A Deep Learning-Based Method for Molecule Screening of Environmentally Friendly Insulating Gases
By using a deep learning-based method to screen gas molecules, the problem of time-consuming and labor-intensive screening of insulating gases in existing technologies has been solved. This method enables rapid and accurate gas molecule evaluation and screening, provides a new approach for environmentally friendly alternative gases, reduces experimental costs, and improves screening efficiency.
Patent Information
- Application Number
- CN202311676067.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-01
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-12-01
AI Technical Summary
Existing methods for screening insulating gas molecules mainly rely on experimental trial and error, which is time-consuming and labor-intensive. Furthermore, traditional machine learning algorithms are not good at problems with high-dimensional input variables or complex classification problems, making it difficult to find environmentally friendly alternative gases with comprehensive performance comparable to SF6.
A deep learning-based approach was adopted to establish a basic library of electrical and environmental properties of gas molecules. A data matrix was generated by molecular descriptors, and a deep learning model was built for training and testing. New small molecular structures were generated by combining functional group substitution methods, potential insulating alternative gases were screened, and their performance indicators were evaluated by the deep learning model.
This approach enables rapid and accurate evaluation of the key electrical and environmental performance of gas molecules, reduces the time cost of experimental trial and error, improves screening efficiency, provides new ideas for the research of novel insulating gas molecules, reduces large-scale experimental testing, and shortens the R&D cycle.
Smart Images

Figure CN117669342B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical insulation medium technology, specifically to an environmentally friendly insulating gas molecule screening method based on deep learning. Background Technology
[0002] Sulfur hexafluoride (SF6) was initially used as a fire extinguishing gas. Due to its excellent electrical properties and good chemical stability, it is currently the best gaseous insulating medium used in various high-voltage power equipment. However, SF6 has an extremely strong greenhouse effect, with a global warming potential (GWP) 24,300 times that of CO2. It is one of the greenhouse gases whose emissions are explicitly restricted in international conventions such as the Kyoto East Protocol and the Paris Agreement. The latest amendments to the EU's greenhouse gas regulations have clearly defined a timetable for the SF6 ban, stipulating that medium and high-voltage equipment must not use insulating media with a GWP greater than 10. The development of environmentally friendly alternative insulating media is imperative.
[0003] The development of electrical insulating gases is exceptionally difficult, requiring the simultaneous satisfaction of high electrical strength, low liquefaction temperature, low greenhouse effect, and high chemical stability. For over half a century, despite arduous experimental and theoretical explorations, no environmentally friendly alternative gas has been found that fully meets or surpasses SF6, and no substantial breakthrough has been achieved.
[0004] For the development of new insulating gas molecules, the experimental trial-and-error method is the most primitive and direct way to find target molecules. From proposing hypotheses to experimental verification, it is necessary to continuously iterate and get closer to the target molecule. Each verification requires tens of millions of yuan in project support. This method is time-consuming and labor-intensive, and so far, no environmentally friendly insulating gas molecule with comprehensive performance comparable to SF6 has been found.
[0005] In recent years, with the rapid development of computing technology, extracting knowledge from big data has become a mainstream form of scientific development. Machine learning has become a widely used and successful method. Traditional machine learning algorithms mainly include random forests, support vector machines, k-nearest neighbors, Naive Bayes, logistic regression, and DBSCAN. However, for finding environmentally friendly alternatives to SF6, a large amount of molecular structure information is needed as input. These traditional machine learning methods are not good at problems with high-dimensional input variables or complex classification problems. Summary of the Invention
[0006] The technical problem to be solved by this invention is that the main problem with existing insulating gas molecule prediction technologies is that most of them adopt experimental trial and error methods, which are time-consuming and labor-intensive. Traditional machine learning algorithms are not good at problems with high-dimensional input variables or complex classification problems.
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A deep learning-based method for screening environmentally friendly insulating gas molecules includes the following steps: S100, establish a basic library of known electrical and environmental properties of gas molecules; the performance indicators of the electrical and environmental properties include: insulation performance, liquefaction temperature and GWP; S200, represent the gas molecules in the basic library using molecular descriptors to generate a molecular data matrix; and divide the molecular data matrix into a training library and a test library; S300, Build a deep learning model. The training library is used to train the deep learning model. The expert system in the trained deep learning model is used to evaluate the performance index of gas molecules in the test library and compare it with the true value to evaluate the prediction accuracy. Continuously update the parameters of the deep learning model until the evaluation prediction accuracy reaches the prediction standard. The deep learning model at this time is taken as the best deep learning model. S400, gas molecules in the basic library are used to generate new small molecule structures in batches through functional group substitution method to form a gas molecule target library; S500: The optimal deep learning model is used to screen the gas molecules in the gas molecule target library, a comprehensive screening performance index is set, the preliminary screening results are sorted by the comprehensive screening performance index, an evaluation boundary is set, and potential insulating alternative gases within the evaluation boundary are output. S600: The potential insulating alternative gases are evaluated for their performance indicators, sorted again using the comprehensive screening performance indicators, and an evaluation line is set to output high-performance potential insulating alternative gases that are within the evaluation line.
[0008] Advantages: This invention enables rapid and accurate evaluation of key electrical and environmental performance indicators of gas molecules, which can greatly reduce the time cost of experimental trial and error methods, improve the efficiency of screening, and provide new ideas for the current research on novel insulating gas molecules.
[0009] This invention provides a method based on a combination of big data and deep learning algorithms. Through high-throughput search, it predicts novel environmentally friendly alternative gases for SF6 with superior performance in various aspects, offering new insights for the exploration of SF6 environmentally friendly alternative gases. This method can reduce large-scale experimental testing and significantly accelerate the development cycle of new molecules. However, this example is limited to providing theoretical support for achieving environmentally friendly substitution of SF6 and reducing the time cost of experimental trial and error. The final predicted gas molecules still need experimental verification.
[0010] In one embodiment of the present invention, the structure type of the gas molecules in the base library is generated in the following manner: a) Research and search literature, technical reports and patents on insulating gas research and development over the past half-century, and collect relevant data on insulating molecules; b) Search literature, molecular physicochemical property databases and physical chemistry handbooks to collect electrical and environmentally relevant physicochemical properties of small molecules, regardless of whether they can be used as insulating gases; c) For typical gas molecule types that lack relevant electrical and environmental protection physicochemical property data, high-precision quantum chemical and molecular dynamics simulation methods are used for evaluation.
[0011] In one embodiment of the present invention, gas molecules lacking liquefaction temperature are obtained through the following steps: S111, establish an amorphous unit cell for multiple gas molecules to obtain the initial gas crystal structure; S112, Based on density functional theory, optimize the bonding parameters and unit cell parameters of the initial gas crystal structure to obtain a stable molecular mechanical environment gas crystal structure; S113, using the Compass II force field, the initial gas crystal structure is energy initialized to obtain a stable force field energy environment gas crystal structure; S114, calculate the average interaction force and density of each single molecule in the molecular mechanical environment gas crystal structure and the force field energy environment gas crystal structure, respectively; S115, compare the average interaction forces and densities of the molecular mechanical environment gas crystal structure and the force field energy environment gas crystal structure respectively to obtain the initial liquefaction temperature prediction model; S116, continuously adjust the force field parameters of the Compass II force field of the initial gas crystal structure in the initial liquefaction temperature prediction model, update the initial liquefaction temperature prediction model, and obtain the best-fitting liquefaction temperature prediction model. S117, the amorphous unit cells of multiple gas molecules to be tested are introduced into the optimal liquefaction temperature prediction model to perform heating and cooling simulations, obtain the liquefaction temperature curve, and obtain the liquefaction temperature range based on the liquefaction temperature curve.
[0012] In one embodiment of the present invention, obtaining the optimal liquefaction temperature prediction model includes: updating the initial liquefaction temperature prediction model once for each adjustment of the force field parameters of the initial gas crystal structure, and performing an energy initialization on the initial gas crystal structure once for each update, and calculating the corresponding average interaction force and density, until the average interaction force and density are close to the average interaction force and density of the gas crystal structure in the molecular mechanical environment, and then taking the fitted liquefaction temperature prediction model at this time as the optimal liquefaction temperature prediction model.
[0013] In one embodiment of the present invention, the following steps are included when performing a temperature rise simulation: The optimal liquefaction temperature prediction model is initialized with energy, and the amorphous unit cells of multiple gas molecules to be tested are brought into the optimal liquefaction temperature prediction model after energy initialization. Under a canonical ensemble, the initial temperature and heating step size are set to construct a dynamic simulation environment; At the initial temperature, for each heating step increase, a molecular dynamics simulation is performed on the environment of the optimal liquefaction temperature prediction model to obtain the stable structure at this temperature, and the density of the environment of the optimal liquefaction temperature prediction model at this temperature is recorded. Repeat the previous step until the liquid system of the gas molecules to be tested boils and completely vaporizes. The density mutation point during the gradual heating process will be determined, and the first mutation temperature at this point will be recorded. The cooling simulation includes the following steps: Set the cooling step size, and gradually cool down after the gas molecules to be measured have completely vaporized; For each cooling step reduced, a molecular dynamics simulation is performed on the environment in which the optimal liquefaction temperature prediction model is located, and the stable structure at this temperature is obtained, and the density of the environment in which the optimal liquefaction temperature prediction model is located at this temperature is recorded. Repeat the previous step until the gaseous system of the gas molecules to be tested is completely liquefied, determine the density abrupt change point during the gradual cooling process, and record the second abrupt change temperature at this point. The density data obtained during the heating simulation and the density data obtained during the cooling simulation are used to plot a liquefaction temperature curve; in the liquefaction temperature curve, the range between the first abrupt change temperature and the second abrupt change temperature is the liquefaction temperature range.
[0014] In one embodiment of the present invention, gas molecules lacking GWP are obtained through the following steps: S121, Determine the optimized basis set, and perform structural optimization on gas molecules and water molecules based on the optimized basis set to obtain the gas molecule-water molecule optimized structure and the gas molecule optimized structure with the lowest energy and the most stable gas molecule-water molecule optimized structure. S122, the gas molecules and water molecules in the gas molecule-water molecule optimized structure are mixed in proportion to construct an initial gas-water eutectic structure; S123, Optimize the initial gas-water eutectic structure under the lowest total energy space group according to first principles to obtain the optimal gas-water eutectic structure; S124. Based on molecular dynamics and density functional methods, the reaction path of the optimal gas-water eutectic structure is simulated to obtain all decomposition paths of the optimal gas-water eutectic structure. S125, obtain the rate constants of all decomposition paths and compare them, and select the one with the smallest rate constant as the rate-determining step of the optimal gas-water eutectic structure reaction; S126, obtain the reaction rate constant of the rate-determining step of the optimal gas-water eutectic structure reaction; and obtain the atmospheric lifetime of the gas based on the reaction rate constant; S127, Obtain the infrared spectrum of the optimized structure of the gas molecules, obtain the band based on the infrared spectrum, and obtain the spectral-resolved radiative forcing per unit cross-section based on the band. S128, Based on the wavelength and the spectral resolution of the radiative forcing, obtain the radiative efficiency of the optimized gas molecule structure; S129, Based on the atmospheric lifetime and the radiation efficiency, obtain the global warming potential value of the gas.
[0015] In one embodiment of the present invention, the construction of the initial gas-water eutectic structure includes: obtaining the interaction force between the optimized structure of the gas molecules and the water molecules under a global force field; and based on the Monte Carlo algorithm, predicting the combination and stacking of all interactions under the C2 / c space group, P1 space group, P21 space group, P2 / c space group and P212121 space group respectively, obtaining multiple initial gas-water eutectic structures, and sorting the total energy of the initial gas-water eutectic structures under each space group to obtain the initial gas-water eutectic structure under the space group with the lowest total energy.
[0016] In one embodiment of the present invention, an ultrasoft pseudopotential is used to bring the initial gas-water eutectic structure under the lowest total energy space group into force equilibrium, thereby obtaining the initial gas-water eutectic structure with the lowest energy under the lowest total energy space group, which is taken as the optimal insulating gas-water eutectic structure.
[0017] In one embodiment of the present invention, obtaining all decomposition paths of the optimal gas-water eutectic structure includes: The energy of the optimal gas-water eutectic structure is initialized to ensure that the environment in which the optimal gas-water eutectic structure is located is in equilibrium before heating. Under a canonical ensemble, the initial temperature, simulation steps, and temperature intervals are set. The initial temperature is used as the first temperature for kinetic simulation. Then, the environment in which the optimal gas-water eutectic structure is located is gradually heated at a determined temperature interval until all decomposition pathways of hydroxyl radicals in the gas and water molecules are obtained.
[0018] In one embodiment of the present invention, the formation of the gas molecule target library includes: Input gas molecules from the base library; To establish substitution rules and generate new small molecule structures, it is necessary to satisfy Lewis's octet rule and hybrid orbital theory. The molecular structure can be replaced in batches by using the functional group substitution method, while satisfying the substitution rules. A new gas molecule structure conforming to the substitution rule is output to form the gas-forming molecule target library; and all gas molecules in the gas-forming molecule target library adopt B3LYP / 6-311G. The method performs structural optimization to determine its ground-state geometry.
[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes for the first time a method for fitting molecular force fields by comparing molecular structure and energy data. The fitting parameters include bonding and non-bonding parameters, and the complete force field parameters obtained can be used to predict liquefaction temperatures. It is the first time that the force field parameters fitted by individual molecules have been used to calculate the liquefaction temperature of mixed gases, which is cost-effective and solves the problem that the liquefaction temperature of insulating gases is difficult to achieve experimentally. It can accurately predict the liquefaction temperature of gases and also has the advantages of simple process, stable cost, and high computational efficiency, providing a valuable reference for the search for insulating gases.
[0020] For calculating radiative efficiency, this application considers the effect of polarization bonds in the molecule and uses an optimized basis set to calculate molecular bond distances, obtaining more accurate wavenumbers. For atmospheric lifetime calculation, the reaction pathways between the insulating gas and atmospheric components are realistically simulated based on the Car-Parrinello molecular dynamics method, and the reaction rate constants of the rate-determining step are calculated to obtain the atmospheric lifetime value. Finally, the obtained radiative efficiency and atmospheric lifetime values are substituted into the GWP calculation formula to obtain the GWP of the insulating gas, which has the advantages of high reliability, high computational efficiency, high accuracy, and good repeatability.
[0021] Molecular dynamics, based on first-principles calculations, starts from fundamental laws of physics and derives and calculates without relying on assumptions or empirical fitting. It directly analyzes the structure of molecules to obtain the behavioral patterns of molecular motion, simulating the dynamic evolution of the system. This allows for simpler ab initio calculations of chemical reactions, realistically simulating all reaction paths. Density functional theory (DFT) describes the state of a system using electron density as the fundamental variable. The total energy and ground-state properties of the reaction system can be obtained through the electron density distribution function in any region. This invention, for the first time, employs CPMD and DFT methods to simulate the real reaction path of insulating gas molecules in the atmosphere, starting from the structure of the insulating gas molecules, and calculates the rate constant for each path, thereby determining the rate-determining step of the reaction. This method reduces the risks and time costs associated with experimental methods, avoids the high uncertainty of traditional transition state methods, and features cost stability, simplicity, and high accuracy.
[0022] Traditional transition state theory directly guesses the possible pathways of a chemical reaction, then calculates the activation energy based on the guessed pathways to determine the rate-determining step. This method is highly subjective and yields significant uncertainty. This invention starts directly from molecular structure, enabling simpler ab initio calculations of chemical reactions and realistically simulating all reaction pathways. Compared to experimental methods and traditional transition state theory, this invention can more accurately simulate all chemical reaction pathways, offering advantages such as cost stability, high reliability, and high precision.
[0023] The method mainly employs computational chemistry, starting from the molecular structure and considering the characteristics of each atom, to calculate the reaction rate of the rate-determining step of the reaction between the insulating gas and atmospheric components, thereby obtaining the atmospheric lifetime value.
[0024] The results from the comparative examples and embodiments show that the GWP obtained by the present invention is closer to the literature value, avoids over-reliance on experience, and has a simple process, high computational efficiency, and stable cost. Attached Figure Description
[0025] Figure 1 This is a flowchart of an environmentally friendly insulating gas molecule screening method based on deep learning, according to an embodiment of the present invention.
[0026] Figure 2 This is a schematic diagram of the C4F7N crystal structure according to an embodiment of the present invention.
[0027] Figure 3 This is a schematic diagram of the liquefaction temperature curve in an embodiment of the present invention.
[0028] Figure 4 This is a schematic diagram of the insulating gas-water eutectic structure according to an embodiment of the present invention.
[0029] Figures 5 to 7 These are the three decomposition paths of C4F7N and -OH in this embodiment of the invention.
[0030] Figure 8 The infrared spectrum of the optimized gas molecule structure in an embodiment of the present invention. Detailed Implementation
[0031] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.
[0032] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0033] Please see Figure 1 As shown, this invention provides a method for screening environmentally friendly insulating gas molecules based on deep learning, comprising the following steps: S100, establish a basic library of known electrical and environmental properties of gas molecules; the performance indicators of the electrical and environmental properties include: insulation performance, liquefaction temperature and GWP.
[0034] In this embodiment, the base library contains as many molecular structure types as possible to achieve full coverage of "big data" samples. It is generated primarily through the following methods: a) Research and search literature, technical reports and patents on insulating gas research and development over the past half-century, and collect relevant insulating molecule data.
[0035] b) Search literature, molecular physicochemical property databases, and physical chemistry handbooks to collect electrical and environmentally relevant physicochemical properties of small molecules, regardless of whether they can be used as insulating gases.
[0036] c) For typical gas molecule types that lack electrical and environmentally relevant physicochemical property data, high-precision quantum chemical and molecular dynamics simulation methods are used for evaluation.
[0037] Please see as follows Figures 1 to 3 As shown, in one embodiment of the present invention, in method c), for gas molecules lacking liquefaction temperature, the following steps are performed: S111, establish an amorphous unit cell for multiple gas molecules to obtain the initial gas crystal structure.
[0038] Please see Figure 2 As shown, in this embodiment, C4F7N is used as an example. In the Amorphous Cell, the ambient temperature is set to 268K and the density is set to 1.5 g / cm³. 3 The initial environment for constructing the initial gas crystal structure is obtained, and then 100 amorphous unit cells of C4F7N molecules are built in the initial environment to obtain the initial C4F7N crystal structure.
[0039] S112, based on density functional theory, optimize the bonding parameters and unit cell parameters of the initial gas crystal structure to obtain a stable molecular mechanical environment gas crystal structure.
[0040] S113, using the Compass II force field, the initial gas crystal structure is energy initialized to obtain a stable force field energy environment gas crystal structure.
[0041] In this embodiment, since the Compass II force field is a high-precision force field capable of performing atomic-scale simulation studies on the system, the Compass II force field is selected to initialize the energy of the initial C4F7N crystal structure, that is, to perform initial operation on the disordered crystal structure, so that the energy is reduced and vibrates within a reasonable range, thus obtaining a reasonable structure that can be used for molecular dynamics simulation.
[0042] S114, calculate the average interaction force and density of each single molecule in the molecular mechanical environment gas crystal structure and the force field energy environment gas crystal structure, respectively.
[0043] Specifically, the average interaction force is obtained using the following formula: ; In the formula, For the average interaction force, The total energy of the gas crystal. The energy of a single molecule, This represents the number of molecules.
[0044] The density can be obtained using the following formula: ρ=m / V; In the formula, ρ is the density of the crystal structure, m is the mass of the crystal in the crystal structure, and V is the volume of the crystal structure.
[0045] Boiling of a gas is the process of the disappearance of intermolecular forces, while liquefaction is the regeneration of intermolecular forces. Therefore, liquefaction temperature is closely related to intermolecular forces. Intermolecular forces depend on the molecular structure; different structures result in different molecular masses. In addition, the volume difference is determined by the intermolecular distance. Density is also a very important evaluation parameter. Therefore, it is necessary to obtain the average interaction force of each single molecule in the crystal structure and the density of the crystal structure.
[0046] S115, compare the average interaction forces and densities of the molecular mechanical environment gas crystal structure and the force field energy environment gas crystal structure respectively to obtain the initial liquefaction temperature prediction model.
[0047] In this embodiment, the average interaction force of the gas crystal structure in the molecular mechanical environment is obtained through the average interaction force and density formula. And density ρ1, and the average interaction force of the gas crystal structure in the force field energy environment. By subtracting the average interaction force and density ρ2 from the average interaction force and density, an initial liquefaction temperature prediction model is obtained.
[0048] S116, continuously adjust the force field parameters of the Compass II force field of the initial gas crystal structure in the initial liquefaction temperature prediction model, update the initial liquefaction temperature prediction model, and obtain the best-fitting liquefaction temperature prediction model.
[0049] In this embodiment, obtaining the optimal liquefaction temperature prediction model includes: updating the initial liquefaction temperature prediction model once for each adjustment of the force field parameters of the initial gas crystal structure, and performing an energy initialization on the initial gas crystal structure once for each update, and calculating the corresponding average interaction force and density, until the average interaction force and density are close to the average interaction force and density of the gas crystal structure in the molecular mechanical environment, and then using the fitted liquefaction temperature prediction model at this time as the optimal liquefaction temperature prediction model.
[0050] Specifically, the similarity criterion is that the average interaction force and density calculated from the updated gas crystal structure have an error of less than 2% compared with the average interaction force and density of the gas crystal structure in the molecular mechanical environment.
[0051] The force field parameters of the Compass II force field in this step include bonding parameters and non-bonding parameters. The bonding parameters include bond length, bond angle, and dihedral angle, while the non-bonding parameters include molecular charge and van der Waals force.
[0052] Molecular dynamics, based on Newton's laws of motion, effectively records the trajectories of molecules within a system. The accuracy of the molecular force field energy directly determines the success or failure of computational simulations; therefore, obtaining highly accurate force field parameters is crucial for simulation. For a chemical reaction, if a suitable force field can be fitted, the actual reaction can be simulated under specific temperature and pressure conditions. The molecular force field is a set of parameterized empirical potential functions describing the interactions between atoms in a system. It mainly consists of two parts: bonding parameters describing intramolecular bonding interactions and non-bonding parameters describing intermolecular interactions, which describe the interactions between unbonded atoms or groups. Based on these parameters, a new force field can be constructed for a specific system like an insulating gas. Using molecular dynamics methods, the gradual cooling process of the gas can be simulated to observe the distances, densities, and forces between molecules, thereby obtaining the relationship between temperature and density, and ultimately determining the liquefaction temperature of the insulating gas. Therefore, this example predicts the liquefaction temperature of a gas based on molecular force field fitting.
[0053] S117, the amorphous unit cells of multiple gas molecules to be tested are introduced into the optimal liquefaction temperature prediction model to perform heating and cooling simulations, obtain the liquefaction temperature curve, and obtain the liquefaction temperature range based on the liquefaction temperature curve.
[0054] The following steps are included when performing a temperature rise simulation: The optimal liquefaction temperature prediction model is initialized with energy, and the amorphous unit cells of multiple gas molecules to be tested are introduced into the energy-initialized optimal liquefaction temperature prediction model.
[0055] Under a canonical ensemble, the initial temperature and heating step size are set to construct a dynamic simulation environment.
[0056] At the initial temperature, for each heating step, a molecular dynamics simulation is performed on the environment of the optimal liquefaction temperature prediction model to obtain the stable structure at this temperature, and the density of the environment of the optimal liquefaction temperature prediction model at this temperature is recorded.
[0057] Repeat the previous step until the liquid system of the gas molecules to be tested boils and completely vaporizes. This will determine the density abrupt change point during the gradual heating process, and record the first abrupt change temperature at this point.
[0058] In this embodiment, the amorphous unit cell of 1000 C4F7N molecules is incorporated into the optimal liquefaction temperature prediction model after energy initialization.
[0059] The cooling simulation includes the following steps: Set the cooling step size, and gradually cool down after the gas molecules to be measured have completely vaporized.
[0060] For each cooling step decrease, a molecular dynamics simulation is performed on the environment in which the optimal liquefaction temperature prediction model is located, and the stable structure at this temperature is obtained. The density of the environment in which the optimal liquefaction temperature prediction model is located at this temperature is also recorded.
[0061] Repeat the previous step until the gaseous system of the gas molecules to be tested is completely liquefied. Determine the density abrupt change point during the gradual cooling process and record the second abrupt change temperature at this point.
[0062] Specifically, the initial temperature is 260K, the temperature increment is 1K, and the duration of each temperature setting during the simulation is 1000ps, with a total simulation duration of 11ns. The cooling increment is equal to the heating increment.
[0063] The density data obtained during the heating simulation and the density data obtained during the cooling simulation are used to plot a liquefaction temperature curve. In the liquefaction temperature curve, the range between the first abrupt change temperature and the second abrupt change temperature is the liquefaction temperature range. Please refer to [link / reference]. Figure 3 Therefore, the predicted liquefaction temperature of C4F7N molecules is between 266K and 268K, while the liquefaction temperature of C4F7N molecules obtained from the literature is 268K, which means that the prediction accuracy of the present invention is high.
[0064] Please see as follows Figure 1 , Figures 4 to 8As shown, in one embodiment of the present invention, gas molecules lacking GWP are obtained through the following steps: S121, Determine the optimized basis set, and perform structural optimization on gas molecules and water molecules based on the optimized basis set to obtain the gas molecule-water molecule optimized structure and the gas molecule optimized structure with the lowest energy and the most stable structure.
[0065] Please see Figure 4 As shown, in this embodiment, the insulating gas is C4F7N. To determine the optimal basis set for optimizing the C4F7N molecular structure, this example compares the selected bond distances (Å) of the C4F7N molecule under different functional method basis sets, as shown in Table 1.
[0066] Table 1. Comparison of selected bond distances (Å) of C4F7N under different functional method basis sets.
[0067] Based on a comparison of bond distances for the insulating gas under different functional method basis sets, the basis set with high bond distance accuracy was selected as the optimization basis set. Table 1 shows that, as a benchmark, under the same basis set, the BP86 / def2tzvp basis set method takes the shortest time in structural description, but has the weakest accuracy. The results of the B3LYP / def2tzvp basis set method are even weaker than those of the BP86 / def2tzvp basis set method. Compared with the B3LYP / def2tzvp basis set method, the M06-2X / def2tzvp basis set method improves bond distance accuracy, and its time error is relatively negligible. Therefore, the M06-2X / def2tzvp basis set method was selected for optimization.
[0068] The degradation of greenhouse gases in the atmosphere mainly depends on their reaction with hydroxyl radicals (-OH). To find the rate-determining step in the reaction pathway between greenhouse gases and -OH, considering the addition of -OH is a prerequisite for obtaining a reasonable decomposition pathway. In this example, H₂O is used as the source of -OH generation.
[0069] S122, the gas molecules and water molecules in the optimized gas-water molecule structure are mixed in proportion to construct an initial gas-water eutectic structure.
[0070] In this embodiment, optimized C4F7N molecules and H2O molecules are mixed in a 1:1 ratio to construct an initial C4F7N-H2O eutectic structure. Specifically, constructing the initial gas-water eutectic structure includes: obtaining the interaction forces between the optimized gas molecule structure and the water molecules under a global force field. Based on the Monte Carlo algorithm, the combined stacking of all interactions under the C2 / c, P1, P21, P2 / c, and P212121 space groups is predicted to obtain multiple initial gas-water eutectic structures. The total energy of the initial gas-water eutectic structures under each space group is sorted, and the initial gas-water eutectic structure under the space group with the lowest total energy is obtained.
[0071] In this embodiment, a suitable force field is chosen to simulate the interaction between C4F7N and H2O molecules in the system. The universal force field (UFF) is applicable to all elements in the periodic table and can be applied to any atom and molecule. The potential energy E is represented as the sum of valence interactions or bonding interactions and non-bonding interactions.
[0072] E=E R +E θ +E +E ω +E vdw +E el Among them, E R E θ E E ω E vdw and E el These are the bond stretching energy, bond angle bending energy, dihedral angle torsion energy, bond angle out-of-plane bending energy, van der Waals energy, and electrostatic energy of the gas-water eutectic structure. Based on the six forces in the UFF force field in the above formulas, the interaction energy between C4F7N molecules and H2O molecules was calculated.
[0073] Based on this, using the Monte Carlo (MC) algorithm, the packing of gas-water eutectic structures in five space groups—C2 / c, P1, P21, P2 / c, and P212121—is predicted, yielding the minimum total energy, van der Waals energy, predicted number of structures, and crystal parameters.
[0074] In this study, C4F7N and H2O are defined as rigid rotors, meaning that during stacking based on the Monte Carlo algorithm, the internal structure of the rotor remains unchanged, acting only as a point mass, while the relative positions between the rotors change. Since most crystals belong to the five space groups C2 / c, P1, P21, P2 / c, and P212121, the possible stacking configurations of C4F7N and H2O in these five space groups were predicted, as shown in Table 2 below.
[0075] Table 2 Stacking of space group insulating gas-water eutectic structures
[0076] In Table 2, a, b, and c represent the three sets of edge lengths of the unit cell, and α, β, and γ represent the inter-edge angles of the unit cell. The tens of thousands of predicted eutectic structures were sorted by energy, and the C4F7N-H2O eutectic structure with the lowest energy was selected as the object of subsequent research.
[0077] S123, optimize the initial gas-water eutectic structure under the lowest total energy space group according to first principles to obtain the optimal gas-water eutectic structure.
[0078] In this embodiment, an ultrasoft pseudopotential is used to bring the initial gas-water eutectic structure under the lowest total energy space group into force equilibrium, thereby obtaining the initial gas-water eutectic structure with the lowest energy under the lowest total energy space group, which is taken as the optimal insulating gas-water eutectic structure.
[0079] S124. Based on molecular dynamics and density functional methods, the reaction path of the optimal gas-water eutectic structure is simulated to obtain all decomposition paths of the optimal gas-water eutectic structure.
[0080] Please see Figures 5 to 8 As shown, in this embodiment, all decomposition paths for obtaining the optimal gas-water eutectic structure include: S1241, initialize the energy of the optimal gas-water eutectic structure to ensure that the environment in which the optimal gas-water eutectic structure is located is in equilibrium before heating.
[0081] Specifically, the time step was first set to 0.1 fs, and the atomic force field was tested to converge to 10 per atom. -1 ~10 -2 au, perform energy initialization Molecular dynamics simulations were used to verify the thermodynamic stability of the system. Under a canonical ensemble, dynamic simulations were used to determine that the structure was stable at low temperatures, and then dynamic simulations were performed by gradually increasing the temperature.
[0082] S1242, under the canonical ensemble, set the initial temperature, simulation steps and temperature interval, use the initial temperature as the first temperature for kinetic simulation, and then gradually increase the temperature of the environment in which the optimal gas-water eutectic structure is located at a determined temperature interval until the complete decomposition path of hydroxyl radicals in the gas and water molecules is obtained.
[0083] Specifically, under a canonical ensemble, a kinetic simulation of 20,000 steps (2 ps) was performed with 500 K as the first temperature, followed by temperature increases of 500 K until reaching 3000 K. To accelerate the reaction process and obtain the complete decomposition trajectory more quickly, the number of simulation steps was reduced to 20,000-40,000 steps below 2000 K, while the simulation step size was longer at higher temperatures, ranging from 60,000 to 80,000 steps. The total simulation time was 26 ps (260,000 steps), obtaining the complete decomposition path of the optimal gas-water eutectic structure.
[0084] Based on the decomposition process, all reactions between C4F7N and -OH can be summarized into three decomposition pathways: The first decomposition path: In the initial stage of the reaction, H2O decomposes to generate -OH, and the chain reaction a involves -OH attacking the weakest C-CN bond to generate the active fragments HOCN and -CF(CF3)2.
[0085] The second decomposition path: chain reaction b is the lone pair electrons of the oxygen atom in the water molecule approaching the -CF(CF3)2 active segment lacking the central electron, in which the cleavage of the C-(CF3) bond is accompanied by the formation of the C-OH2 bond.
[0086] The third decomposition pathway: Due to the synergistic interaction between oxygen and the active segment of -CF(CF3)2, the reaction in which the H radical and the -CF(CF3)2 segment compete for F atoms is triggered, and finally -H and -F combine to form HF.
[0087] The three pathways represent the possible reaction mechanisms between C4F7N and -OH. Different triggering mechanisms require different rate-determining steps and transition states, and therefore the reaction barriers they overcome are also different.
[0088] S125, obtain the rate constants of all decomposition paths and compare them, and select the one with the smallest rate constant as the rate-determining step of the optimal gas-water eutectic structure reaction.
[0089] The rate constant is obtained using the following formula: ; In the formula, Expressed as a rate constant, Represented as the Arrhenius constant, It is expressed as the decomposition activation energy. Expressed as molar gas constant, It is expressed as thermodynamic temperature, and e is the base of the natural logarithm.
[0090] In this embodiment, the rate constants of the first decomposition path Path(a), the second decomposition path Path(b), and the third decomposition path Path(c) at 298 K are calculated to be 3.14 × 10⁻⁶. -15 6.26×10 -4 1.92×10 -8 The first decomposition path, Path(a), has the smallest rate constant, and is therefore determined to be the rate-determining step of the reaction.
[0091] S126, obtain the reaction rate constant of the rate-determining step of the optimal gas-water eutectic structure reaction; and obtain the atmospheric lifetime of the gas based on the reaction rate constant.
[0092] Based on the transition state theory (TST), the rate constant of the rate-determining step of the optimal gas-water eutectic structure reaction is calculated. Specifically, the reaction rate constant is obtained using the following formula: ; formula, The reaction rate constant is... Boltzmann constant, is Planck's constant. c 'Standard gas concentration' Expressed as molar gas constant, Expressed as thermodynamic temperature, Δ r ≠ G Θ For the standard molar activation Gibbs free energy, It is an exponential function.
[0093] The atmospheric lifetime of C4F7N was calculated using the atmospheric lifetime calculation formula proposed by ESCKewok and R. Atkinson. Specifically, the atmospheric lifetime of the insulating gas was obtained using the following formula: ; In the formula, Indicated as atmospheric lifetime, The reaction rate constant is... The concentration of -OH in the atmosphere is 1 × 10⁶ mol·cm⁻¹ -3 , The empirical parameter is set to 2.1. The calculated atmospheric lifetime of C4F7N is 21.2 years.
[0094] S127, Obtain the infrared spectrum of the optimized structure of the gas molecules, obtain the band based on the infrared spectrum, and obtain the spectral-resolved radiative forcing per unit cross-section based on the band.
[0095] In this embodiment, the optimized gas molecule structure can be automatically acquired in an optimized environment, such as... Figure 8 As shown. Specifically, the optimized environment is Gaussian 16.
[0096] Based on the peaks of the infrared spectrum, determine the wavelength band, and then look up the Wm per unit cross-section based on the wavelength band. 2 cm (cm) 2 molecule 1 ) 1 Spectral-resolved radiative forcing.
[0097] S128, Based on the wavelength and the spectral resolution radiative forcing, obtain the radiative efficiency of the optimized gas molecule structure.
[0098] Specifically, the radiation efficiency of the optimized gas molecule structure is obtained using the following formula: ; In the formula, For radiation efficiency, For spectrally resolved radiative forcing per unit cross-section, For bands, This represents the wavenumber increment, with a value of 1 cm. 1 , For band In wavenumber increment The absorption cross section with the average value taken above.
[0099] S129, Based on the atmospheric lifetime and the radiation efficiency, obtain the global warming potential value of the gas.
[0100] The global warming potential (GWP) of insulating gases is obtained using the following formula: ; In the formula, The time period is typically 100 years. The global warming potential (GWP) of C4F7N over 100 years is 9.17 × 10⁻¹⁴ W·m⁻²·yr⁻¹·kg⁻¹, according to the IPCC report in 2018. The final calculated GWP for C4F7N is 2153.
[0101] Comparative Example 1: GWP Calculation Method of C4F7N Based on Group Contribution Method The rate constant of the reaction between the C4F7N molecule and OH radicals in the atmosphere was determined using the group contribution method. Calculations are performed, in which, , This represents the rate at which the OH radical absorbs H atoms from the CH and OH bonds. This represents the reaction rate of OH radicals with unsaturated carbon-carbon double and triple bonds in the gas. This represents the rate of reaction between OH radicals and aromatic rings in the gas. This represents the reaction rate of OH radicals with functional groups containing N, P, and S elements in a gaseous atmosphere. The reaction rates are calculated based on the reaction rates of different functional groups in the molecule with -OH radicals. k OH ((CF3)2CFCN)=6.1× 10 16 cm 3 molecule 1 s 1 Using the atmospheric lifetime of CH3CCl3 as a reference, the atmospheric lifetime of C4F7N is estimated according to the following formula.
[0102] ; In the formula, The atmospheric lifetime of C4F7N calculated based on the group contribution method, The value is 6.03 × 10 15 cm 3 molecule 1 s 1 , for The atmospheric lifetime of C4F7N is 5 years, and the atmospheric lifetime of C4F7N calculated based on the group contribution method is about 49 years.
[0103] The wavelength range of the C4F7N molecular structure was obtained, and the spectral-resolved radiative forcing per unit cross-section was calculated based on this wavelength range. Then, using the radiative efficiency calculation formula, the radiative efficiency of the C4F7N molecular structure was determined to be 0.21 W·(m²). 2 ·ppbv) -1 The atmospheric lifetime and radiative efficiency of C4F7N, calculated using the global warming potential (GWP) formula, yield a GWP of 680 based on the group contribution method.
[0104] Comparative Example 2: GWP Calculation Method of C4F7N Based on Traditional Transition State Method Based on the fact that OH radicals can form complexes with C4F7N through hydrogen bonding, and that highly reactive OH radicals can catalyze the decomposition of C4F7N under atmospheric catalysis, a series of possible decomposition pathways of C4F7N under the action of OH radicals can be predicted. At the M062X / Aug-cc-pVTZ computational level, the reaction energy barriers and rate-determining step reaction rate constants of all possible pathways are calculated using traditional transition state theory. Then, the atmospheric lifetime of C4F7N is calculated based on the reaction rate constants. Using the method proposed by Pinnock to calculate radiative efficiency, the radiative efficiency of C4F7N is calculated to be 0.21. Finally, substituting the atmospheric lifetime and radiative efficiency values into the GWP formula, the atmospheric lifetime of C4F7N is calculated to be 47 years. Finally, substituting the radiative efficiency and atmospheric lifetime values into the GWP calculation formula, the GWP of C4F7N based on the traditional transition state method is 1597.
[0105] Please refer to Table 3. Compared with the invention, the GWP estimated by the present invention is closer to the experimental data in Comparative Examples 1 and 2.
[0106] Table 3 Comparison of GWP between Comparative Examples 1 and 2 and the estimated GWP of this invention
[0107] In one embodiment of the present invention, gas molecules lacking insulating strength are obtained through the following steps: using M06-2X / 6-311++G The method calculates the dipole moment of the molecule. and adiabatic ionization energy Then, the insulation strength value is calculated using the following formula. : ; in, , , For different correction coefficients, the values of molecules with known insulation strength can be fitted. In this embodiment, the method for obtaining the insulation strength of gas molecules is existing technology and will not be described in detail here.
[0108] S200, the gas molecules in the basic library are represented by molecular descriptors to generate a molecular data matrix; and the molecular data matrix is divided into a training library and a test library.
[0109] In this embodiment, the structure of gas molecules determines their properties, but computers cannot directly identify their molecular coordinates. It is necessary to extract the intrinsic topological information of their structure to generate a data matrix as a molecular descriptor for the deep learning model. Specifically, a small molecule descriptor generation program is written using Python: It takes the molecular coordinates of gas molecules from the base library as input, converts the molecular structure into various molecular descriptors as output, and presents the molecular descriptors in the form of a data matrix, thereby achieving batch generation of small molecule descriptors. The multiple molecular descriptors determine the accuracy of the input information. To balance efficiency and accuracy, a combination of various physical and chemical information molecular descriptors is used. Specifically, these multiple molecular descriptors include the Coulomb matrix, the CORINA descriptor, the MACCS fingerprint, and the Morgan fingerprint.
[0110] S300, Build a deep learning model. The training library is used to train the deep learning model. The expert system in the trained deep learning model is used to evaluate the performance indicators of gas molecules in the test library and compare them with the true values to evaluate the prediction accuracy. Continuously update the parameters of the deep learning model until the evaluation prediction accuracy reaches the prediction standard. The deep learning model at this time is taken as the best deep learning model.
[0111] In this embodiment, the deep learning model platform uses Python and is programmed under a Linux system. It mainly consists of three parts: 1) An input layer, which uses the molecular descriptor as input. 2) An intermediate hidden layer, which is the selected neural network model, determining the learning level of the deep learning model. A multi-layer fully connected DNN neural network model is used. The hidden layer determines the intelligence of the deep learning model. A simple network model results in fast training speed but difficulty in guaranteeing prediction accuracy; an overly complex network model results in prohibitive training costs. A suitable balance is found through continuous testing. 3) An output layer, which outputs performance indicators related to electrical and environmental protection properties. The built deep learning model is continuously tested using a test library and compared with real values. The mean absolute error (MAE) and coefficient of determination R are calculated. 2 To evaluate the performance of deep learning models, parameters are continuously adjusted to achieve optimal prediction accuracy.
[0112] S400, gas molecules in the basic library are used to generate new small molecule structures in batches through functional group substitution method to form a gas molecule target library.
[0113] In this embodiment, the gas molecule target library is based on an existing base library of gas molecules. A smart gas molecule generation program is written in Python, and a functional group substitution method is used to generate new small molecule structures in batches based on the original gas molecules. The original gas molecules are continuously rearranged and combined to form new gas molecule structures, and the rearrangement and combination of the original gas molecules conforms to the Lewis octet rule and hybridization theory. Specifically, the gas molecules in the base library are input; substitution rules are set, and the generated new small molecule structures must satisfy the Lewis octet rule and hybridization theory; the functional group substitution method is used to replace molecular structures in batches based on satisfying the substitution rules; new gas molecule structures that conform to the substitution rules are output, forming the gas molecule target library; and the gas molecules in the gas molecule target library all adopt B3LYP / 6-311G. The method performs structural optimization to determine the ground-state geometry. During batch generation, functional group substitutions are not scored; instead, an exhaustive method is used to generate as many molecular structures as possible.
[0114] S500: The optimal deep learning model is used to screen the gas molecules in the gas molecule target library, a comprehensive screening performance index is set, the preliminary screening results are sorted by the comprehensive screening performance index, an evaluation boundary is set, and potential insulating substitute gases within the evaluation boundary are output.
[0115] In this embodiment, when the optimal deep learning model filters gas molecules in the gas molecule target library, it first obtains the molecular coordinates of the gas molecules in the gas molecule target library, then converts the molecular coordinates of the gas molecules in the gas molecule target library into various molecular descriptors as output, and inputs them into the input layer of the optimal deep learning model.
[0116] The comprehensive performance indicators for screening are: insulation performance: 60%; liquefaction temperature: 20%; GWP: 20%. In this embodiment, the evaluation boundary is the top 500 gas molecules after sorting the preliminary screening results by comprehensive performance indicators. These top 500 gas molecules are considered potential insulating alternative gases and are output for further evaluation.
[0117] S600: The potential insulating alternative gases are evaluated for their performance indicators, sorted again using the comprehensive screening performance indicators, and an evaluation line is set to output high-performance potential insulating alternative gases that are within the evaluation line.
[0118] In this embodiment, the potential insulating alternative gases are evaluated for their performance indicators. The method used is consistent for obtaining gas molecules lacking liquefaction temperature, gas molecules lacking GWP, and gas molecules lacking insulating strength. The evaluation criteria are as follows: after evaluating the performance indicators of these 500 potential insulating alternative gases, the comprehensive performance indicators are ranked, and the top 5 potential insulating alternative gases are considered high-performance potential insulating alternative gases.
[0119] In this embodiment, for complex problems such as predicting the properties of gas molecules, deep learning can automatically find the characteristics of gas molecules by learning the features of data samples without manual design. Furthermore, by adding hidden layers to the neural network, deeper features can be extracted, significantly improving the model's predictive and classification capabilities.
[0120] Deep learning has made significant breakthroughs in recommendation systems, classification, object detection, speech recognition, image segmentation, and drug activity prediction. Common deep learning models are mainly divided into unsupervised and supervised learning models. In supervised learning, the computer learns from past data and applies the results to current data to predict future events. Unsupervised learning, on the other hand, learns independently using unclassified and unlabeled data. The machine must be able to classify the data without being provided with any prior information. For molecular information about insulating gases, unsupervised learning can break away from conventional learning methods, finding useful insights from the data, making it more suitable for predicting SF6 as an environmentally friendly alternative gas. Deep neural networks (DNNs), as the most common unsupervised deep learning method, are neural networks with at least one hidden layer. They can model complex nonlinear systems, but the additional layers provide a higher level of abstraction, thus improving the model's predictive ability. Therefore, DNNs can be used to build deep learning systems.
[0121] Deep learning-based insulating gas molecule screening systems first require a large number of known cases for training. The reliability and quantity of data affect the training quality and determine the final performance of the system. Therefore, a basic library of insulating gas molecules needs to be established before adopting a screening system. Relevant insulating gas molecule data and their physicochemical properties can be found through research and extensive literature review; however, classic gas molecules often lack relevant physicochemical properties. Quantum chemical and molecular dynamics simulations can analyze gas molecules from an electronic and atomic perspective, predicting molecular properties. Therefore, for gas molecules lacking relevant physicochemical properties, high-precision quantum chemical and molecular dynamics methods are used for evaluation.
[0122] The structure of gas molecules determines their properties. When using deep learning systems to predict insulating gas systems, computers cannot directly identify molecular coordinates; they need to extract the intrinsic topological information of the molecular structure to generate a data matrix as the molecular descriptor for the deep learning system. Molecular descriptors are divided into physical descriptors such as Coulomb matrices and multitension representations, and chemical information descriptors such as molecular access system structure keys, topological fingerprints, and Morgan fingerprints. If the molecular descriptor is too simple, it will affect the prediction accuracy of deep learning. If it is too complex, the deep learning system will be difficult to train successfully and will struggle to converge. Therefore, it is necessary to use Python to write programs that, by taking molecular coordinates as input, output various molecular descriptors, thereby achieving the generation of batches of small molecular descriptors.
[0123] While training a deep learning expert system is extremely time-consuming, it can predict the electrical and environmental performance indicators of gas molecules very quickly after training, enabling a "mass selection" of all possible gas molecules. Therefore, the first step is to establish the "mass selection" targets, namely, a target library of gas molecules. This target library can be built by writing an intelligent gas molecule generation program, which uses functional group substitution methods to generate new small molecule structures in batches based on existing small molecule libraries.
[0124] Therefore, this invention provides a method based on a combination of big data and deep learning algorithms. Through high-throughput search, it predicts novel environmentally friendly alternative gases for SF6 with superior performance in various aspects, offering new ideas for exploring environmentally friendly alternative gases for SF6. This can reduce large-scale experimental testing and significantly accelerate the development cycle of new molecules. However, this example is limited to providing theoretical support for realizing environmentally friendly alternatives to SF6 and reducing the time cost of experimental trial and error. The final predicted gas molecules still need to be experimentally verified.
[0125] Environmentally friendly alternative gases that fully achieve SF6 compliance may exist within a vast body of literature and small molecule databases, but their physicochemical properties as insulating gases remain unexplored and await further investigation. Alternatively, environmentally friendly alternative gases that fully achieve SF6 compliance may exist only theoretically, but have not yet been studied and are not recorded in literature or small molecule databases, yet they could be constructed using high-throughput molecular design methods. Therefore, by screening literature and small molecule data and employing high-throughput molecular design, environmentally friendly alternative gases that can fully achieve SF6 compliance can be identified.
[0126] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0127] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.
Claims
1. A method for screening environmentally friendly insulating gas molecules based on deep learning, characterized in that, Includes the following steps: S100, establish a basic library of known electrical and environmental properties of gas molecules; the performance indicators of the electrical and environmental properties include: insulation performance, liquefaction temperature, and gas volatile matter content (GWP); the structural types of gas molecules in the basic library are generated in the following ways: c) For typical gas molecule types that lack electrical and environmentally relevant physicochemical property data, high-precision quantum chemical and molecular dynamics simulations are used for evaluation; for gas molecules lacking liquefaction temperatures, the following steps are taken to obtain them: S111, establish an amorphous unit cell for multiple gas molecules to obtain the initial gas crystal structure; S112, Based on density functional theory, optimize the bonding parameters and unit cell parameters of the initial gas crystal structure to obtain a stable molecular mechanical environment gas crystal structure; S113, using the Compass II force field, the initial gas crystal structure is energy initialized to obtain a stable force field energy environment gas crystal structure; S114, calculate the average interaction force and density of each single molecule in the molecular mechanical environment gas crystal structure and the force field energy environment gas crystal structure, respectively; S115, compare the average interaction forces and densities of the molecular mechanical environment gas crystal structure and the force field energy environment gas crystal structure respectively to obtain the initial liquefaction temperature prediction model; S116, continuously adjust the force field parameters of the Compass II force field of the initial gas crystal structure in the initial liquefaction temperature prediction model, update the initial liquefaction temperature prediction model, and obtain the best fitting liquefaction temperature prediction model. This includes: updating the initial liquefaction temperature prediction model once for each adjustment of the force field parameters of the initial gas crystal structure, and performing an energy initialization on the initial gas crystal structure once for each update, and calculating the corresponding average interaction force and density, until the average interaction force and density are close to the average interaction force and density of the gas crystal structure in the molecular mechanical environment. The liquefaction temperature prediction model fitted at this time is the best liquefaction temperature prediction model. S117, Introduce the amorphous unit cells of multiple gas molecules to be tested into the optimal liquefaction temperature prediction model, perform heating and cooling simulations, obtain the liquefaction temperature curve, and obtain the liquefaction temperature range based on the liquefaction temperature curve. S200, represent the gas molecules in the basic library using molecular descriptors to generate a molecular data matrix; and divide the molecular data matrix into a training library and a test library; S300, Build a deep learning model. The training library is used to train the deep learning model. The expert system in the trained deep learning model is used to evaluate the performance index of gas molecules in the test library and compare it with the true value to evaluate the prediction accuracy. Continuously update the parameters of the deep learning model until the evaluation prediction accuracy reaches the prediction standard. The deep learning model at this time is taken as the best deep learning model. S400, gas molecules in the basic library are used to generate new small molecule structures in batches through functional group substitution method to form a gas molecule target library; S500: The optimal deep learning model is used to screen the gas molecules in the gas molecule target library, a comprehensive screening performance index is set, the preliminary screening results are sorted by the comprehensive screening performance index, an evaluation boundary is set, and potential insulating alternative gases within the evaluation boundary are output. S600: The potential insulating alternative gases are evaluated for their performance indicators, sorted again using the comprehensive screening performance indicators, and an evaluation line is set to output high-performance potential insulating alternative gases that are within the evaluation line.
2. The method for screening environmentally friendly insulating gas molecules based on deep learning according to claim 1, characterized in that, The structural types of gas molecules in the basic library also include those generated from the following: a) Research and search literature, technical reports and patents on insulating gas research and development over the past half-century, and collect relevant data on insulating molecules; b) Search literature, molecular physicochemical property databases, and physical chemistry handbooks to collect electrical and environmentally relevant physicochemical properties of small molecules, regardless of whether they can be used as insulating gases.
3. The method for screening environmentally friendly insulating gas molecules based on deep learning according to claim 1, characterized in that, The following steps are included when performing a temperature rise simulation: The optimal liquefaction temperature prediction model is initialized with energy, and the amorphous unit cells of multiple gas molecules to be tested are brought into the optimal liquefaction temperature prediction model after energy initialization. Under a canonical ensemble, the initial temperature and heating step size are set to construct a dynamic simulation environment; At the initial temperature, for each heating step increase, a molecular dynamics simulation is performed on the environment of the optimal liquefaction temperature prediction model to obtain the stable structure at this temperature, and the density of the environment of the optimal liquefaction temperature prediction model at this temperature is recorded. Repeat the previous step until the liquid system of the gas molecules to be tested boils and completely vaporizes. The density mutation point during the gradual heating process will be determined, and the first mutation temperature at this point will be recorded. The cooling simulation includes the following steps: Set the cooling step size, and gradually cool down after the gas molecules to be measured have completely vaporized; For each cooling step reduced, a molecular dynamics simulation is performed on the environment in which the optimal liquefaction temperature prediction model is located, and the stable structure at this temperature is obtained, and the density of the environment in which the optimal liquefaction temperature prediction model is located at this temperature is recorded. Repeat the previous step until the gaseous system of the gas molecules to be tested is completely liquefied, determine the density abrupt change point during the gradual cooling process, and record the second abrupt change temperature at this point. The density data obtained during the heating simulation and the density data obtained during the cooling simulation are used to plot a liquefaction temperature curve; in the liquefaction temperature curve, the range between the first abrupt change temperature and the second abrupt change temperature is the liquefaction temperature range.
4. The method for screening environmentally friendly insulating gas molecules based on deep learning according to claim 1, characterized in that, For gas molecules lacking GWP, obtain them through the following steps: S121, Determine the optimized basis set, and perform structural optimization on gas molecules and water molecules based on the optimized basis set to obtain the gas molecule-water molecule optimized structure and the gas molecule optimized structure with the lowest energy and the most stable gas molecule-water molecule optimized structure. S122, the gas molecules and water molecules in the gas molecule-water molecule optimized structure are mixed in proportion to construct an initial gas-water eutectic structure; S123, Optimize the initial gas-water eutectic structure under the lowest total energy space group according to first principles to obtain the optimal gas-water eutectic structure; S124. Based on molecular dynamics and density functional methods, the reaction path of the optimal gas-water eutectic structure is simulated to obtain all decomposition paths of the optimal gas-water eutectic structure. S125, obtain the rate constants of all decomposition paths and compare them, and select the one with the smallest rate constant as the rate-determining step of the optimal gas-water eutectic structure reaction; S126, obtain the reaction rate constant of the rate-determining step of the optimal gas-water eutectic structure reaction; and obtain the atmospheric lifetime of the gas based on the reaction rate constant; S127, Obtain the infrared spectrum of the optimized structure of the gas molecules, obtain the band based on the infrared spectrum, and obtain the spectral-resolved radiative forcing per unit cross-section based on the band. S128, Based on the wavelength and the spectral resolution of the radiative forcing, obtain the radiative efficiency of the optimized gas molecule structure; S129, Based on the atmospheric lifetime and the radiation efficiency, obtain the global warming potential value of the gas.
5. The method for screening environmentally friendly insulating gas molecules based on deep learning according to claim 4, characterized in that, The construction of the initial gas-water eutectic structure includes: obtaining the interaction force between the optimized structure of the gas molecules and the water molecules under a global force field; and based on the Monte Carlo algorithm, predicting the combination and stacking of all interactions under the C2 / c space group, P1 space group, P21 space group, P2 / c space group and P212121 space group respectively, obtaining multiple initial gas-water eutectic structures, and sorting the total energy of the initial gas-water eutectic structures under each space group to obtain the initial gas-water eutectic structure under the space group with the lowest total energy.
6. The method for screening environmentally friendly insulating gas molecules based on deep learning according to claim 4, characterized in that, An ultrasoft pseudopotential is used to bring the initial gas-water eutectic structure under the lowest total energy space group into force equilibrium, thereby obtaining the initial gas-water eutectic structure with the lowest energy under the lowest total energy space group, which is taken as the optimal insulating gas-water eutectic structure.
7. The method for screening environmentally friendly insulating gas molecules based on deep learning according to claim 6, characterized in that, All decomposition paths for obtaining the optimal gas-water eutectic structure include: The energy of the optimal gas-water eutectic structure is initialized to ensure that the environment in which the optimal gas-water eutectic structure is located is in equilibrium before heating. Under a canonical ensemble, the initial temperature, simulation steps, and temperature intervals are set. The initial temperature is used as the first temperature for kinetic simulation. Then, the environment in which the optimal gas-water eutectic structure is located is gradually heated at a determined temperature interval until all decomposition pathways of hydroxyl radicals in the gas and water molecules are obtained.
8. The method for screening environmentally friendly insulating gas molecules based on deep learning according to claim 1, characterized in that, The target library for forming gas molecules includes: Input gas molecules from the base library; To establish substitution rules and generate new small molecule structures, it is necessary to satisfy Lewis's octet rule and hybrid orbital theory. The molecular structure can be replaced in batches by using the functional group substitution method, while satisfying the substitution rules. A new gas molecule structure conforming to the substitution rule is output to form the gas-forming molecule target library; and all gas molecules in the gas-forming molecule target library adopt B3LYP / 6-311G. The method performs structural optimization to determine its ground-state geometry.
Citation Information
Patent Citations
Method for simulating cracking of silicone rubber material at high temperature in flashover development
CN115295085A
Intelligent screening method for low-melting-point high-stability fused salt heat storage material
CN116013436A
