Method for simulating thermal decomposition process of insulating gas based on deep learning potential function

Abstract

The invention discloses a method for simulating an insulating gas thermal decomposition process based on a deep learning potential function, and relates to the technical field of insulating gas molecular dynamics simulation. The method comprises the following steps of: simulating a target insulating gas system by adopting a de novo calculation molecular dynamics method based on a density functional theory, and acquiring atomic coordinates, total energy corresponding to each frame of atomic configuration, acting force borne by atoms and a Virie tensor to form an initial data set; constructing an initial deep learning potential function model, training, verifying and testing, and outputting to obtain an optimal deep learning potential function model; and performing molecular dynamics simulation on the thermal decomposition process of the unknown insulating gas by using the optimal deep learning potential function model to obtain a molecular dynamics simulation track. According to the invention, super-large-scale molecular dynamics simulation of the thermal decomposition process of the complex insulating gas is realized, and a temperature-dependent reaction dynamics law and a chemical bond fracture mechanism which cannot be obtained by a traditional simulation method can be disclosed.

Description

Technical Field

[0001] This invention relates to the field of molecular dynamics simulation technology for insulating gases, and in particular to a method for simulating the thermal decomposition process of insulating gases based on deep learning potential functions. Background Technology

[0002] Insulating gases are crucial for power systems. Sulfur hexafluoride (SF6) is widely used in switchgear, transmission pipelines, and other equipment due to its excellent insulation and arc-quenching properties. However, SF6's extremely high global warming potential (GWP=23900) and ultra-long atmospheric lifetime (3200 years) have caused long-term and irreversible environmental impacts, necessitating the exploration of green and environmentally friendly alternative insulating media.

[0003] Researchers have conducted extensive studies on various SF6 alternative gases. Currently, candidate gases include CF3I and C5F. 10 While O, C4F7N, and their mixtures possess high insulating properties, these candidate gases may decompose in electrical equipment due to localized overheating or discharge faults, producing byproducts. This could potentially affect the equipment's insulation performance and pose potential environmental risks. Therefore, a thorough understanding of the thermal decomposition mechanism of these candidate gases is essential before engineering applications. Currently, at the experimental level, the difficulty in real-time, in-situ reproduction and precise control of extreme operating conditions (such as high temperature, high pressure, and discharge environments) makes direct observation of reaction dynamics at the atomic scale extremely challenging. Related experiments also face challenges such as high equipment requirements, long cycles, high costs, and potential safety risks, further limiting their widespread application.

[0004] In terms of simulating the thermal decomposition path of insulating gases, the existing mainstream methods have obvious shortcomings: (1) Ab initio molecular dynamics (AIMD), although it can accurately describe the changes in electronic structure and chemical reaction processes based on the principles of quantum mechanics, consumes a lot of computational resources and is usually limited to simulations at the scale of hundreds of atoms and picoseconds, making it difficult to cope with the evolution simulation of actual insulating gas systems at real time and space scales; (2) Methods based on empirical force fields or reactions, although they have high computational efficiency and can handle larger systems and longer time periods, their accuracy depends heavily on the applicability of the parameter set, and their reliability is often insufficient when dealing with key chemical processes such as bond breaking, charge transfer, and reaction barriers.

[0005] Therefore, there is an urgent need in this field for a new method that can maintain the accuracy of quantum computing while achieving the efficiency of large-scale molecular dynamics simulations, in order to accurately predict the thermal decomposition pathways and products of insulating gases. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for simulating the thermal decomposition process of insulating gases based on a deep learning potential function. The method provided by this invention can be used to simulate the thermal decomposition process of environmentally friendly insulating gases with high precision and efficiency. This method combines first-principles calculations with active learning techniques to train a deep learning potential function model that rivals the precision of quantum computing while possessing the efficiency of classical molecular dynamics simulations. Using this deep learning potential function model, the decomposition mechanism, pathway, and product distribution of insulating gases and their mixtures with buffer gases (such as CO2 and N2) under high temperature and high pressure can be studied in depth at the atomic scale. This provides key theoretical basis and technical tools for the design, safety assessment, and engineering application of new environmentally friendly insulating gases to replace SF6. The technical means of this invention are specifically implemented through the following techniques.

[0007] A method for simulating the thermal decomposition process of insulating gases based on deep learning potential functions includes the following steps:

[0008] Ab initio molecular dynamics based on density functional theory was used to simulate the target insulating gas system at predetermined temperatures. Atomic coordinates, total energy corresponding to each atomic configuration, force on each atom, and virial tensor were collected to form an initial dataset, which was then divided into training, validation, and test sets.

[0009] An initial deep learning potential function model is constructed, and the initial deep learning potential function model is trained, validated, and tested using the training set, validation set, and test set. Finally, the optimal deep learning potential function model is obtained.

[0010] By using a deep learning potential function model, molecular dynamics simulations of the thermal decomposition process of an unknown insulating gas are performed to obtain the molecular dynamics simulation trajectory and analyze the results.

[0011] Furthermore, the method for obtaining and preprocessing the initial dataset is as follows:

[0012] An initial unit cell is constructed, and a molecular dynamics simulation based on quantum mechanics is performed. All evolved configurations and their physical properties are recorded to obtain the original output simulation trajectory. All configurations and their physical properties include atomic coordinates, the total energy corresponding to each frame of atomic configuration, the force on each atom, and the virial tensor.

[0013] Several frames of discretized, static atomic configurations and their full set of quantum mechanical precision physical quantity labels are sampled from the simulated trajectory of the original output results. The total energy corresponding to each frame of atomic configuration, the force on each atom, and the virial tensor are extracted to form the initial dataset.

[0014] Furthermore, the initial unit cell comprises a plurality of the target insulating gases and a plurality of buffer gases.

[0015] Furthermore, the method for molecular dynamics simulation based on quantum mechanics is as follows:

[0016] AIMD calculations were performed on the initial unit cell using VASP, and the simulation was conducted under the NVT ensemble.

[0017] The temperature was controlled using a Nose-Hoover thermal bath, with the temperature range set to 300-3200 K. Simulations were run for 20 ps at each temperature point, with a time step of 0.5 fs.

[0018] The forces acting on atoms include the interaction between ions and electrons, as well as the interaction between electrons; the interaction between ions and electrons is described using the projected fused wave pseudopotential, and the interaction between electrons is handled using the PBE functional; the cutoff energy of the plane wave basis set is set to 520 eV.

[0019] The convergence criteria for energy and the forces acting on atoms are set to 10. -6 eV and 0.01 eV / Å.

[0020] Furthermore, the initial dataset is divided into a training set, a validation set, and a test set in proportions of 60%, 20%, and 20%, respectively.

[0021] Furthermore, the method for constructing the initial deep learning potential function model is as follows:

[0022] The initial deep learning potential function model was built and trained using the DeePMD-kit toolkit; the "se_e2_a" descriptor type was selected, and the cutoff radius was set to 10.00 Å; the descriptor embedding network adopted a three-layer structure with 25, 50, and 100 neurons respectively; the fitting network adopted a three-layer structure with 250 neurons in each layer.

[0023] Training employed an exponentially decaying learning rate strategy, initially set at 0.001 and eventually reduced to 3.51 × 10⁻⁶. -8 ; Assign loss function weights to the total energy corresponding to each frame's atomic configuration, the force acting on each atom, and the virial tensor;

[0024] Based on the total energy loss, atomic force loss, and virial tensor loss values ​​calculated for each frame of atomic configuration on the validation set, the weights are adjusted to ensure that the weighted loss terms are on the same order of magnitude until training converges.

[0025] Furthermore, the initial ratio of the total energy, the force on each atom, and the loss weight coefficient of the virial tensor corresponding to each frame of atomic configuration is set to (0.5-2):(500-2000):(0.01-0.2).

[0026] Furthermore, the method for training, validating, and testing the initial deep learning potential function model using the training set, and finally outputting the optimal deep learning potential function model, is as follows:

[0027] Using the training set and the same deep neural network architecture, multiple deep learning potential function models with identical structures are trained in parallel by independently assigning different random initial weight parameters to each deep learning potential function model, thus forming a deep learning potential function model set.

[0028] Calculate the standard deviation of the force prediction of the atoms in each new configuration by the deep learning potential function model, and select configurations with a standard deviation higher than a preset threshold.

[0029] First-principles calculations were used to compute the selected configurations to expand the dataset;

[0030] Exploratory molecular dynamics simulations were performed in LAMMPS using a set of deep learning potential function models. The maximum standard deviation ε of the forces on atoms predicted by the set of deep learning potential function models was used as an uncertainty measure to screen out molecular configurations with 0.05 eV / Å ≤ ε < 0.50 eV / Å. Accurate DFT calculations were performed using VASP and these configurations were added to the training set.

[0031] The training continues in a loop until the training set can fully describe all relevant reaction paths and configuration space, at which point the loop training terminates.

[0032] The deep learning potential function model after the termination of the loop training is evaluated using the test set, and the optimal deep learning potential function model is output.

[0033] Furthermore, the method of expanding the dataset by using first-principles calculations to compute the selected configurations is as follows:

[0034] VASP was used, based on DFT, to generate the same pseudopotential, functional, and cutoff energy parameters as the initial data;

[0035] High-precision single-point energy calculations are performed on the selected atomic configurations to obtain accurate values ​​for their total energy, atomic force, and virial tensor.

[0036] The configuration and its corresponding physical quantity are added as a new data sample to the training set.

[0037] Furthermore, the method for determining whether the training set can adequately describe all relevant reaction pathways and configuration space is as follows:

[0038] When the maximum standard deviation ε of the forces on atoms predicted by the deep learning potential function model set is consistently lower than the preset threshold, it is considered that the current training set can fully cover all relevant reaction paths and configuration spaces of the target system under simulated conditions, and the active learning loop can be terminated.

[0039] Compared with the prior art, the advantages of the present invention are:

[0040] 1. Achieving a balance between high precision and high efficiency. A deep learning potential function (DLP) was developed, which achieves the accuracy of first-principles calculations while maintaining computational efficiency comparable to empirical potential function simulations, thereby enabling the simulation of ultra-large-scale molecular dynamics for complex gas-phase reaction systems.

[0041] 2. Explaining microscopic mechanisms. The deep learning molecular dynamics framework provided by this invention can reveal temperature-dependent reaction kinetics and chemical bond breaking mechanisms that are impossible to obtain using traditional simulation methods.

[0042] 3. Guiding engineering applications. Simulation results can provide key data support for optimizing the formulation of environmentally friendly insulating gases (such as the optimal mixing ratio), setting equipment operating conditions, and conducting safety assessments, thereby accelerating the SF6 replacement process. Attached Figure Description

[0043] Figure 1 A flowchart illustrating the workflow for deep learning potential energy training and calculation of insulating gas decomposition products at different temperatures.

[0044] Figure 2 This is a graph showing the accuracy verification of the deep learning potential function model in step two.

[0045] Figure 3 The graphs show the decomposition kinetics of CF3SO2F under different conditions in step three.

[0046] Figure 4 This is a schematic diagram of the different decomposition paths of CF3SO2F in CO2 and N2 environments in step three.

[0047] Figure 5 Let g(r) be the radial distribution function of CF3SO2F / N2 at 1400 K (g) plot, 2200 K (h) plot, and 3200 K (i) plot. Detailed Implementation

[0048] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The present invention provides a method for simulating the thermal decomposition process of insulating gas based on deep learning potential function, comprising the following steps: acquisition and preprocessing of initial dataset, training, verification and testing of deep learning potential function model, molecular dynamics simulation of the thermal decomposition process of unknown insulating gas, and result analysis.

[0050] Step 1: Obtaining and preprocessing the initial dataset.

[0051] We employ an ab initio molecular dynamics (AIMD) method based on density functional theory (DFT) to simulate target insulating gas systems (e.g., a mixture of CF3SO2F and buffer gases CO2 / N2) at different temperatures (e.g., 300-3200 K). We collect atomic coordinates, the total energy corresponding to each atomic configuration, the forces acting on each atom, and the virial tensor to form an initial dataset for subsequent training, validation, and testing of deep learning potential function models.

[0052] Step 2: Training, validation, and testing of the deep learning potential function (DLP) model.

[0053] 1. Construct the initial deep learning potential function model; divide the initial dataset obtained in step one into a training set, a validation set, and a test set.

[0054] 2. Train the initial deep learning potential function model using the training set.

[0055] Specifically, atomic coordinates are used as input, and the predicted total energy of the system, the forces acting on the atoms, and the virial tensor are used as output targets. Atomic coordinates include: (1) a list of element types in the target system, i.e., all types of atoms in the target system; (2) an atomic coordinate file containing the initial molecular arrangement, i.e., the initial atomic configuration; and (3) an initial configuration file for the deep learning potential function that defines the network architecture and training parameters.

[0056] 3. Use the validation set to monitor the training process of the initial deep learning potential function model to prevent overfitting until training and validation are completed.

[0057] 4. Use an independent test set to evaluate the accuracy of the trained deep learning potential function model, ensuring that the root mean square error (RMSE) of the predictions of total energy, forces on atoms, and virial tensors are in high agreement with the DFT calculation results.

[0058] Step 3: Perform molecular dynamics simulations on the thermal decomposition process of the unknown insulating gas.

[0059] The trained optimal deep learning potential function model is embedded into a large-scale atomic / molecular parallel simulator (LAMMPS) to construct a simulation system containing thousands to tens of thousands of atoms. Different initial conditions (temperature, pressure, gas mixing ratio) are set, and molecular dynamics simulation is run to simulate the thermal decomposition process of insulating gas and obtain the molecular dynamics simulation trajectory.

[0060] Step 4: Results Analysis.

[0061] The following information was extracted and analyzed from molecular dynamics simulation trajectories:

[0062] (1) Curve of the evolution of the decomposition rate of gas molecules over time;

[0063] (2) The quantity and proportion of different decomposition products (such as CF4, SO2, COF2, SOF2, CF3, S, C, etc.);

[0064] (3) The process of chemical bond breaking and formation (analyzed by radial distribution function RDF);

[0065] (4) Reveal the effects of temperature, pressure, type of buffer gas (CO2 or N2) and mixing ratio on decomposition mechanism and kinetics.

[0066] Example

[0067] The method for simulating the thermal decomposition process of insulating gas provided in this embodiment uses the insulating gas CF3SO2F as the object to simulate the thermal decomposition process. The overall process is as follows: Figure 1 As shown, the specific steps are as follows.

[0068] Step 1: Obtaining and preprocessing the initial dataset.

[0069] This step uses ab initio molecular dynamics (AIMD) based on density functional theory (DFT) to simulate the target insulating gas system and generate an initial dataset. All calculations were performed using the Vienna Ab initio Simulation software package (VASP).

[0070] 1. Construct a small system (initial unit cell) containing several molecules as the initial model system.

[0071] Specifically, this initial unit cell is a periodic unit cell containing 3 trifluoromethanesulfonyl fluoride (CF3SO2F) molecules and 6 buffer gas (CO2 or N2) molecules. This initial unit cell is used as the initial model system to effectively cover all possible atomic configurations of the CF3SO2F mixed gas.

[0072] 2. Perform AIMD simulation based on DFT on the above initial model system, record all evolved forms (configurations) and their physical properties, and obtain the original output simulation trajectory.

[0073] All morphologies (configurations) and their physical properties include: atomic coordinates (i.e., the distribution of atoms in a unit cell expressed in coordinates (x, y, z) of a spatial rectangular coordinate system), the total energy corresponding to each frame of atomic configuration, the force acting on each atom, and the virial tensor.

[0074] Specifically, AIMD calculations were performed on the initial unit cell system using VASP, and the simulation was conducted under the NVT ensemble. A Nose-Hoover thermoelectric bath was used to control the temperature; the temperature range was broadly set to 300-3200 K (1400 K is the temperature at which significant thermal decomposition of the gas occurs, therefore...). Figure 1 The simulation covers a wide temperature range (not including plots below 1400K) to encompass various thermal conditions from normal operation to arc faults; simulations are run for 20 ps at each temperature point with a time step of 0.5 fs.

[0075] Specifically, the forces acting on atoms include the interaction between ions and electrons, as well as the interaction between electrons. The interaction between ions and electrons is described using a projector augmented wave (PAW), while the interaction between electrons is handled using a PBE functional. The Brillouin zone integral is performed at the Γ point, and the cutoff energy of the plane-wave basis set is set to 520 eV.

[0076] Specifically, the convergence criteria for energy (i.e., the change in electron energy between two iterations during the electron self-consistent field iteration process) and the force exerted on the atom are set to a strict 10. -6 eV and 0.01 eV / Å provide a high-quality data foundation for subsequent machine learning.

[0077] The original output simulation trajectory after AIMD simulation is a complete record of a series of atomic configurations and their physical properties, including the time dimension. It records the coordinates of all atoms in the initial model system, the unit cell vector, and the total energy of the corresponding configuration, the force on each atom, and the virial tensor obtained by real-time DFT calculation at each simulation step (i.e., each frame) in a complete time sequence.

[0078] 3. Randomly sample and discretize a set of data points from the simulated trajectory of these original output results.

[0079] Thousands of static atomic configurations and their full set of quantum mechanical precision physical quantity labels are randomly selected. The total energy, the force on each atom, and the virial tensor corresponding to each atomic configuration are extracted to form the initial dataset for subsequent training, validation, and testing of the initial deep learning potential function model.

[0080] Step 2: Training, validating, and testing the deep learning potential function (DLP) model.

[0081] 1. Divide into training set, validation set and test set.

[0082] The initial dataset obtained in step one above is divided into training set, validation set and test set in proportions of 60%, 20% and 20% respectively.

[0083] 2. Construction of the initial deep learning potential function (DLP) model.

[0084] (1) Use the DeePMD-kit toolkit to build and train the initial deep learning potential function model.

[0085] The "se_e2_a" descriptor type was selected, and a cutoff radius of 10.00 Å was set to capture interatomic interactions. The descriptor embedding network uses a three-layer structure with 25, 50, and 100 neurons respectively. The fitting network uses a three-layer structure, with each layer containing 250 neurons.

[0086] (2) The training adopted an exponential decay learning rate strategy, with the initial value set at 0.001 and eventually reduced to 3.51×10. -8 .

[0087] (3) Assign appropriate loss function weights to the total energy corresponding to each frame's atomic configuration, the force on each atom, and the virial tensor to optimize the training process.

[0088] In this step, the loss function is a weighted sum of the total energy corresponding to each frame's atomic configuration, the forces acting on each atom, and the virial tensor loss. The training process is optimized by configuring the weight coefficients of the loss function.

[0089] Specifically, the initial ratio of the total energy corresponding to each frame of atomic configuration, the force on each atom, and the loss weight coefficient of the virial tensor is set to (0.5-2):(500-2000):(0.01-0.2).

[0090] 3. Implementation of the Active Learning Cycle (DP-GEN).

[0091] DP-GEN is an iterative process, with each cycle consisting of training, exploration, and labeling. Specifically, it involves the following three steps:

[0092] (1) Train a set of deep learning potential function models using the current training set.

[0093] Using the current training set and the same deep learning neural network architecture, multiple deep learning potential function models with identical structures are trained in parallel by independently assigning different random initial weight parameters to each model. These models form a deep learning potential function model set.

[0094] Calculate the standard deviation of the force prediction of the atoms in each new configuration by the deep learning potential function model, and select configurations with standard deviations higher than a preset threshold.

[0095] First-principles calculations are used to precisely compute the selected configurations to expand the training set. The purpose of expanding the training set is to optimize the model when the trained model fails to meet our requirements, such as having too large an error. The expanded training set can train a new deep learning potential function model that maintains high accuracy in a wider configuration space.

[0096] During training, the weights are dynamically adjusted based on the magnitudes of each loss term on the validation set (i.e., the total energy loss, atomic force loss, and virial tensor loss calculated on the validation set), ensuring that the weighted loss terms are of the same order of magnitude, until training converges. When the loop converges (i.e., no more high-uncertainty configurations appear during exploration), the final high-precision deep learning potential function model and its corresponding complete training set are obtained.

[0097] Specifically, the method of using first-principles calculations to accurately calculate the selected configurations to expand the dataset is as follows: ① Using VASP, based on DFT, the same pseudopotential, functional, and cutoff energy parameters as the initial data are generated; ② Performing high-precision single-point energy calculations on the selected atomic configurations to obtain the accurate values ​​of their total energy, atomic force, and virial tensor; ③ Adding the configuration and its corresponding physical quantities as a new data sample to the training set.

[0098] (2) Using a set of deep learning potential function models, exploratory molecular dynamics simulations were performed in LAMMPS (Large-scale Atomic / Molecular Massively Parallel Simulator), and the maximum standard deviation (ε) of the forces on atoms predicted by the set of deep learning potential function models was used as the uncertainty measure. Molecular configurations with high uncertainty (0.05 eV / Å≤ε<0.50 eV / Å) were selected, and accurate DFT calculations were performed using VASP, and these were added to the training set.

[0099] Specifically, the exploratory molecular dynamics simulation method is as follows: During the exploration process, NVT iterations are performed using LAMMPS, and the maximum standard deviation ε of the forces acting on atoms predicted by the deep learning potential function model set is used as the error metric. Atomic configurations with model bias in the range of 0.05 eV / A ≤ ε < 0.50 eV / A are selected for the annotation process.

[0100] Specifically, the method for accurate DFT calculation using VASP is as follows: Using VASP software, based on density functional theory, employing the PAW pseudopotential and PBE functional, at a plane wave cutoff energy of 520 eV and 10... -6 Under the energy convergence criterion of eV, single-point energy calculations or structural relaxations are performed on the selected atomic configurations to obtain their accurate total energy, the forces acting on the atoms, and the virial tensor.

[0101] (3) The above steps (1)-(2) are repeated continuously until the training set can fully describe all relevant reaction paths and configuration space.

[0102] The criterion for determining whether the training set can adequately describe all relevant reaction pathways and configuration spaces (i.e., the condition for terminating training) is as follows: In exploratory DE molecular dynamics simulations, the prediction uncertainty of newly generated atomic configurations (measured by the maximum standard deviation ε of the forces acting on the atoms), as determined by the deep learning potential function model set, remains consistently below a preset threshold (0.05 eV / Å). At this point, it is considered that the current training set can adequately cover all relevant reaction pathways and configuration spaces of the target system under simulation conditions, and the active learning loop terminates.

[0103] 4. After training is completed, the trained deep learning potential function model is rigorously evaluated using an independent test set.

[0104] The method for rigorously evaluating the model using an independent test set is as follows: calculate the root mean square error between the model's predicted values ​​and the DFT calculated values; the core accuracy metrics are: total energy error 1.3 meV / atom, atomic force error 0.023 eV / Å, and virial tensor error 2.5 meV / atom.

[0105] like Figure 2 As shown in the accuracy verification figure of the deep learning potential function (DLP) model, the root mean square error (RMSE) of the total energy prediction for the CF3SO2F molecular system is as low as 1.3 meV / atom, the RMSE of the predicted forces on the atoms is 0.023 eV / Å, and the RMSE of the virial tensor is 2.5 meV / atom. The error distribution closely revolves around zero.

[0106] This indicates that the trained deep learning potential function model has the ability to simulate the chemical reaction of the CF3SO2F system with high precision, achieving an accuracy comparable to DFT calculation, and can be output as the optimal deep learning potential function model.

[0107] Step 3: Molecular dynamics simulation of the thermal decomposition process of the unknown insulating gas.

[0108] The trained optimal deep learning potential function model is deployed to LAMMPS (Large-scale Atomic / Molecular Massively Parallel Simulator) for large-scale simulation.

[0109] 1. To approximate macroscopic properties, this embodiment constructs a large-scale mixed gas system containing 500 CF3SO2F molecules and 3667 CO2 molecules (a total of 15,001 atoms), i.e. Figure 1 The supercellular process in [the study] was investigated. This large mixed gas system was placed in a cubic box with a side length of 553.24 Å and a density of 0.002328 g / cm³. 3 It precisely corresponds to the physical state of a 12% CF3SO2F / 88% CO2 (volume ratio / molar ratio) gas mixture at 0.1 MPa and 25°C.

[0110] 2. Before large-scale simulation, the large mixed gas system is first relaxed for 5 ps under the NPT ensemble to stabilize its density. Subsequently, the large mixed gas system is relaxed for 10 ps at 1000 K under the NVT ensemble to minimize energy.

[0111] 3. After the large mixed gas system has relaxed, an NVT production simulation for 1000 ps is conducted at the target temperatures (1400 K, 2200 K, 3200 K). The simulation uses a Nose-Hoover thermostat to control the temperature, with a temperature damping parameter set to 0.1 ps, and a time step of 0.5 fs to accurately integrate the atomic motion equations.

[0112] Throughout the simulation, the forces acting on the atoms are provided by the pre-trained optimal deep learning potential function model via the LAMMPS pair_style deepmd command.

[0113] 4. Simulated molecular dynamics trajectories are output periodically, such as... Figure 4 As shown, this is used for subsequent results analysis.

[0114] Step 4: Results Analysis.

[0115] By analyzing the molecular dynamics simulation trajectory output in step three, the curve of the change in the number of CF3SO2F molecules over time is extracted.

[0116] By analyzing the output molecular dynamics simulation trajectory, the number of complete CF3SO2F molecules is identified and counted based on interatomic bond lengths.

[0117] The specific analysis method is as follows:

[0118] (1) Read the atomic coordinates and element type of each frame, and construct a bonding connection diagram between atoms based on whether the distance between atoms is less than the preset covalent bond length threshold (e.g., CS bond ≤ 1.90 Å, SO bond ≤ 1.55 Å).

[0119] (2) Use graph theory algorithms to identify all interconnected atomic groups (connected components);

[0120] (3) Based on precise stoichiometry and topological rules, namely that the atomic group must strictly contain the atomic composition of C1S1O2F3 and form the characteristic linkage structure of CF3-SO2F, complete CF3SO2F molecules are screened out.

[0121] (4) Traverse all time frames and count their number, and finally plot the curve of the change of the number of CF3SO2F molecules over time.

[0122] like Figure 3 As shown in figure a, the decomposition behavior of CF3SO2F exhibits a strong temperature dependence: at 1400 K, CF3SO2F decomposes slowly, with a decomposition rate of only 7% after 1000 ps; when the temperature rises to 2200 K, the decomposition rate of CF3SO2F accelerates significantly, eventually reaching a decomposition rate of 77%; and under the extreme condition of 3200 K, the decomposition of CF3SO2F is extremely rapid, with a decomposition rate as high as 94% within 1000 ps. This indicates that it decomposes rapidly in an electric arc channel.

[0123] like Figure 3 As shown in b, the decomposition behavior of CF3SO2F exhibits a strong positive correlation with the environmental system pressure. Under conditions of 2200 K and a 12% mixing ratio, at a pressure of 0.1 MPa, the decomposition rate is 77% within 1000 ps; when the pressure increases to 0.3 MPa, the decomposition rate rapidly increases to 87%; and when the pressure reaches 0.5 MPa, the decomposition is extremely complete, with a decomposition rate as high as 94%. It is evident that increasing the environmental system pressure dramatically accelerates the decomposition reaction kinetics by significantly increasing the molecular collision frequency.

[0124] like Figure 3As shown in Figure c, the decomposition behavior of CF3SO2F exhibits a significant proportion-dependent characteristic. Under conditions of 2200 K and 0.1 MPa, when the proportion of CF3SO2F is 12%, the decomposition rate reaches 77% within 1000 ps; when the proportion of CF3SO2F increases to 14%, the final decomposition rate decreases to 74%; and when the proportion of CF3SO2F increases to 20%, the decomposition rate further decreases to 64%. This indicates that, under the same operating conditions, a higher mixing proportion of CF3SO2F actually inhibits its own decomposition process.

[0125] At 1400 K, the main products are CF4 and SO2. When the temperature rises to 2200 K, these primary products will react further to generate secondary products such as SO, CF3, and CF2. At 3200 K, the system becomes extremely complex, producing a variety of small molecules and free radicals, including C, S, F, and O atoms.

[0126] By tracing the evolution of specific chemical bonds, key decomposition mechanisms were revealed. For example... Figure 5 As shown, radial distribution function (RDF) analysis reveals a significant decrease in the intensity of characteristic peaks representing SF, CS, and SO bonds as the temperature increases from 1400 K to 3200 K, statistically confirming the stepwise decomposition of the CF3SO2F molecule. Two main initial pathways exist: one is a concerted reaction, where the CS bond breaks simultaneously with the transfer of F atoms, directly generating CF4 and SO2; the other, at higher temperatures, involves the direct homolytic cleavage of the CS bond to generate highly reactive CF3· and ·SO2F radicals, which further initiate chain reactions.

[0127] The above detailed embodiments describe the implementation of the present invention; however, the present invention is not limited to the specific details described in the above embodiments. Within the scope of the claims and technical concept of the present invention, various simple modifications and changes can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

Claims

1. A method for simulating the thermal decomposition process of insulating gases based on deep learning potential functions, characterized in that, Includes the following steps: Ab initio molecular dynamics based on density functional theory was used to simulate the target insulating gas system at predetermined temperatures. Atomic coordinates, total energy corresponding to each atomic configuration, force on each atom, and virial tensor were collected to form an initial dataset, which was then divided into training, validation, and test sets. An initial deep learning potential function model is constructed, and the initial deep learning potential function model is trained, validated, and tested using the training set, validation set, and test set. Finally, the optimal deep learning potential function model is obtained. By using a deep learning potential function model, molecular dynamics simulations of the thermal decomposition process of an unknown insulating gas are performed to obtain the molecular dynamics simulation trajectory and analyze the results.

2. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 1, characterized in that, The method for obtaining and preprocessing the initial dataset is as follows: An initial unit cell is constructed, and a molecular dynamics simulation based on quantum mechanics is performed. All evolved configurations and their physical properties are recorded to obtain the original output simulation trajectory. All configurations and their physical properties include atomic coordinates, the total energy corresponding to each frame of atomic configuration, the force on each atom, and the virial tensor. Several frames of discretized, static atomic configurations and their full set of quantum mechanical precision physical quantity labels are sampled from the simulated trajectory of the original output results. The total energy corresponding to each frame of atomic configuration, the force on each atom, and the virial tensor are extracted to form the initial dataset.

3. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 2, characterized in that, The initial unit cell contains several target insulating gases and several buffer gases.

4. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 2, characterized in that, The method for molecular dynamics simulation based on quantum mechanics is as follows: AIMD calculations were performed on the initial unit cell using VASP, and the simulation was conducted under the NVT ensemble. The temperature was controlled using a Nose-Hoover thermal bath, with the temperature range set to 300-3200 K. Simulations were run for 20 ps at each temperature point, with a time step of 0.5 fs. The forces acting on atoms include the interaction between ions and electrons, as well as the interaction between electrons; the interaction between ions and electrons is described using the projected fused wave pseudopotential, and the interaction between electrons is handled using the PBE functional; the cutoff energy of the plane wave basis set is set to 520 eV. The convergence criteria for energy and the forces acting on atoms are set to 10. -6 eV and 0.01 eV / Å.

5. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 1, characterized in that, The initial dataset was divided into training set, validation set, and test set in proportions of 60%, 20%, and 20%, respectively.

6. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 1, characterized in that, The method for constructing the initial deep learning potential function model is as follows: The initial deep learning potential function model was built and trained using the DeePMD-kit toolkit; the "se_e2_a" descriptor type was selected and the cutoff radius was set to 10.00 Å; the descriptor embedding network adopted a three-layer structure with 25, 50 and 100 neurons respectively; the fitting network adopted a three-layer structure with 250 neurons in each layer. Training employed an exponentially decaying learning rate strategy, initially set at 0.001 and eventually reduced to 3.51 × 10⁻⁶. -8 ; Assign loss function weights to the total energy corresponding to each frame's atomic configuration, the force acting on each atom, and the virial tensor; Based on the total energy loss, atomic force loss, and virial tensor loss values ​​calculated for each frame of atomic configuration on the validation set, the weights are adjusted to ensure that the weighted loss terms are on the same order of magnitude until training converges.

7. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 6, characterized in that, The initial ratio of the total energy, the force on each atom, and the loss weight coefficient of the virial tensor corresponding to each frame atomic configuration is set to (0.5-2):(500-2000):(0.01-0.2).

8. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 1, characterized in that, The method for training, validating, and testing the initial deep learning potential function model using the training set, and finally outputting the optimal deep learning potential function model, is as follows: Using the training set and the same deep neural network architecture, multiple deep learning potential function models with identical structures are trained in parallel by independently assigning different random initial weight parameters to each deep learning potential function model, thus forming a deep learning potential function model set. Calculate the standard deviation of the deep learning potential function model's prediction of the forces acting on atoms in each new configuration, and select configurations with standard deviations higher than a preset threshold. First-principles calculations were used to compute the selected configurations to expand the dataset; Exploratory molecular dynamics simulations were performed in LAMMPS using a set of deep learning potential function models. The maximum standard deviation ε of the forces on atoms predicted by the set of deep learning potential function models was used as an uncertainty measure to screen out molecular configurations with 0.05 eV / Å ≤ ε < 0.50 eV / Å. Accurate DFT calculations were performed using VASP and these configurations were added to the training set. The training continues in a loop until the training set can fully describe all relevant reaction paths and configuration space, at which point the loop training terminates. The deep learning potential function model after the termination of the loop training is evaluated using the test set, and the optimal deep learning potential function model is output.

9. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 8, characterized in that, The method for expanding the dataset by using first-principles calculations on the selected configurations is as follows: VASP was used, based on DFT, to generate the same pseudopotential, functional, and cutoff energy parameters as the initial data; High-precision single-point energy calculations are performed on the selected atomic configurations to obtain accurate values ​​for their total energy, atomic force, and virial tensor. The configuration and its corresponding physical quantity are added as a new data sample to the training set.

10. The method for simulating the thermal decomposition process of insulating gas based on a deep learning potential function according to claim 8, characterized in that, The method for determining whether a training set can adequately describe all relevant reaction pathways and configuration space is as follows: When the maximum standard deviation ε of the forces on atoms predicted by the deep learning potential function model set is consistently lower than the preset threshold, it is considered that the current training set can fully cover all relevant reaction paths and configuration spaces of the target system under simulated conditions, and the active learning loop can be terminated.

Images