A method for improving the accuracy of a machine learning potential function for pyrolysis of energetic compounds

Abstract

The application discloses a method for improving the accuracy of a pyrolysis machine learning potential function of an energetic compound, comprising the following steps: obtaining energetic compound molecular crystal structure data and optimizing the data by using a first principle method; constructing a supercell structure, simulating the undecomposed state and the pyrolysis state of the supercell structure by using a first principle molecular dynamics method, and obtaining a molecular dynamics trajectory; taking the trajectory as a training set, and generating a plurality of machine learning potential functions by using DeepPot-SE; simulating the pyrolysis process by using classical molecular dynamics, and generating a test set; calculating the relative deviation of atomic force and the relative deviation of potential energy of the supercell structure in the test set; comparing the proportion of the number of two deviation indexes simultaneously lower than the lower limit with a threshold value, and determining whether to use the first principle method to recalculate part of the samples and update the training set; if the proportion is greater than the set threshold value, continuing to calculate the molecular dynamics trajectory until the cumulative time length reaches a preset time length, and obtaining a final high-precision pyrolysis machine learning potential function.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of machine learning and computational energetic materials research, and relates to a method for improving the accuracy of the pyrolysis machine learning potential function of energetic compounds. Background Technology

[0002] In recent years, high-energy-density energetic materials have been a key research focus in the international academic community. Scholars both at home and abroad are dedicated to finding new energetic materials with higher energy, better safety, and greater application potential. With the rapid development of computational energetic materials technology, research on the relationship between the molecular structure and energetic properties of novel energetic compounds using quantum chemical methods has matured. Among these methods, molecular dynamics simulation is a crucial method for studying the energy release laws and ignition propagation safety of energetic compounds. It serves as an important basis for exploring the application potential of new energetic materials and has significant reference value in the molecular design and laboratory research stages of energetic compounds.

[0003] Molecular dynamics methods commonly used in studying energetic compounds can be divided into first-principles molecular dynamics (or ab initio molecular dynamics, AIMD) and classical molecular dynamics based on empirical force fields. First-principles molecular dynamics requires solving the Kohn-Sham equations to obtain the potential energy experienced by ions during their motion, resulting in high computational complexity; the number of atoms in the simulation system is typically no more than 500. Due to the limited size of the computational system, first-principles molecular dynamics cannot adequately describe the reaction mechanisms of energetic compounds, such as combustion-to-detonation and ignition propagation. In contrast, classical molecular dynamics based on empirical force fields maps the force field to the potential energy experienced by ions during their motion, resulting in a much lower computational complexity than first-principles molecular dynamics; the number of atoms in the simulation system can exceed 1000. 5Empirical force fields that can describe the reactions of energetic compounds are usually called reaction force fields. ReaxFF is a common reaction force field. ReaxFF was initially used to study the chemical reactions of hydrocarbon compound systems and has now been extended to systems such as metal oxides and energetic materials. ReaxFF-lg corrects long-range interaction terms, improving the description of intermolecular interactions. Therefore, it improves the description of lattice parameters of RDX, terane, triaminotrinitrobenzene, and nitromethane, while not changing the description of reactions by ReaxFF. It is now the main force field used to study the thermal decomposition of explosives. However, while ReaxFF reaction force fields have good versatility, they are not suitable for novel energetic compounds and have large systematic errors in describing the potential energy of simulated systems. Machine learning potential functions can well make up for the shortcomings of ReaxFF and establish force fields with better accuracy based on the first-principles molecular dynamics trajectories of energetic compounds (The Journal of Physical Chemistry Letters 2023 14 (32), 7141-7148). The literature reports a machine learning potential function for the impact decomposition of hexanitrohexaazaisowurtzite. This potential function, combined with molecular dynamics methods, has both the computational accuracy of first-principles molecular dynamics methods and can achieve a faster speed than the ReaxFF empirical force field, thus effectively solving the dilemma of efficiency and accuracy in the impact detonation simulation of hexanitrohexaazaisowurtzite.

[0004] However, the existing technology has the following four technical defects: (1) The ReaxFF force field is not applicable to new energetic compounds and has a large deviation in describing the potential energy of the system; (2) The training set sample size of the existing impact decomposition machine learning potential function training is insufficient; (3) The existing impact decomposition machine learning potential function training focuses on the accuracy of the potential function in describing the forces on the atoms in the system, and has a deviation in describing the potential energy of the system; (4) The existing impact decomposition machine learning potential function training introduces the dynamic trajectory of impact decomposition, which is easy to interfere with the pyrolysis machine learning potential function training.

[0005] The above four technical defects affect the accuracy of the machine learning potential function for the pyrolysis of energetic compounds, making the classical molecular dynamics based on empirical force fields unreliable for simulating the pyrolysis process of novel energetic compounds. This makes it difficult to accurately reflect the energy release and ignition growth process of novel energetic compounds, and thus cannot support the molecular design and laboratory research of novel energetic compounds. Summary of the Invention

[0006] This invention provides a method to improve the accuracy of the pyrolysis machine learning potential function of energetic compounds, in order to solve the technical problems of poor applicability of the ReaxFF force field and the inability of existing impact decomposition machine learning potential functions to reliably describe the pyrolysis process. This will improve the scientific nature of molecular design and develop a new method for applying machine learning combined with molecular dynamics to the calculation of energetic materials.

[0007] This invention provides a method for improving the accuracy of the pyrolysis machine learning potential function of energetic compounds, comprising:

[0008] Step 1: Obtain the molecular crystal structure data of the energetic compound molecular crystal, and optimize the molecular crystal structure data using the first-principles method according to the predefined parameters;

[0009] Step 2: Based on the optimized molecular crystal structure data, establish a supercell structure with a total number of atoms not exceeding 400, and use first-principles molecular dynamics to simulate the undecomposed and pyrolysis states of the energetic compound to obtain the molecular dynamic trajectory;

[0010] Step 3: Using molecular dynamics trajectories as the training set, DeepPot-SE is used to train and generate n machine learning potential functions;

[0011] Step 4: Randomly select one from n machine learning potential functions, use classical molecular dynamics methods to simulate the pyrolysis process of energetic compounds at a specific time and temperature, obtain molecular dynamic trajectories, and use the supercell structure in the molecular dynamic trajectory as the test set;

[0012] Step 5: Substitute the test set into the n machine learning potential functions to calculate the supercell potential energy and the forces on all atoms in the supercell;

[0013] Step 6: Calculate the relative force deviation and the relative potential energy deviation of all atoms in the supercell in the test set, and set the upper and lower limits of the relative force deviation and the relative potential energy deviation of the atoms; calculate the proportion of supercell structures whose relative force deviation and relative potential energy deviation of the atoms are both less than the corresponding lower limits.

[0014] Step 7: If the proportion of the number is less than or equal to the set threshold, use the first-principles method to calculate the potential energy and the force on all atoms of the supercell structure that meets the requirements of the relative deviation of atomic force and the relative deviation of supercell potential energy, and put them into the training set. Retrain the DeepPot-SE model to obtain n machine learning potential functions, and return to step 5; if the proportion of the number is greater than the set threshold, proceed to step 8.

[0015] Step 8: Return to Step 4 and continue simulating the pyrolysis process at a specific time and temperature based on the existing molecular dynamics trajectory to obtain the continued molecular dynamics trajectory. Use the supercell structure in the continued molecular dynamics trajectory as the test set. Continue until the sum of the specific time reaches the preset duration to obtain the final machine learning potential function of the energetic compound pyrolysis.

[0016] The present invention provides a method for improving the accuracy of the pyrolysis machine learning potential function of energetic compounds, which has the following beneficial effects:

[0017] (1) The method of the present invention can be used to train the potential function of the pyrolysis machine learning of novel energetic compounds, and solves the problem that the ReaxFF force field is not applicable to novel energetic compounds and has a large deviation in describing the potential energy of the system.

[0018] (2) The method of the present invention adjusts the sample size of the training set of the pyrolysis machine learning potential function of the novel energetic compound based on the structure of the novel energetic compound (i.e., the number of carbon, hydrogen and oxygen atoms). This solves the problem of insufficient training set sample size for the existing impact decomposition machine learning potential function training, especially when the number of oxygen atoms is small, the carbon-nitrogen polymer product structure is complex, and the insufficient sample size of carbon-nitrogen polymer structure in the training set leads to a decrease in the accuracy of the pyrolysis machine learning potential function.

[0019] (3) The method of the present invention requires the generation of at least 6 potential functions in each training session. Each training session adopts the dual standard of relative deviation of atomic force and relative deviation of supercell potential energy to constrain the accuracy of the pyrolysis machine learning potential function, and balances the accuracy of the pyrolysis machine learning potential function in describing the atomic force and system potential energy in the system.

[0020] (4) The training set of the method of the present invention includes the kinetic trajectories of the non-decomposition state and the pyrolysis state of the novel energetic compound, so as to avoid introducing other decomposition states into the training set to interfere with the accuracy of the pyrolysis machine learning potential function. Attached Figure Description

[0021] Figure 1 The following are schematic diagrams of the structures of the energetic compounds in the examples: (a) is a schematic diagram of the structure of NH4N5 (pentazolium ammonium), (b) is a schematic diagram of the structure of (NH3OH)N5 (pentazolyl hydroxylamine), and (c) is a schematic diagram of the structure of (C(NH2)3)N5 (pentazolidineguanidine).

[0022] Figure 2 This is a comparison chart of the total energy drift of the pyrolysis machine learning potential function in the adiabatic (NVE) simulation process in the example;

[0023] Figure 3 The diagrams show the initial and final pyrolysis structures of NH4N5 (pentazonium) based on the pyrolysis machine learning potential function simulation. (a) is the initial pyrolysis structure diagram, and (b) is the final pyrolysis structure diagram. Detailed Implementation

[0024] The present invention provides a method for improving the accuracy of the pyrolysis machine learning potential function of energetic compounds, comprising:

[0025] Step 1: Obtain the molecular crystal structure data of the energetic compound molecular crystal, and optimize the molecular crystal structure data using the first-principles method according to the predefined parameters.

[0026] The molecular crystal structure data includes: cell length, cell orientation, and the original fractional coordinates of all atoms within the molecular crystal. Predefined parameters include: plane wave cutoff energy, Brillouin zone number of points, electron energy convergence criterion, and atomic force convergence criterion. The optimized molecular crystal structure data includes: optimized cell length, optimized cell orientation, and optimized fractional coordinates of all atoms within the molecular crystal.

[0027] In practice, the predefined parameter values ​​are: plane wave cutoff energy of 500~850 eV, and electron energy convergence criterion of 1.0×10⁻⁶ eV. -6 ~1.0×10 -4 eV / atom, the atomic force convergence standard is 0.01~0.05 eV / Å, the Brillouin zone lattice point number expression is L×M×N, where L, M and N are positive integers; let the unit cell lengths of the molecular crystal be a, b and c, then 30 ≤ a×L ≈ b×M ≈ c×N ≤ 70.

[0028] Step 2: Based on the optimized molecular crystal structure data, establish a supercell structure with a total number of atoms not exceeding 400, and use first-principles molecular dynamics to simulate the undecomposed and pyrolytic states of the energetic compound to obtain the molecular dynamic trajectory.

[0029] The supercell structure includes: supercell length, supercell orientation, and Cartesian coordinates of all atoms within the supercell; the molecular dynamics trajectory is a collection of multiple supercell structures, and also includes the supercell potential energy of each supercell structure and the forces acting on all atoms within the supercell.

[0030] The first-principles molecular dynamics method in step 2 employs a canonical ensemble, with the following specific parameter settings: plane wave cutoff energy of 450–500 eV, and electron energy convergence criterion of 1.0 × 10⁻⁶ eV. -4 ~1.0×10 -3 eV / atom, atomic force convergence criterion is 0.03~0.05 eV / Å, simulation step size is 0.5~1.0 fs, simulation duration is 5~10 ps; the simulation temperature of the undecomposed state in step 2 is 200~1000 K, and the simulation temperature of the pyrolysis state is 2500~4500 K; the molecular dynamics trajectory of step 2 is output 2~3 ps after the start of the simulation, including 4~6 undecomposed states with a simulation temperature difference of ≥100 K for each state, and including 2~4 pyrolysis states with a simulation temperature difference of ≥500 K for each state.

[0031] Step 3: Using molecular dynamics trajectories as the training set, DeepPot-SE is used to train and generate n machine learning potential functions.

[0032] Step 4: Randomly select one from n machine learning potential functions, and use classical molecular dynamics methods to simulate the pyrolysis process of energetic compounds at a specific time and temperature to obtain molecular dynamic trajectories. Use the supercell structure in the molecular dynamic trajectories as the test set.

[0033] The classical molecular dynamics method in step 4 uses a canonical ensemble with the following specific parameter settings: simulation step size of 0.1~0.25 fs; specific time l of 3~5 ps; specific temperature T of 3500~4500 K; the molecular dynamic trajectory obtained in step 4 includes 2~3 pyrolysis processes and the simulated temperature difference of each process is greater than or equal to 500 K.

[0034] In step 4, the supercell structure of the energetic compound is first simulated at 300 K for 10 ps. The molecular dynamics trajectory generated by this simulation is not included in the test set. The final structure is used as the initial state for simulating the pyrolysis process within a specific time l and at a specific temperature T.

[0035] Step 5: Substitute the test set into the n machine learning potential functions to calculate the supercell potential energy and the forces on all atoms in the supercell.

[0036] Step 6: Calculate the relative force deviation and the relative potential energy deviation of all atoms within the supercell in the test set, and set upper and lower limits for the relative force deviation and the relative potential energy deviation of the atoms; calculate the percentage of supercell structures whose relative force deviation and relative potential energy deviation are both less than the corresponding lower limits, specifically:

[0037] Step 6.1: Calculate the relative deviation of atomic forces and the relative deviation of supercell potential energy according to the following formula:

[0038]

[0039]

[0040] Among them, D F D represents the relative deviation of the forces acting on the atoms. E This represents the relative deviation of the supercell potential energy; The force on the i-th atom in any direction within the supercell structure; This represents the average force exerted on the i-th atom in any direction within the supercell structure, calculated using n machine learning numerical potential functions; n is the number of machine learning numerical potential functions generated through DeepPot-SE training; and m is a constant ranging from 1 to 1.5. The supercell potential energy is calculated for the j-th type of machine learning potential function. The supercell potential energy is calculated for the kth type of machine learning potential function.

[0041] Step 6.2: Define the lower limit of the relative force deviation of all atoms in the supercell as D.F(low) The upper limit is D F(high) The lower limit of the relative deviation of the supercell potential energy is defined as D. E(low) The upper limit is D E(high) ; Calculate the set of test cases that satisfy D E <D E(low) And D F <D F(low) The proportion of supercell structures R.

[0042] In specific implementation, D E(low) The value is 0.5~1.0 eV, D E(high) 1.5~2.0 eV; D F(low) It is 5%~10%, D F(high) It is 40% to 50%.

[0043] Step 7: If the proportion of such cases is less than or equal to a set threshold, calculate the potential energy and the forces on all atoms of the supercell structure that meet the requirements of both the relative deviation of atomic forces and the relative deviation of supercell potential energy falling within the corresponding upper and lower limits using first-principles methods. Add these to the training set and retrain the DeepPot-SE model to obtain n machine learning potential functions. Return to Step 5. If the proportion of such cases is greater than the set threshold, proceed to Step 8. Specifically:

[0044] Step 7.1: Set the threshold value to R low The value is 95%~98%, if R>R low Proceed to step 8.

[0045] Step 7.2: If R ≤ R low The first-principles method is used to calculate the result that conforms to D. E(low) ≤D E <D E(high) And D F(low) ≤D F <D F(high) The potential energy of the supercell structure and the forces acting on all atoms are then included in the training set.

[0046] Step 7.3: Retrain the DeepPot-SE model to obtain n machine learning potential functions, and return to step 5.

[0047] Step 8: Return to Step 4 and continue simulating the pyrolysis process at a specific time and temperature based on the existing molecular dynamics trajectory to obtain the continued molecular dynamics trajectory. Use the supercell structure in the continued molecular dynamics trajectory as the test set. Continue until the sum of the specific time reaches the preset duration to obtain the final machine learning potential function of the energetic compound pyrolysis.

[0048] In specific implementation, step 8 continues to simulate the pyrolysis process within a specific time l and at a specific temperature T based on the existing molecular dynamics trajectory, using the final state structure simulated in step 4 as the initial state for continued simulation; the ensemble, simulation step size, specific time l, and specific temperature T are consistent with step 4.

[0049] Let the preset duration in step 8 be L, N C N represents the number of carbon atoms in an energetic compound. O The number of oxygen atoms in an energetic compound, N H The number of hydrogen atoms in an energetic compound; when N O ≥N H / 2+N C or N C When N = 0, L is 50~100 ps; when N = 0, L is 50~100 ps. C ≠ 0 and N O <N H / 2+N C When L is 100~200 ps; when N C ≠0 and N O <N H When the value is 2 / 2, L is 200~300 ps.

[0050] The present invention will be described below through embodiments. Figure 1 This is a schematic diagram of the structure of the energetic compounds in three embodiments of the present invention.

[0051] Example 1

[0052] This embodiment provides a method for improving the accuracy of the pyrolysis machine learning potential function of energetic compounds. In this embodiment, the energetic compound is NH4N5 (pentazolium). The method includes the following steps:

[0053] Step 1: Obtain the crystal structure data of NH4N5 crystal, and optimize the molecular crystal structure data using the first-principles method according to the predefined parameters.

[0054] In this embodiment, the plane wave cutoff energy is 700 eV, and the electron energy convergence standard is 1.0 × 10⁻⁶ eV. -5 eV / atom, atomic force convergence standard 0.05 eV / Å, van der Waals weak interaction correction method adopted DFT-D3 method, Brillouin zone lattice number of 3×8×3.

[0055] Step 2: Based on the optimized NH4N5 crystal structure data, a supercell structure with a total number of atoms not exceeding 400 was established. First-principles molecular dynamics was used to simulate the undecomposed and pyrolysis states of the energetic compound to obtain the molecular dynamic trajectory.

[0056] In this embodiment, a 1×4×1 supercell structure of NH4N5 crystal was established, with a total of 160 atoms. The first-principles molecular dynamics simulation conditions included: NVT ensemble, plane wave cutoff energy of 500 eV, and electron energy convergence criterion of 1.0 × 10⁻⁶ eV. -3 The simulation step size was 0.5 eV / atom, the atomic force convergence standard was 0.05 eV / Å, the simulation step size was 0.5 fs, the simulation duration was 10 ps, ​​the six temperatures for the undecomposed state simulation were 200 K, 300 K, 400 K, 500 K, 600 K and 800 K, the three temperatures for the pyrolysis state simulation were 3000 K, 3500 K and 4000 K, and the molecular dynamics trajectory was output 2 ps after the simulation started.

[0057] Step 3: Using molecular dynamics trajectories as the training set, DeepPot-SE simulation training was used to obtain 6 machine learning numerical potential functions.

[0058] Step 4: Randomly select one of the six machine learning potential functions, and use classical molecular dynamics methods to simulate the pyrolysis process of NH4N5 within a specific time l and at a specific temperature T to obtain the molecular dynamic trajectory. Use the supercell structure in the molecular dynamic trajectory as the test set.

[0059] In this embodiment, the classical molecular dynamics simulation conditions include: a canonical ensemble, a simulation step size of 0.25 fs, simulating the NH4N5 supercell structure at 300 K for 10 ps, ​​the molecular dynamics trajectory generated by this simulation is not included in the test set, the final structure is used as the initial state for simulating the pyrolysis process, the specific time is 4 ps, and the three specific temperatures are 3000 K, 3500 K and 4000 K.

[0060] Step 5: Substitute the test set into the six machine learning potential functions to calculate the supercell potential energy and the forces on all atoms in the supercell.

[0061] Step 6: Calculate the relative force deviation and the relative potential energy deviation of all atoms in the supercell in the test set, and set the upper and lower limits of the relative force deviation and the relative potential energy deviation of the atoms; calculate the proportion of supercell structures whose relative force deviation and the relative potential energy deviation of the atoms are both less than the corresponding lower limits.

[0062] The relative force deviations of all atoms within the NH4N5 supercell calculated using six machine learning numerical potential functions are defined as D. F The relative deviation D of the supercell potential energy E Define the lower limit of the relative deviation of the forces on all atoms in the supercell as D. F(low) The upper limit is D F(high) The lower limit of the relative deviation of the supercell potential energy is defined as D. E(low) The upper limit is DE(high) In this embodiment, m=1, that is;

[0063]

[0064]

[0065] Calculate the test set that satisfies D E <D E(low) And D F <D F(low) The percentage of NH4N5 supercell structures, R. In this embodiment, D E(low) 0.5 eV, D F(low) It is 6%.

[0066] Step 7: If R ≤ R low The first-principles method is used to calculate the result that conforms to D. E(low) ≤D E <D E(high) And D F(low) ≤D F <D F(high) The potential energy of the NH4N5 supercell structure and the forces acting on all atoms are collected and placed into the training set. Six machine learning potential functions are obtained by training the DeepPot-SE model, and the process returns to step 5. If R > R... low Proceed to step 8;

[0067] In this embodiment, R low 98%, D E(high) 1.5 eV, D F(high) It is 40%.

[0068] Step 8: Return to Step 4 and continue simulating the pyrolysis process within a specific time l and at a specific temperature T based on the existing molecular dynamics trajectory to obtain the continued molecular dynamics trajectory. Use the supercell structure in the continued molecular dynamics trajectory as the test set. Continue until the sum of the specific time reaches a certain duration L to obtain the final machine learning potential function for the pyrolysis of the energetic compound.

[0069] In this embodiment, the specific time is 4 ps, and the three specific temperatures are 3000 K, 3500 K, and 4000 K, respectively. The N in the NH4N5 supercell structure... O =0 and N C =0, the total duration reached at a specific time is 60 ps.

[0070] In this embodiment, the deviations between the description of the NH4N5 unit cell structure and the X-ray single-crystal diffraction analysis results by the pyrolysis machine learning potential function of NH4N5 are listed in Table 1. The total energy drift of the adiabatic (NVE) simulation of the NH4N5 supercell based on the pyrolysis machine learning potential function is as follows: Figure 2 As shown.

[0071] Example 2

[0072] In this embodiment, the energetic compound is (NH3OH)N5 (pentazolidinyl hydroxylamine).

[0073] In this embodiment, step 1 is basically the same as step 1 in embodiment 1, except that the (NH3OH)N5 crystal uses a Brillouin zone lattice with a number of 8×2×4.

[0074] In this embodiment, step 2 is basically the same as step 2 in embodiment 1. The difference is that a 3×1×1 supercell structure of (NH3OH)N5 crystal is established, with a total number of atoms of 132. The three temperatures for pyrolysis simulation are 2000 K, 3000 K and 4000 K.

[0075] In this embodiment, step 4 is basically the same as step 4 in embodiment 1, except that the specific time is 3 ps.

[0076] In this embodiment, step 8 is basically the same as step 8 in embodiment 1, except that N in the (NH3OH)N5 supercell structure O =12, N O <N H / 2=24, N C =0, the total duration reached at a specific time is 80 ps.

[0077] In this embodiment, the deviations between the pyrolysis machine learning potential function of (NH3OH)N5 and the X-ray single-crystal diffraction analysis results of the (NH3OH)N5 cell structure description are listed in Table 1. The total energy drift of the adiabatic (NVE) simulation of the (NH3OH)N5 supercell based on the pyrolysis machine learning potential function is as follows: Figure 2 As shown.

[0078] Example 3

[0079] In this embodiment, the energetic compound is (C(NH2)3)N5 (pentazocineguanidine).

[0080] In this embodiment, step 1 is basically the same as step 1 in embodiment 1, except that the (C(NH2)3)N5 crystal uses a Brillouin zone lattice with a number of 4×4×8.

[0081] In this embodiment, step 2 is basically the same as step 2 in embodiment 1, except that a 1×1×2 supercell structure of (C(NH2)3)N5 crystal is established, with a total number of atoms of 240. The plane wave cutoff energy under the first-principles molecular dynamics simulation conditions is 400 eV, and the electron energy convergence criterion is 1.0 × 10⁻⁶ eV. -3eV / atom, atomic force convergence standard is 0.05 eV / Å, simulation step size is 1 fs, simulation duration is 10 ps, ​​the three temperatures for pyrolysis simulation are 3500 K, 4000 K and 4500 K, and molecular dynamics trajectory is output 3 ps after the simulation starts.

[0082] In this embodiment, step 4 is basically the same as step 4 in embodiment 1. The difference is that the simulation step size of the classical molecular dynamics simulation is 0.1 fs. The (C(NH2)3)N5 supercell structure is first simulated at 300 K for 10 ps. The molecular dynamic trajectory generated by this simulation is not included in the test set. The final structure is used as the initial state for simulating the pyrolysis process. The specific time is 5 ps, and the three specific temperatures are 3500 K, 4000 K and 4500 K.

[0083] In this embodiment, step 6 is basically the same as step 6 in embodiment 1, except that D E(low) =1 eV, D F(low) It is 10%.

[0084] In this embodiment, step 7 is basically the same as step 7 in embodiment 1, except that R low 95%, D E(high) 2.0 eV, D F(high) It is 50%.

[0085] In this embodiment, step 8 is basically the same as step 8 in embodiment 1, except that N in the (C(NH2)3)N5 supercell structure O =0, N C =16, N H =96, N O <N H / 2, where the total duration of a specific time is 250 ps.

[0086] In this embodiment, the deviations between the description of the (C(NH2)3)N5 unit cell structure by the pyrolysis machine learning potential function and the X-ray single-crystal diffraction analysis results are listed in Table 1. The total energy drift of the adiabatic (NVE) simulation of the (C(NH2)3)N5 supercell based on the pyrolysis machine learning potential function is as follows: Figure 2 As shown.

[0087] Table 1. Cell length, cell volume, and cell density of Examples 1 to 3

[0088]

[0089] To verify the reliability of the pyrolysis machine learning potential function in describing the supercell energy of the pyrolysis reaction, 3×7×3 (2520 atoms) supercells, 6×2×4 (2112 atoms) supercells, and 2×2×5 (2400 atoms) supercells were established for NH4N5 (pentazolidinium), (NH3OH)N5 (pentazolidinyl hydroxylamine), and (C(NH2)3)N5 (pentazolidinyl guanidine), respectively. Based on the pyrolysis machine learning potential function trained in this invention, NVE simulations with a time of 100 ps were performed at 4000 K with step sizes of 0.1 fs, 0.2 fs, and 0.5 fs, respectively. NVE simulations with a time of 100 ps were also performed at 4000 K based on the ReaxFF force field with a step size of 0.05 fs, respectively. The energy drift curves are shown below. Figure 2 As shown.

[0090] Figure 3 This is a schematic diagram of the initial and final structures of NH4N5 (pentazolium ammonium) pyrolysis simulated using a pyrolysis machine learning potential function. The simulated energy drift under three time step conditions is within 2 kcal / mol, and the energy drift is related to the time step. At a time step of 0.1 fs, the energy drift is within 0.25 kcal / mol, which is much smaller than the energy drift simulated using a ReaxFF force field with a time step of 0.05 fs.

[0091] As can be seen from Examples 1 to 3:

[0092] (1) The pyrolysis machine learning potential function can accurately describe the chemical bonds and weak intermolecular interactions in the structure of pentazol organic salts, with an accuracy comparable to that of density functional calculation.

[0093] (2) The pyrolysis machine learning potential function can accurately describe the energy of pentazol organic salts in the adiabatic (NVE) simulation process and is suitable for the pyrolysis molecular dynamics simulation of this type of compound.

[0094] (3) The pyrolysis machine learning potential function has a much higher accuracy in describing the energy of pentazol organic salts in the molecular dynamics simulation process than ReaxFF.

[0095] In summary, the above analysis shows that training the pyrolysis machine learning potential function of energetic materials according to the method of the present invention can improve the accuracy. The trained pyrolysis machine learning potential function has a much higher accuracy in describing the energy in the molecular dynamics simulation of pentazol organic salt energetic compounds than ReaxFF, indicating that the method of the present invention for improving the accuracy of the pyrolysis machine learning potential function of energetic compounds is reliable.

[0096] This invention provides technical support for machine learning potential function modeling of pyrolysis of novel energetic compounds, provides a basis for molecular design of novel energetic compounds, improves the scientific nature of molecular design, and develops a new method for applying machine learning combined with molecular dynamics to the computation of energetic materials.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of improving the accuracy of a machine-learned potential function for pyrolysis of an energetic compound, the method comprising: include: Step 1: Obtain the molecular crystal structure data of the energetic compound molecular crystal, and optimize the molecular crystal structure data using the first-principles method according to the predefined parameters; Step 2: Based on the optimized molecular crystal structure data, establish a supercell structure with a total number of atoms not exceeding 400, and use first-principles molecular dynamics to simulate the undecomposed and pyrolysis states of the energetic compound to obtain the molecular dynamic trajectory; Step 3: Using molecular dynamics trajectories as the training set, DeepPot-SE is used to train and generate n machine learning potential functions; Step 4: Randomly select one from n machine learning potential functions, use classical molecular dynamics methods to simulate the pyrolysis process of energetic compounds at a specific time and temperature, obtain molecular dynamic trajectories, and use the supercell structure in the molecular dynamic trajectory as the test set; Step 5: Substitute the test set into the n machine learning potential functions to calculate the supercell potential energy and the forces on all atoms in the supercell; Step 6: Calculate the relative force deviation and the relative potential energy deviation of all atoms in the supercell in the test set, and set the upper and lower limits of the relative force deviation and the relative potential energy deviation of the atoms; calculate the proportion of supercell structures whose relative force deviation and relative potential energy deviation of the atoms are both less than the corresponding lower limits. Step 7: If the proportion of the number is less than or equal to the set threshold, use the first-principles method to calculate the potential energy and the force on all atoms of the supercell structure that meets the requirements of the relative deviation of atomic force and the relative deviation of supercell potential energy, and put them into the training set. Retrain the DeepPot-SE model to obtain n machine learning potential functions, and return to step 5; if the proportion of the number is greater than the set threshold, proceed to step 8. Step 8: Return to step 4 and continue to simulate the pyrolysis process at a specific time and temperature based on the existing molecular dynamics trajectory to obtain the continued molecular dynamics trajectory. Use the supercell structure in the continued molecular dynamics trajectory as the test set. The machine learning potential function for the pyrolysis of the energetic compound is obtained when the total time reaches a preset duration.

2. The method of claim 1, wherein, The molecular crystal structure data includes: cell length, cell orientation, and original fractional coordinates of all atoms within the molecular crystal; predefined parameters include: plane wave cutoff energy, number of Brillouin lattice points, electron energy convergence criterion, and atomic force convergence criterion; the optimized molecular crystal structure data includes: optimized cell length, optimized cell orientation, and optimized fractional coordinates of all atoms within the molecular crystal. The predefined parameter values ​​range as follows: plane wave cutoff energy of 500~850 eV, and electron energy convergence criterion of 1.0×10⁻⁶ eV. -6 ~1.0×10 -4 eV / atom, the atomic force convergence standard is 0.01~0.05 eV / Å, the Brillouin zone lattice point number expression is L×M×N, where L, M and N are positive integers; let the unit cell lengths of the molecular crystal be a, b and c, then 30 ≤ a×L ≈ b×M ≈ c×N ≤ 70.

3. The method of claim 1, wherein, The supercell structure includes: supercell length, supercell orientation, and Cartesian coordinates of all atoms within the supercell; the molecular dynamics trajectory is a collection of multiple supercell structures, and also includes the supercell potential energy of each supercell structure and the forces acting on all atoms within the supercell.

4. The method of claim 1, wherein, The first-principles molecular dynamics method in step 2 employs a canonical ensemble, with the following specific parameter settings: plane wave cutoff energy of 450–500 eV, and electron energy convergence criterion of 1.0 × 10⁻⁶ eV. -4 ~1.0×10 -3 eV / atom, atomic force convergence criterion is 0.03~0.05 eV / Å, simulation step size is 0.5~1.0 fs, simulation duration is 5~10 ps; the simulation temperature of the undecomposed state in step 2 is 200~1000 K, and the simulation temperature of the pyrolysis state is 2500~4500 K; the molecular dynamics trajectory of step 2 is output 2~3 ps after the start of the simulation, including 4~6 undecomposed states with a simulation temperature difference of ≥100 K for each state, and including 2~4 pyrolysis states with a simulation temperature difference of ≥500 K for each state.

5. The method of claim 1, wherein, The classical molecular dynamics method in step 4 uses a canonical ensemble with the following specific parameter settings: simulation step size of 0.1~0.25 fs; specific time l of 3~5 ps; specific temperature T of 3500~4500 K; the molecular dynamic trajectory obtained in step 4 contains 2~3 pyrolysis processes and the temperature difference between each process is greater than or equal to 500 K. In step 4, the supercell structure of the energetic compound is first simulated at 300 K for 10 ps. The molecular dynamics trajectory generated by this simulation is not included in the test set. The final structure is used as the initial state for simulating the pyrolysis process within a specific time l and at a specific temperature T.

6. The method of claim 3, wherein the machine-learned potential function is a neural network. Step 6 specifically involves: Step 6.1: Calculate the relative deviation of atomic forces and the relative deviation of supercell potential energy according to the following formula: Among them, D F D represents the relative deviation of the forces acting on the atoms. E This represents the relative deviation of the supercell potential energy; The force on the i-th atom in any direction within the supercell structure; This represents the average force exerted on the i-th atom in any direction within the supercell structure, calculated using n machine learning numerical potential functions; n is the number of machine learning numerical potential functions generated through DeepPot-SE training; and m is a constant ranging from 1 to 1.

5. The supercell potential energy is calculated for the j-th type of machine learning potential function. The supercell potential energy is calculated for the kth type of machine learning potential function. Step 6.2: Define the lower limit of the relative force deviation of all atoms in the supercell as D. F(low) The upper limit is D F(high) The lower limit of the relative deviation of the supercell potential energy is defined as D. E(low) The upper limit is D E(high) ; Calculate the set of test cases that satisfy D E <D E(low) And D F <D F(low) The proportion of supercell structures R.

7. The method of claim 6, wherein the machine-learned potential function is a neural network. D E(low) is 0.5-1.0 eV, D E(high) is 1.5-2.0 eV;D F(low) is 5%-10%, D F(high) is 40%-50%.

8. The method of claim 6, wherein the machine-learned potential function is a neural network. Step 7 specifically involves: Step 7.1: Set the threshold value to R low The value is 95%~98%, if R>R low Proceed to step 8; Step 7.2: If R ≤ R low The first-principles method is used to calculate the result that conforms to D. E(low) ≤D E <D E(high) And D F(low) ≤D F <D F(high) The potential energy of the supercell structure and the forces acting on all atoms are then included in the training set. Step 7.3: Retrain the DeepPot-SE model to obtain n machine learning potential functions, and return to step 5.

9. The method for improving the accuracy of a machine-learned potential function for pyrolysis of an energetic compound of claim 1, wherein, Step 8 continues to simulate the pyrolysis process within a specific time l and at a specific temperature T based on the existing molecular dynamics trajectory, using the final state structure simulated in each step 4 as the initial state for further simulation. The ensemble, simulation step size, specific time l, and specific temperature T remain consistent with step 4; Let the preset duration in step 8 be L, N C N represents the number of carbon atoms in an energetic compound. O The number of oxygen atoms in an energetic compound, N H The number of hydrogen atoms in an energetic compound; when N O ≥N H / 2+N C or N C When N = 0, L is 50~100 ps; when N = 0, L is 50~100 ps. C ≠0 and N O <N H / 2+N C When L is 100~200 ps; when N C ≠0 and N O <N H When the value is 2 / 2, L is 200~300 ps.

Images