has obvious advantages and broad application prospects in the design and application of new semiconductor materials.

[0042] According to this invention, step S1 is first performed, using VASP software to conduct first-principles molecular dynamics simulations of semiconductor materials at multiple temperatures and pressures. It should be noted that obtaining the dataset from the conformational space defined by temperature and pressure is a common practice in developing molecular dynamics force field models. In other embodiments, other parameters such as density, state of matter, and reaction coordinates can also be used.

[0043] Specifically, the parameters such as functional file, iteration precision, K-point, initial conformation, time step, and ensemble are first set. Then, a first-principles molecular dynamics simulation is started. Finally, the conformation, potential energy, and atomic force information of each frame in the molecular dynamics simulation are obtained from the VASP software. The first-principles molecular dynamics simulation uses the NPT ensemble, the time step is 0.5 fs, the cutoff energy calculated by the first principles is 400 eV, the K-point is selected based on the size of the calculated model, ensuring that the ratio of the K-point to the three sides of the box is greater than 20 Å, and the electronic convergence precision is 1E⁻³.

[0044] Then, in step S2, each frame of conformation, along with its potential energy and atomic force information obtained in step S1, is used as a database. The conformations are treated as samples, and their potential energy and atomic force information are used as labels. The Adam method is then used to train a neural network, training several sets of machine learning force fields. The neural network typically uses three hidden layers, each with 240 nodes. The initial learning rate is typically 0.005, and the final learning rate is typically around 1.0E. -8 and 1.0E -7 The training batch size is typically 1,000,000. Additionally, these force field training sets use the same hyperparameters but different random numbers.

[0045] In step S3, multiple sets of machine learning force fields trained in step S2 are used to perform molecular dynamics simulations of the calibration properties of semiconductor materials in the LAMMPS software. For example, if it is expected that the strength properties of semiconductor materials will be described by machine learning force fields, then molecular dynamics simulations of the strength properties are performed. During the molecular dynamics simulation, parameters such as the initial conformation, time step, and ensemble are first set, then the molecular dynamics simulation begins, and finally the molecular dynamics simulation trajectory is obtained from the LAMMPS software.

[0046] In step S4, the potential energy and atomic force information of each conformation in the molecular dynamics simulation trajectory are calculated using all the force fields trained in step S2. The potential energy and atomic force of different machine learning force fields in each conformation are compared, and the target conformation whose potential energy and atomic force deviate from the average value is selected.

[0047] Preferably, comparing the information of each frame's conformation under different machine learning force fields includes:

[0048] The average value of the information of each frame conformation under different machine learning force fields is obtained respectively;

[0049] Then, the deviation of each frame's configuration from the average value under different machine learning force fields is obtained;

[0050] Compare the various deviation values, and take the maximum value among the deviation values ​​as the maximum deviation value between the information of each frame configuration under different machine learning force fields.

[0051] For example, referring to Table 1, assuming that three sets of machine learning force fields are obtained through training in step S2, which are denoted as force field a, force field b, and force field c respectively for ease of description, then proceeding to step S3, for a certain property of the semiconductor material, such as strength, molecular dynamics simulations are performed using these three sets of force fields, and three sets of molecular dynamics simulation trajectories are obtained. Information (including potential energy and atomic force information) for each frame of these three trajectories is calculated. Assuming there are two frames of conformation, conformation 1 and conformation 2, the potential energy and atomic force information of conformation 1 under the three sets of force fields are calculated, abbreviated as information 1a, information 1b, and information 1c. Similarly, for conformation 2, the potential energy and atomic force information of conformation 2 under the three sets of force fields are calculated, yielding information 2a, information 2b, and information 2c.

[0052] Referring to Table 2, the maximum deviation values ​​among the conformational information in conformation 1 are compared. The method involves first obtaining the average value *m* of information 1a, information 1b, and information 1c, then calculating the deviation values ​​of information 1a, information 1b, and information 1c relative to the average value *m*, and obtaining the maximum value among these deviations, denoted as *1max*. This represents the maximum deviation value among the conformational information obtained by conformation 1 under the three force fields. Following this method, the maximum deviation value *2max* among the conformational information obtained by conformation 2 under the three force fields is calculated. Then, the maximum deviation values ​​of each conformation are compared, and the conformation with the largest maximum deviation value is selected as the target conformation.

[0053] Table 1

[0054]

[0055]

[0056] Table 2

[0057]

[0058] In step S5, single-point energy calculations are performed on the target conformation selected in step S4 using VASP software. First, the functional file, iteration precision, and K-point are set. Then, the single-point energy calculation begins, calculating the potential energy and atomic force information of the target conformation. The cutoff energy is 400 eV. The K-point is selected based on the size of the calculated model, ensuring that the ratio of the K-point to the dimensions of the three sides of the box is greater than 20 angstroms. The electron convergence precision is 1 E. -3 .

[0059] In step S6, the potential energy and atomic force information of each conformation in the first-principles molecular dynamics simulation trajectory obtained in step S1, and the potential energy and atomic force information of the target conformation obtained in step S5, are used as a database. The conformations are treated as samples, and their potential energy and atomic force information are used as labels. The Adam method is used to train a neural network to train the final machine learning force field. The neural network typically uses three hidden layers with 240 nodes per layer. The initial learning rate is typically 0.005, and the final learning rate is typically around 1.0E. -8 and 1.0E -7 Between these values, the training batch size is typically 1,000,000.

[0060] During the training process in step S6, the curves showing the change in error between potential energy and atomic force information on the validation set with the training batch are as follows: Figure 2 As shown, where, Figure 2 The x-axis represents the training batch, and the y-axis represents the mean square error of force and the mean square error of energy, respectively. Figure 2 It can be concluded that the potential energy and atomic force errors gradually decrease with the increase of training batches until they converge.

[0061] The final machine learning force field developed in this embodiment is used as the force field for molecular dynamics simulation of semiconductor device-related materials. The relevant properties of the materials are calculated and compared with existing empirical force fields (e.g., BKS, Tersoff). The calculations show that:

[0062] First, the machine learning force field model developed in this embodiment can accurately describe the static properties of semiconductor device-related materials, such as lattice constants, binding energy, and elastic constants, and is superior to existing empirical force fields. Second, the machine learning force field developed in this embodiment can also accurately describe the polycrystalline content of semiconductor materials, which is better than existing empirical force fields (such as BKS and Tersoff). Third, the machine learning force field developed in this embodiment can describe the radial distribution function (RDF), a more complex property that reflects the dynamics of the system.

[0063] Specifically, the relevant properties of low-quartz silica single crystals in the material were calculated using the machine learning force field, the force field with the potential function Vashishta, and the force field with the potential function Tersoff, respectively. The relevant properties of low-quartz silica single crystals in the material were also calculated through experiments and first-principles calculations, respectively. The results are shown in Table 3 and... Figure 3 As shown, Figure 3 In this context, DFT represents the first-principles calculation result, ML represents the machine learning force field calculation result developed in this embodiment, and BKS and tersoff represent the calculation results of existing empirical force fields.

[0064] Table 3

[0065]

[0066] From Table 3 and Figure 3 It can be seen that the lattice constant (α, in units of α) of the silica low-quartz phase single crystal obtained by machine learning force field calculation in this embodiment is... (angstrom), ), binding energy (E) c (unit: eV), elastic constant (C) 11 C 12 C 13 C 14 C 33 C 44 The value (in GPa) is basically consistent with the value calculated using first principles. Figure 3 The machine learning force field results (ML) developed in this embodiment almost overlap with the first-principles calculation results (DFT), which shows that the force field trained through this embodiment has an accuracy close to that of the first-principles calculation.

[0067] Using the machine learning force field developed in this embodiment, the polycrystalline amount of the Low-quartz phase of silicon dioxide in the material was calculated. The results are shown in Table 4. It can be seen that the machine learning force field developed in this embodiment can accurately describe the polycrystalline amount of semiconductor device-related materials and is superior to existing empirical force fields.

[0068] Table 4

[0069]

[0070] Using the machine learning force field developed in this embodiment, molecular dynamics simulations of amorphous silica in an NVT ensemble at 300K were performed. The average radial distribution function of silica is as follows: Figure 4As shown, it can be seen that the machine learning force field developed in this embodiment can describe the relatively complex property of the radial distribution function (RDF), which reflects the dynamics of the system, and the machine learning force field results (ML) developed in this embodiment almost overlap with the first-principles calculation results (DFT).

[0071] While the disclosure is as stated above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of this disclosure, and all such changes and modifications will fall within the protection scope of this invention.

Claims

1. A method for developing machine learning force fields, characterized in that, include: Step S1: Use VASP software to perform first-principles molecular dynamics simulation of semiconductor materials to obtain each frame of the molecular dynamics simulation and the information of the conformation, including potential energy and atomic force information; Step S2: Using the conformation and its information as a database, multiple sets of machine learning force fields are obtained through training; The process of obtaining multiple sets of machine learning force fields through training includes: using the conformation as a sample, using the information of the conformation as a label, and using the Adam method to train a neural network to obtain multiple sets of machine learning force fields. Step S3: Using multiple sets of machine learning force fields, perform molecular dynamics simulations on the calibration properties of the semiconductor material to obtain multiple molecular dynamics simulation trajectories; Step S4: Calculate the information of each frame conformation in the multiple molecular dynamics simulation trajectories, compare the information of each frame conformation under different machine learning force fields, and take the conformation whose information deviates from the average value as the target conformation; The comparison of information about the conformation in each frame under different machine learning force fields includes: The average value of the information of the conformation in each frame under different machine learning force fields is obtained respectively; Obtain the deviation of the information of the conformation in each frame from the average value under different machine learning force fields; The deviation values ​​are compared, and the maximum value among the deviation values ​​is taken as the maximum deviation value between the information of the conformation in each frame under different machine learning force fields; The step of using a conformation whose information deviates from the average value as the target conformation includes: The maximum deviation values ​​of each conformation are compared, and the conformation with the largest maximum deviation value is selected as the target conformation. Step S5: Use VASP software to perform single-point energy calculation on the target conformation to obtain information about the target conformation; Step S6: Use the conformational information of each frame of the molecular dynamics simulation and the information of the target conformation as a database, and obtain the final machine learning force field through training.

2. The machine learning force field development method according to claim 1, characterized in that, The neural network training includes three hidden layers, each containing 240 nodes. The initial learning rate is 0.005, and the final learning rate is 1.0E. 8 1.0E 7 The training batch size is 1,000,000.

3. The machine learning force field development method according to claim 1, characterized in that, Obtaining multiple sets of machine learning force fields through training also includes using the same hyperparameters and different random numbers during training to obtain multiple different sets of machine learning force fields.

4. The machine learning force field development method according to claim 1, characterized in that, The first-principles molecular dynamics simulation of semiconductor materials using VASP software includes: performing first-principles molecular dynamics simulations of the semiconductor materials at multiple temperatures and multiple pressures using the VASP software.

5. The machine learning force field development method according to claim 4, characterized in that, The ensemble used for the first principle molecular dynamics simulation is the NPT ensemble, with a time step of 0.5 fs and a cutoff energy of 400 eV.

6. The machine learning force field development method according to claim 1, characterized in that, The step of performing molecular dynamics simulations on the calibration properties of the semiconductor materials includes: setting an initial conformation, time step, and ensemble; performing molecular dynamics simulations in LAMMPS software; and obtaining the molecular dynamics simulation trajectory from the LAMMPS software.

7. The machine learning force field development method according to claim 1, characterized in that, The single-point energy calculation of the target conformation using VASP software includes: setting the functional file, iteration precision, and K-point, and performing the single-point energy calculation. The cutoff energy for the single-point energy calculation is 400 eV, the K-point is selected based on the size of the calculated model, and the electronic convergence precision is 1 eV. 3 .

Images