A vibration spectrum simulation system using a domestic deep learning accelerator card

Abstract

The present invention discloses a vibration spectrum simulation system using a domestic deep learning accelerator card, and relates to the technical field of vibration spectrum simulation. The system includes: an initialization module for system modeling; an electronic structure calculation module for obtaining the electronic structure of the ground state through self-consistent field iteration; a dipole moment calculation module for calculating the dipole moment according to the converged electron density; a dynamic behavior evolution module for calculating the force and updating the speed and position of the atom according to the electronic structure of the ground state; an iterative control module for cyclically calling each module to obtain velocity sequence data and dipole moment sequence data; a spectrum calculation module for processing time series data to obtain power spectrum data and infrared spectrum data. By establishing an architecture and performing dynamic evolution, simulating vibration, generating spectra, and realizing automated spectrum simulation, the calculation time is reduced by computers and deep learning accelerator cards, which greatly reduces manpower and time costs.

Description

Technical Field

[0001] The present invention relates to the technical field of vibration spectrum simulation, and in particular to a vibration spectrum simulation system using a domestic deep learning accelerator card. Background Art

[0002] Each material and molecule has its corresponding physical and chemical properties, and vibrational spectroscopy can be used as the "identity card" of materials and molecules. Different materials and molecules have different characteristic frequencies and spectral shapes. Vibrational spectroscopy has many application scenarios, including identification and analysis of compounds, chemical reaction monitoring, characterization of chemical structure of materials, detection of drugs, atmospheric pollutants, etc. There are many experimental methods such as Fourier transform spectrometer (FTIR) and surface enhanced Raman scattering.

[0003] Existing methods for acquiring vibrational spectra require complex steps such as sample preparation, instrument calibration, background scanning, sample scanning, data acquisition, baseline correction, and quantitative analysis, which consumes a lot of manpower and time. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of consuming a large amount of manpower, material resources and time costs mentioned in the above background technology, and to propose a vibration spectrum simulation system using a domestic deep learning accelerator card.

[0005] The first aspect of the present invention provides a vibration spectrum simulation system using a domestic deep learning accelerator card, the system comprising:

[0006] An initialization module is used to model the atomic structure of a preset material or molecular system to obtain an initial model of the system; the initial model includes atomic positions, periodic boundary conditions, and initial electron density and electron wave function;

[0007] The electronic structure calculation module is used to use the self-consistent field iteration method to call the deep learning accelerator card to solve the electron density and potential energy function of the system converged in the current ion step to obtain the electronic structure of the ground state;

[0008] The dipole moment calculation module is used to calculate the dipole moment of the system at the current ion step based on the converged electron density;

[0009] The dynamic behavior evolution module is used to calculate the forces on atoms based on the electronic structure of the ground state and to update the speed and position of atoms according to the preset evolution algorithm.

[0010] An iterative control module, used for cyclically calling the electronic structure calculation module, the dipole moment calculation module and the dynamic behavior evolution module until a preset number of ion steps is reached, thereby obtaining velocity sequence data and dipole moment sequence data;

[0011] A spectrum calculation module is used to obtain power spectrum data and infrared spectrum data according to the velocity sequence data and the dipole moment sequence data;

[0012] The visualization module is used to draw a spectrum graph based on the power spectrum data and the infrared spectrum data.

[0013] Optionally, it is characterized in that the calculation process of the electronic structure calculation module includes:

[0014] Step 1: construct the Hamiltonian according to the current electron density;

[0015] Step 2: According to the preset exchange-correlation functional and diagonalization algorithm, the generalized eigenvalue solver, matrix multiplication and fast Fourier transform corresponding to the deep learning card are called to solve the eigenvalue equation of the Hamiltonian to obtain the eigenvalue electron wave function and eigenvalue energy;

[0016] Step 3: Update the electron density according to the intrinsic electron wave function;

[0017] Step 4, calculating the potential energy difference between the total potential energy of the current system and the total potential energy of the previous self-consistent process; if the potential energy difference is less than a preset first accuracy threshold, it is determined that the iteration reaches the potential energy convergence accuracy, and the process proceeds to step 5; otherwise, it returns to step 1;

[0018] Step 5: If the preset exchange-correlation functional is a hybrid functional, proceed to step 6; otherwise, exit the self-consistent field iteration process;

[0019] Step six, fix the converged intrinsic electron wave function and electron density in the inner loop, enter the iterative process of the precise exchange term, and in each iteration of the precise exchange term, calculate the exchange energy difference between the current precise exchange energy and the precise exchange energy of the previous iteration; if the exchange energy difference is less than the preset second precision threshold, it is judged that the iteration reaches the exchange energy convergence accuracy, and the self-consistent field iteration process is exited.

[0020] Optionally, it is characterized in that the calculation process of the dipole moment calculation module includes:

[0021] Read the nuclear charge data from the pseudopotential, and combine it with the electron density converged in the current ion step to calculate the dipole moment of the system in the current ion step:

[0022] ;

[0023] in, is the required dipole moment; is the electron charge; represents the nuclear charge; are atomic coordinates; is the expected value of the electron's position; represents the electron density; is the coordinate symbol.

[0024] Optionally, it is characterized in that the calculation process of the dynamic behavior evolution module includes:

[0025] According to the potential energy function of the system and the current atomic position, the potential energy gradient is solved by numerical method to obtain the force on each atom;

[0026] According to the preset first selection parameter, calling the corresponding evolutionary algorithm; the evolutionary algorithm is one of the VelocityVerlet algorithm and the Nosé-Hoover algorithm;

[0027] The current velocity, position and force information of the atoms are used as inputs of the evolutionary algorithm to calculate the velocity and position of each atom after a preset time step.

[0028] Optionally, it is characterized in that the evolutionary algorithm corresponding to the first selection parameter is a Velocity Verlet algorithm; the evolutionary algorithm includes:

[0029] Update the position of the target atom according to its current velocity, position and force:

[0030] ;

[0031] in, is the target atom at the new time step The target atom is any atom in the system; is the current position of the target atom; is the current velocity of the target atom; is the preset time step; is the force on the target atom; is the mass of the target atom;

[0032] According to the preset second selection parameter, calling a corresponding extrapolation algorithm; the extrapolation algorithm is one of linear, proaspc2 and proaspc3;

[0033] The historical force information of the atom is used as the input of the extrapolation algorithm to obtain the force of the target atom at the new time step;

[0034] Update the velocity of the target atom based on the current force and the force at the new time step:

[0035] ;

[0036] in, is the target atom at the new time step speed; is the target atom at the new time step The force.

[0037] Optionally, it is characterized in that the calculation process of the spectrum calculation module includes:

[0038] According to the dipole moment sequence, the dipole moment change rate is obtained:

[0039] ;

[0040] in, is the desired rate of change of dipole moment; is the time step The dipole moment of is the time step The dipole moment of

[0041] According to the preset autocorrelation function formula, the velocity sequence and the dipole moment change rate are processed to obtain the autocorrelation function of the velocity sequence and the autocorrelation function of the rate of change of the dipole moment ;

[0042] Perform Fourier transform on the autocorrelation function of the velocity series to obtain the power intensity:

[0043] ;

[0044] in, is the desired power intensity; is the atomic mass;

[0045] The autocorrelation function of the dipole moment change rate is Fourier transformed to obtain the infrared intensity:

[0046] ;

[0047] in, is the desired infrared intensity; represents the Boltzmann constant, represents the speed of light, represents the refractive index, Represents temperature.

[0048] Beneficial effects of the present invention:

[0049] The present invention proposes a vibration spectrum simulation system using a domestic deep learning accelerator card, which includes: an initialization module, which is used to model according to the atomic structure of a preset material or molecular system to obtain an initial model of the system; the initial model includes atomic positions, periodic boundary conditions, and initial electron density and electron wave function; an electronic structure calculation module, which is used to call the deep learning accelerator card to solve the electron density and potential energy function of the system at the current ion step through a self-consistent field iteration method to obtain the electronic structure of the ground state; a dipole moment calculation module, which is used to calculate the system at the current ion step according to the converged electron density. The dipole moment of the previous ion step; the dynamic behavior evolution module is used to calculate the atomic force according to the electronic structure of the ground state; and update the speed and position of the atom according to the preset evolution algorithm; the iterative control module is used to cyclically call the electronic structure calculation module, the dipole moment calculation module and the dynamic behavior evolution module until the preset number of ion steps is reached to obtain the velocity sequence data and the dipole moment sequence data; the spectrum calculation module is used to obtain the power spectrum data and the infrared spectrum data according to the velocity sequence data and the dipole moment sequence data; the visualization module is used to draw the spectrum according to the power spectrum data and the infrared spectrum data.

[0050] By establishing an architecture and performing dynamic evolution, simulating vibrations, generating spectra, and realizing automated spectral simulation, and reducing computing time through computer equipment and deep learning accelerator cards, the manpower and time costs are greatly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention will be further described below in conjunction with the accompanying drawings.

[0052] Figure 1 An architecture diagram of a vibration spectrum simulation system using a domestic deep learning accelerator card is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The embodiment of the present invention provides a vibration spectrum simulation system using a domestic deep learning accelerator card. Figure 1 , Figure 1 The architecture diagram of a vibration spectrum simulation system using a domestic deep learning accelerator card provided in an embodiment of the present invention. The system includes:

[0055] The initialization module is used to model according to the atomic structure of the preset material or molecular system to obtain the initial model of the system.

[0056] The electronic structure calculation module is used to call the deep learning accelerator card to solve the electron density and potential energy function of the system converged in the current ion step through the self-consistent field iteration method to obtain the electronic structure of the ground state.

[0057] The dipole moment calculation module is used to calculate the dipole moment of the system in the current ion step based on the converged electron density.

[0058] The dynamic behavior evolution module is used to calculate the forces on atoms based on the electronic structure of the ground state, and to update the speed and position of atoms according to the preset evolution algorithm.

[0059] The iterative control module is used to cyclically call the electronic structure calculation module, the dipole moment calculation module and the dynamic behavior evolution module until a preset number of ion steps is reached to obtain velocity sequence data and dipole moment sequence data.

[0060] The spectrum calculation module is used to obtain power spectrum data and infrared spectrum data according to the velocity sequence data and the dipole moment sequence data.

[0061] The visualization module is used to draw spectral spectra based on power spectrum data and infrared spectrum data.

[0062] The initial model includes atomic positions, periodic boundary conditions, initial electron density and electron wave function; the system is used to represent the material or molecular system.

[0063] A vibration spectrum simulation system using a domestic deep learning accelerator card is provided based on an embodiment of the present invention. By establishing an architecture and performing dynamic evolution, vibrations are simulated, spectra are generated, and automated spectral simulation is achieved. The calculation time is reduced by using computer equipment and a deep learning accelerator card, thereby greatly reducing manpower and time costs.

[0064] In one embodiment, the calculation process of the electronic structure calculation module includes:

[0065] Step 1: Construct the Hamiltonian based on the current electron density.

[0066] Step 2: According to the preset exchange-correlation functional and diagonalization algorithm, call the generalized eigenvalue solver, matrix multiplication and fast Fourier transform corresponding to the deep learning card to solve the eigenvalue equation of the Hamiltonian and obtain the eigenvalue electron wave function and eigenvalue energy.

[0067] Step three: Update the electron density based on the intrinsic electron wave function.

[0068] Step 4, calculate the potential energy difference between the total potential energy of the current system and the total potential energy of the previous self-consistent process; if the potential energy difference is less than the preset first accuracy threshold, it is determined that the iteration has reached the potential energy convergence accuracy and proceed to step 5; otherwise, return to step 1.

[0069] Step 5: If the preset exchange-correlation functional is a hybrid functional, proceed to step 6; otherwise, exit the self-consistent field iteration process.

[0070] Step six, fix the converged intrinsic electron wave function and electron density in the inner loop, enter the iterative process of the precise exchange term, and in each iteration of the precise exchange term, calculate the exchange energy difference between the current precise exchange energy and the precise exchange energy of the previous iteration; if the exchange energy difference is less than the preset second precision threshold, it is judged that the iteration has reached the exchange energy convergence accuracy, and the self-consistent field iteration process is exited.

[0071] In one implementation, the process significantly accelerates the solution of the Hamiltonian's eigenvalue equation by calling the generalized eigenvalue solver (Sygvd), matrix multiplication (hipBlas::GEMM) and fast Fourier transform (hipFFT) corresponding to the deep learning card (DCU). These acceleration techniques can significantly reduce the calculation time.

[0072] In one implementation, during the iteration process, original code is written for the calculation process to accelerate the DCU hardware, such as kernel_vtot, kernel_teter, kernel_get_psiphiRow, etc., which are used to calculate the Hamiltonian acting on the orbit. The specific implementation form is as follows:

[0073] __global__ void kernal_vtot( double* psi, double * gkk, int len) / *The kinetic energy in the total potential energy is calculated by the kinetic energy operator on the DCU, which can also be used for other wave function multiplications* /

[0074] {

[0075] int tid = blockIdx.x * blockDim.x + threadIdx.x;

[0076] if(tid <len)

[0077] {

[0078] psi[tid] = psi[tid]* gkk[tid];

[0079] }

[0080] }

[0081] __global__ void kernal_teter( cuDoubleComplex * psi, double * teter, int len) / *Add preconditions to implement wave function on DCU* /

[0082] {

[0083] int tid = blockIdx.x * blockDim.x + threadIdx.x;

[0084] if(tid <len)

[0085] {

[0086] psi[tid] = psi[tid]* teter[tid];

[0087] }

[0088] }

[0089] __global__ void kernel_get_psiphiRow(

[0090] int nocc_,

[0091] int ncband,

[0092] int nvband,

[0093] int ntotLocal,

[0094] cuDblNumMat cu_psiRow,

[0095] cuDblNumMat&cu_psiphiRow) / *Used to calculate the orbital pair product existing on the exchange term molecule on the DCU* /

[0096] {

[0097] int id = threadIdx.x + blockIdx.x * blockDim.x;

[0098] int k = nocc_ - nvband + id / nvband;

[0099] int j = nocc_ + id % ncband;

[0100] if (id <nvband * ncband ){

[0101] for (Int i = 0; i <ntotLocal; i++){

[0102] cu_psiphiRow(i, id) = cu_psiRow(i, k) * cu_psiRow(i, j);

[0103] } / / for

[0104] } / / if

[0105] }

[0106] In one implementation, the accuracy of the calculation is ensured by iteratively updating the electron density and checking the convergence of the total potential energy. When the potential energy difference is less than a preset first accuracy threshold, it is considered that the iteration has reached the potential energy convergence accuracy, which ensures the stability and accuracy of the calculation result.

[0107] In one implementation, the embodiment supports a variety of exchange-correlation functionals, including local density approximation, generalized gradient approximation, and hybrid functionals. This provides the possibility of considering more physical effects in the calculation, such as the exact exchange interaction between electrons, thereby improving the accuracy of the calculation results. The self-consistent field iteration and the exact exchange term iteration are decoupled to avoid unnecessary complex calculations under ordinary functionals, significantly reducing the amount of calculation. When using hybrid functionals, by fixing the converged intrinsic electron wave function and electron density in the inner loop and entering the iterative process of the exact exchange term, the calculation results can be further refined to ensure that the required accuracy can be achieved when high accuracy is required.

[0108] In one embodiment, the calculation process of the dipole moment calculation module includes:

[0109] Read the nuclear charge data from the pseudopotential, and combine it with the electron density converged in the current ion step to calculate the dipole moment of the system in the current ion step:

[0110] ;

[0111] in, is the required dipole moment; is the electron charge; represents the nuclear charge; are atomic coordinates; is the expected value of the electron's position; represents the electron density; is the coordinate symbol.

[0112] In one implementation, by combining the nuclear charge and the expected value of the electron position, the calculation process can fully reflect the charge distribution of the electrons and nuclei in the system, thereby accurately calculating the dipole moment. This helps to deeply understand the charge distribution characteristics and polarity of the system.

[0113] In one implementation, the electron density and nuclear charge data are directly used to avoid the complicated process of point-by-point summation or one-by-one estimation of the nuclear and electron positions. The dipole moment is directly calculated using the electron density and nuclear charge of the current iteration, simplifying the calculation steps. The calculation of the dipole moment mainly involves matrix operations (electronic part) and vector addition (nuclear part), which can be highly optimized by the parallel capability of the deep learning accelerator card to achieve efficient calculation.

[0114] In one embodiment, the calculation process of the dynamic behavior evolution module includes:

[0115] According to the potential energy function of the system and the current atomic position, the potential energy gradient is solved by numerical method to obtain the force on each atom.

[0116] According to the preset first selection parameter, the corresponding evolutionary algorithm is called.

[0117] The current velocity, position and force information of the atoms are used as input to the evolutionary algorithm to calculate the velocity and position of each atom after a preset time step.

[0118] Among them, the evolutionary algorithm is one of the Velocity Verlet algorithm and the Nosé-Hoover algorithm.

[0119] In one implementation, the force on each atom is obtained by using the potential energy function and the current atomic position to perform a gradient solution. ;in, It is The forces on each atom; It is the Born-Oppenheimer potential; It is for The process is based on the gradient calculation of the system potential energy, providing a high-precision evaluation of the interaction force between atoms. This method starts from the basic physical laws and ensures that the simulation results have good physical consistency.

[0120] In one implementation, the Velocity Verlet algorithm can efficiently simulate long-time scale dynamic behaviors without sacrificing too much accuracy. The Nosé-Hoover algorithm simplifies the complexity of temperature control and improves simulation efficiency by introducing an external heat bath to maintain the system temperature. This embodiment allows the user to set the first selection parameter to select the evolutionary algorithm during initialization. This flexibility allows the user to select the most appropriate algorithm based on the specific needs of the simulation and computing resources.

[0121] In one embodiment, the evolutionary algorithm corresponding to the first selection parameter is a Velocity Verlet algorithm; the specific process of the evolutionary algorithm includes:

[0122] Update the position of the target atom according to its current velocity, position and force:

[0123] ;

[0124] in, is the target atom at the new time step The target atom is any atom in the system; is the current position of the target atom; is the current velocity of the target atom; is the preset time step; is the force on the target atom; is the mass of the target atom.

[0125] According to the preset second selection parameter, the corresponding extrapolation algorithm is called.

[0126] The historical force information of the atom is used as the input of the extrapolation algorithm to obtain the force of the target atom in the new time step.

[0127] Update the velocity of the target atom based on the current force and the force at the new time step:

[0128] ;

[0129] in, is the target atom at the new time step speed; is the target atom at the new time step The force.

[0130] Among them, the extrapolation algorithm is one of linear, proaspc2 and proaspc3.

[0131] In one implementation, the number of potential energy calculations for solving atomic forces is reduced through an extrapolation algorithm, thereby optimizing the use of computing resources. This embodiment allows the user to set a second selection parameter during initialization to call different extrapolation algorithms during the process. This flexibility enables the algorithm to select the most appropriate extrapolation algorithm based on specific simulation requirements and computing resources.

[0132] In one implementation, the state of the new time step obtained by updating is the initial state of the next ion step.

[0133] In one embodiment, the calculation process of the spectrum calculation module includes:

[0134] According to the dipole moment sequence, the dipole moment change rate is obtained:

[0135] ;

[0136] in, is the desired rate of change of dipole moment; is the time step The dipole moment of is the time step The dipole moment of

[0137] According to the preset autocorrelation function formula, the velocity sequence and the dipole moment change rate are processed to obtain the autocorrelation function of the velocity sequence and the autocorrelation function of the rate of change of the dipole moment ;

[0138] Perform Fourier transform on the autocorrelation function of the velocity sequence to obtain the power intensity:

[0139] ;

[0140] in, is the desired power intensity; is the atomic mass;

[0141] The autocorrelation function of the dipole moment change rate is Fourier transformed to obtain the infrared intensity:

[0142] ;

[0143] in, is the desired infrared intensity; represents the Boltzmann constant, represents the speed of light, represents the refractive index, Represents temperature.

[0144] In one implementation, according to the Wiener-Khintchin theorem, the preset autocorrelation function can be defined as:

[0145] ;

[0146] The above formula calculates The Fourier transform of , and its square is calculated in the frequency domain, and then an inverse Fourier transform is used to calculate the given function The above formula gives the basic operation form. For discrete velocity series and the autocorrelation function of the dipole moment change rate, it can be calculated by discrete Fourier transform. There are multiple atoms in the system, and the total result is obtained by summing the calculation results of each atom.

[0147] In one implementation, based on the autocorrelation function of the atomic velocity sequence, the power intensity calculated can directly reflect the vibration mode and frequency characteristics of the system, providing a basis for analyzing the vibration characteristics of materials and molecules. The rate of change of the dipole moment is directly related to the intensity of the infrared spectrum, and the Fourier transform of its autocorrelation function can truly reproduce the infrared absorption characteristics of the system in the vibration mode.

[0148] In one implementation, for discrete data, the fast Fourier transform corresponding to the deep learning accelerator card can be used for processing to accelerate the calculation.

[0149] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A vibration spectrum simulation system using a domestic deep learning accelerator card, characterized in that: The system comprises: An initialization module is used to model the atomic structure of a preset material or molecular system to obtain an initial model of the system; the initial model includes atomic positions, periodic boundary conditions, and initial electron density and electron wave function; The electronic structure calculation module is used to use the self-consistent field iteration method to call the deep learning accelerator card to solve the electron density and potential energy function of the system converged in the current ion step to obtain the electronic structure of the ground state; The dipole moment calculation module is used to calculate the dipole moment of the system at the current ion step based on the converged electron density; The dynamic behavior evolution module is used to calculate the forces on atoms based on the electronic structure of the ground state and to update the speed and position of atoms according to the preset evolution algorithm. An iterative control module, used for cyclically calling the electronic structure calculation module, the dipole moment calculation module and the dynamic behavior evolution module until a preset number of ion steps is reached, thereby obtaining velocity sequence data and dipole moment sequence data; A spectrum calculation module is used to obtain power spectrum data and infrared spectrum data according to the velocity sequence data and the dipole moment sequence data; The visualization module is used to draw a spectrum graph based on the power spectrum data and the infrared spectrum data.

2. A vibration spectrum simulation system using a domestic deep learning accelerator card according to claim 1, characterized in that: The calculation process of the electronic structure calculation module includes: Step 1: construct the Hamiltonian according to the current electron density; Step 2: According to the preset exchange-correlation functional and diagonalization algorithm, the generalized eigenvalue solver, matrix multiplication and fast Fourier transform corresponding to the deep learning card are called to solve the eigenvalue equation of the Hamiltonian to obtain the eigenvalue electron wave function and eigenvalue energy; Step 3: Update the electron density according to the intrinsic electron wave function; Step 4, calculating the potential energy difference between the total potential energy of the current system and the total potential energy of the previous self-consistent process; if the potential energy difference is less than a preset first accuracy threshold, it is determined that the iteration reaches the potential energy convergence accuracy, and the process proceeds to step 5; otherwise, it returns to step 1; Step 5: If the preset exchange-correlation functional is a hybrid functional, proceed to step 6; otherwise, exit the self-consistent field iteration process; Step six, fix the converged intrinsic electron wave function and electron density in the inner loop, enter the iterative process of the precise exchange term, and in each iteration of the precise exchange term, calculate the exchange energy difference between the current precise exchange energy and the precise exchange energy of the previous iteration; if the exchange energy difference is less than the preset second precision threshold, it is judged that the iteration reaches the exchange energy convergence accuracy, and the self-consistent field iteration process is exited.

3. A vibration spectrum simulation system using a domestic deep learning accelerator card according to claim 1, characterized in that: The calculation process of the dipole moment calculation module includes: Read the nuclear charge data from the pseudopotential, and combine it with the electron density converged in the current ion step to calculate the dipole moment of the system in the current ion step: ； in, is the required dipole moment; is the electron charge; represents the nuclear charge; are atomic coordinates; is the expected value of the electron's position; represents the electron density; is the coordinate symbol.

4. A vibration spectrum simulation system using a domestic deep learning accelerator card according to claim 1, characterized in that: The calculation process of the dynamic behavior evolution module includes: According to the potential energy function of the system and the current atomic position, the potential energy gradient is solved by numerical method to obtain the force on each atom; According to the preset first selection parameter, calling the corresponding evolutionary algorithm; the evolutionary algorithm is one of the Velocity Verlet algorithm and the Nosé-Hoover algorithm; The current velocity, position and force information of the atoms are used as inputs of the evolutionary algorithm to calculate the velocity and position of each atom after a preset time step.

5. A vibration spectrum simulation system using a domestic deep learning accelerator card according to claim 4, characterized in that: The evolutionary algorithm corresponding to the first selection parameter is a Velocity Verlet algorithm; the evolutionary algorithm includes: Update the position of the target atom according to its current velocity, position and force: ； in, is the target atom at the new time step The target atom is any atom in the system; is the current position of the target atom; is the current velocity of the target atom; is the preset time step; is the force on the target atom; is the mass of the target atom; According to the preset second selection parameter, calling a corresponding extrapolation algorithm; the extrapolation algorithm is one of linear, proaspc2 and proaspc3; The historical force information of the atom is used as the input of the extrapolation algorithm to obtain the force of the target atom at the new time step; Update the velocity of the target atom based on the current force and the force at the new time step: ； in, is the target atom at the new time step speed; is the target atom at the new time step The force.

6. A vibration spectrum simulation system using a domestic deep learning accelerator card according to claim 1, characterized in that: The calculation process of the spectrum calculation module includes: According to the dipole moment sequence, the dipole moment change rate is obtained: ； in, is the desired rate of change of dipole moment; is the time step The dipole moment of is the time step The dipole moment of According to the preset autocorrelation function formula, the velocity sequence and the dipole moment change rate are processed to obtain the autocorrelation function of the velocity sequence and the autocorrelation function of the rate of change of the dipole moment ; Perform Fourier transform on the autocorrelation function of the velocity series to obtain the power intensity: ； in, is the desired power intensity; is the atomic mass; The autocorrelation function of the dipole moment change rate is Fourier transformed to obtain the infrared intensity: ； in, is the desired infrared intensity; represents the Boltzmann constant, represents the speed of light, represents the refractive index, Represents temperature.

Images