Screening method of insulating gas decomposition product sensing material based on machine learning

Abstract

The invention discloses a screening method of insulating gas decomposition product sensing materials based on machine learning, and relates to the technical field of functional material design and gas sensing. The method provided by the invention comprises the following steps: selecting a first candidate sensing material according to an insulating gas and a decomposition product; doping elements or embedding atoms into a central cavity of the first candidate sensing material to obtain a second candidate sensing material, and constructing a candidate sensing material library; obtaining the most stable adsorption configuration and adsorption energy of gas molecules, constructing feature descriptors, forming an initial data set, and dividing a training set and a test set; constructing and training a model, and evaluating and screening out an optimal model by using a test set; predicting and screening the second candidate sensing material by using the optimal model to obtain a third candidate sensing material; and performing DFT calculation verification and performance index evaluation on the third candidate sensing material, and outputting a final candidate sensing material. The method provided by the invention is high in prediction accuracy and reliability, and a novel sensitive material can be found.

Description

Technical Field

[0001] This invention relates to the fields of functional materials design and gas sensing technology, and in particular to a method for screening sensing materials for insulating gas decomposition products based on machine learning. Background Technology

[0002] Sulfur hexafluoride (SF6) has long been widely used in high-voltage electrical equipment such as gas-insulated switchgear due to its excellent insulation and arc-quenching properties. However, SF6 is an extremely potent greenhouse gas, with a global warming potential (GWP) 23,900 times that of carbon dioxide, and an extremely long atmospheric lifetime. Therefore, it was listed as one of the six greenhouse gases whose use needs to be restricted under the Kyoto Protocol.

[0003] To address climate change, the global power industry is urgently seeking environmentally friendly insulating gases to replace SF6. In recent years, gases such as trifluoromethanesulfonyl fluoride (CF3SO2F) and heptafluoroisobutyronitrile (C3F7CN) have been considered promising alternatives to SF6 due to their high insulation strength, relatively low gas permeable (GWP), and good chemical stability. However, these new insulating gases decompose and produce various byproducts when partial discharge or overheating faults occur inside equipment. The main decomposition products of CF3SO2F include CF4, SO2, SO2F2, and HF. The decomposition products of C3F7CN are more complex, including CF4, C2F4, C2F6, C3F6, C3F8, C2N2, CF3CN, and C2F5CN. These decomposition gases can not only corrode equipment materials and accelerate insulation degradation, threatening the safe operation of the power grid, but some gases are also harmful to human health and the environment. Therefore, online monitoring of these characteristic decomposition gases can effectively diagnose early insulation defects in power equipment, which is a key technological link in realizing condition-based maintenance and the development of smart grids.

[0004] Currently, the core bottleneck in this field lies in the lack of high-performance gas sensing materials. Ideal sensing materials need to possess high sensitivity, high selectivity, and reversible response recovery characteristics to specific gases. Traditional materials research and development heavily relies on the experience of experimentalists and a trial-and-error approach, resulting in cumbersome processes, long cycles, and high costs, which can no longer meet the demand for rapid development of new materials. Although first-principles calculations based on density functional theory (DFT) can reveal the interaction mechanism between gases and materials at the microscopic level and predict changes in adsorption energy and electronic structure, their enormous computational cost severely limits their ability to directly perform large-scale, high-throughput screening. Calculating the energy and electronic structure of a system typically takes hours to days. For the vast combination space composed of dozens of material substrates and multiple gas molecules, performing comprehensive DFT calculations is infeasible in terms of both time and computing power.

[0005] Although machine learning technology has demonstrated powerful data mining and prediction capabilities in the field of materials science, its successful application to the design of such specific gas sensing materials still requires solving the key scientific problems of how to construct feature descriptors that can accurately describe the complex "gas-material" interaction and how to ensure the accuracy and physical interpretability of model predictions.

[0006] Therefore, developing a high-throughput screening method that can integrate the accuracy of first-principles calculations with the efficiency of machine learning is of great scientific significance and engineering application value for accelerating the research and development of next-generation gas sensors for insulation fault diagnosis. Summary of the Invention

[0007] To address the shortcomings of existing methods for screening sensing materials for insulating gas decomposition products, this invention provides a machine learning-based method for such screening. By combining first-principles calculations with machine learning techniques, an optimal machine learning prediction model is established and screened. This model can rapidly select ideal sensing materials from a vast pool of candidate materials that simultaneously meet the requirements of high sensitivity, high selectivity, rapid response and recovery, and good stability for detecting the decomposition products of environmentally friendly insulating gases (such as CF3SO2F and C3F7CN). The invention employs the following technical solutions to achieve the aforementioned technical objectives.

[0008] The present invention provides a method for screening sensing materials for insulating gas decomposition products based on machine learning, comprising the following steps:

[0009] Select the insulating gas to be detected and the corresponding decomposition product gas, as well as the first candidate sensing material;

[0010] Several second candidate sensing materials are obtained by doping with several elements or embedding atoms into the central cavity of the first candidate sensing material.

[0011] Several second candidate sensing materials and decomposition product gases are combined to form several sensing material-gas combinations;

[0012] For each of the aforementioned sensing material-gas combinations, the most stable adsorption configuration of the decomposition product gas molecules is calculated, along with the corresponding charge transfer amount, electronic structure information, and feature descriptors, forming an initial dataset; the initial dataset is then divided into a training set and a test set.

[0013] Select several initial machine learning prediction models, train them using the training set, evaluate them using the test set, and select the optimal machine learning prediction model.

[0014] Using the optimal machine learning prediction model, high-throughput prediction and screening of the sensing material-gas combination are performed to obtain a third candidate sensing material;

[0015] The third candidate sensing material is verified by DFT calculation and its performance index is evaluated to output the final candidate sensing material.

[0016] The method for screening sensing materials for insulating gas decomposition products provided by this invention first requires identifying the insulating gas to be detected, such as CF3SO2F or C3F7CN. Through further theoretical calculations or experimental data, the main types of decomposition products with detection significance are determined, such as CF4, SO2, SO2F2, HF, and CF3CN.

[0017] In addition, two-dimensional materials with high specific surface area and tunable electronic structure are selected as substrates (first candidate sensing materials), such as graphitic carbon nitride g-C3N4, hexagonal aluminum nitride h-AlN, graphene, etc.

[0018] Then, a candidate sensing material library (e.g., TM / g-C3N4, X@h-AlN) is constructed through elemental doping or atomic embedding (e.g., transition metals TM, main group elements X, etc.). The candidate sensing material library should be of sufficiently large size.

[0019] Furthermore, for each of the aforementioned sensing material-gas combinations, the method for calculating and obtaining the most stable adsorption configuration and adsorption energy of the decomposition product gas molecules is as follows:

[0020] Twenty percent of the sensor material-gas combinations were randomly selected, and DFT calculations were performed using VASP. The calculations employed PAW pseudopotentials and PBE functionals, and the DFT-D3 method was introduced to correct for van der Waals forces. The cutoff energy was set to 520 eV, and the K-point mesh was optimized using a Γ-centered 3×3×1 Monkhorst-Pack mesh. The electronic self-consistent iteration energy convergence criterion was 10 eV. -5 eV, atomic force convergence threshold set to 0.02 eV / Å, 20 Å vacuum layer set in the vertical direction;

[0021] For each sensing material-gas combination, calculate the most stable adsorption configuration of the decomposition product gas molecules and its adsorption energy E. ads Adsorption energy E ads The calculation formula is:

[0022] E ads =E total -E material -E gas ;

[0023] Among them, E total E represents the total energy of the sensor material-gas combination after adsorption. material E represents the energy of the second candidate sensing material itself. gas The energy of the isolated decomposition product gas molecules.

[0024] Furthermore, based on knowledge from the field of physical chemistry, a comprehensive set of digital features, or feature descriptors, is constructed for each sensing material-gas combination. The specific data corresponding to each feature descriptor is the multidimensional feature vector data. The feature descriptors include material-side descriptors, gas-side descriptors, and structural deformation descriptors;

[0025] The material-side descriptor includes both specific to the modifying atom (such as TM or X) and the material as a whole. This includes, but is not limited to, the atomic radius (r) of the modifying atom. M ), relative atomic mass (M) M Pauli electronegativity (χ²) M ), the number of electrons in the outermost d / p orbitals (e d / e p ), first ionization energy (I M ), electron affinity (A M ), d / p band center (ε) d / ε p Enthalpy of formation of oxides, enthalpy of formation of fluorides, etc.

[0026] The structural deformation descriptor includes a maximum protrusion distance, which is the maximum vertical displacement that causes the doped atom to protrude beyond the original plane of the first candidate sensing material.

[0027] The gas-side descriptor includes a feature descriptor for the decomposition product gas molecules, a feature descriptor for the adsorbed atoms of the decomposition product gas molecules, a global physical and chemical descriptor, and a quantum chemical descriptor.

[0028] The descriptor of the decomposition product gas molecules includes molecular weight (M). W ), average atomic radius (r Mg ), polar surface area (PSA), average electronegativity (χ²) avg ), lipid-water partition coefficient (Log P), number of hydrogen bond acceptors (HB) a ), and a geometric descriptor based on Smooth Overlapping Atom Positions (SOAP).

[0029] The descriptor of the adsorbed atoms of the gas molecules includes the atomic radius (r) of the adsorbed atoms (such as O, F, N). g ), relative atomic mass (M) g Pauli electronegativity (χ²) g ), the number of electrons in the outermost p orbital (e p ), first ionization energy (I g ), electron affinity (A g ).

[0030] The quantum chemical descriptor is obtained based on the type and number of functional groups in the decomposition product gas molecules and the electronic contributions of key reacting atoms; the quantum chemical descriptor includes charge distribution descriptors, frontier orbital descriptors, and chemical reactivity descriptors.

[0031] Furthermore, the method for obtaining the quantum chemical descriptor is as follows:

[0032] Based on the three-dimensional structure or linear identifier of the gas molecules in the decomposition products, specific functional groups can be identified.

[0033] Multithermal coding is used to generate a functional group existence vector for each decomposition product gas molecule; a functional group counting feature is added to the functional group existence vector;

[0034] Based on partial density of states analysis and frontier orbit analysis, key functional groups were identified; atomic-level electronic structure analysis was performed on the screened sensitive gas adsorption systems to further identify key reaction atoms;

[0035] For the identified key reaction atoms, quantum chemical descriptors are extracted to quantify the electronic behavior of the key reaction atoms.

[0036] Furthermore, various machine learning algorithms are used to train the subsequent initial machine learning prediction model. The initial machine learning prediction model includes, but is not limited to, any one of: Gradient Boosting Regression (GBR), Random Forest Regression (RFR), Support Vector Regression (SVR), k Nearest Neighbor Regression (KNR), Kernel Ridge Regression (KRR), LASSO Regression, Artificial Neural Network (ANN), and Gaussian Process Regression (GPR).

[0037] Furthermore, the model is trained using the training set and evaluated using the test set to select the optimal machine learning prediction model.

[0038] The initial dataset is randomly divided into a training set and a test set at a ratio of 80% and 20% respectively;

[0039] The hyperparameters of the initial machine learning prediction model are optimized on the training set using 5-fold cross-validation to minimize the root mean square error of the prediction.

[0040] Retrain the model using optimized hyperparameters and evaluate performance on an independent test set;

[0041] Furthermore, the method for optimizing the hyperparameters of the initial machine learning prediction model on the training set using 5-fold cross-validation is as follows:

[0042] Calculate the Pearson correlation coefficient matrix among all feature descriptors in the training set, and remove highly correlated redundant features;

[0043] Calculate the Spearman correlation coefficient between the remaining features and the adsorption energy target, and retain the features with the highest correlation to the target within each highly correlated feature group to obtain a feature subset;

[0044] Model performance was evaluated through cross-validation, and hyperparameters were tuned using grid search or Bayesian optimization. Root mean square error and coefficient of determination were used as the main evaluation metrics to select models that met the requirements.

[0045] The method for evaluating performance on independent test sets is as follows:

[0046] The root mean square error and coefficient of determination between the adsorption energy predicted by the computational model and the actual DFT value;

[0047] Plot a scatter plot of the predicted values ​​and the actual values ​​to visually assess the consistency and dispersion of the predictions;

[0048] The training and testing process of the model was repeated hundreds of times, with the initial dataset being randomly re-split each time. The optimal machine learning prediction model was determined by evaluating the mean and standard deviation of the root mean square error and the coefficient of determination.

[0049] Furthermore, the method for using the optimal machine learning prediction model to perform high-throughput prediction and screening of the sensing material-gas combination to obtain the third candidate sensing material is as follows:

[0050] Determine the set temperature and target recovery time range based on actual application requirements;

[0051] Based on the upper and lower limits of the target recovery time range and the set temperature, substitute them into the following calculation formula to solve for the corresponding ideal adsorption energy range;

[0052] τ=ν -1 exp(-E ads / (k B T ));

[0053] Where ν is the frequency of attempts, E ads k is the absolute value of the adsorption energy. B Where is Boltzmann's constant, and T is the operating temperature;

[0054] The predicted adsorption energies output by the optimal machine learning prediction model are sorted from strong to weak. It is determined whether the predicted adsorption energies fall within the ideal adsorption energy range. The second candidate sensing material corresponding to the predicted adsorption energy falling within the range of the ideal adsorption energy is selected as the third candidate sensing material.

[0055] Furthermore, the method for performing DFT calculations and performance evaluations on the third candidate sensing material to output the final candidate sensing material is as follows:

[0056] The predicted adsorption energy of the third candidate sensing material was accurately calculated using DFT to verify whether it was still within the ideal recovery time range and to evaluate the prediction error.

[0057] The performance indicators include sensitivity assessment, selectivity assessment, repeatability assessment, and stability assessment.

[0058] Furthermore, for resistive sensing materials, the sensitivity is evaluated by calculating the change in the material's band gap Eg before and after the adsorption of decomposition product gases, and the sensitivity S is required to be higher than a predetermined threshold. The formula for calculating the sensitivity S is:

[0059] S = |(G-G0) / G0|;

[0060] Wherein, G0 and G are the electrical conductances of the third candidate sensing material in a clean state and after adsorbing and decomposing the gas products, respectively.

[0061] Furthermore, by utilizing density functional theory coupled with non-equilibrium Green's function, the electron transport characteristics are simulated to obtain the drain current-gate voltage transfer characteristic curve and extract the subthreshold slope. By comparing the changes in the subthreshold slope before and after the adsorption of decomposition product gases, the sensitivity λ is calculated, requiring λ for the decomposition product gases to be higher than a set threshold. The formula for calculating the sensitivity λ is:

[0062] λ=|SS * -SS0| / SS0;

[0063] Among them, SS0 and SS * These represent the subthreshold slopes of the device in a clean state and after adsorbing gas, respectively.

[0064] Furthermore, for the selectivity, the sensitivity ratio θ of the decomposition product gas relative to other coexisting interfering gases is calculated according to the following formula;

[0065] θ=λ target / λ interfering ;

[0066] Where, λ target λ represents the sensitivity of the sensing material to the target gas. interfering The sensitivity of the sensing material to a certain interfering gas;

[0067] Preferably, to further quantify the overall selectivity advantage, an average selectivity difference Δθ is introduced, which is the ratio of the average sensitivity of the target gas to the average sensitivity of the interfering gas. The larger Δθ is, the stronger the selectivity of the material for the functional groups of the target gas. The formula for calculating Δθ is:

[0068] ;

[0069] in, The average sensitivity of the third candidate sensing material to the decomposition product gases. The average sensitivity of the third candidate sensing material to a certain interfering gas.

[0070] Furthermore, regarding the repeatability, it is jointly evaluated by the maximum protrusion distance and molecular dynamics AIMD simulation, requiring that gas adsorption does not cause irreversible and significant structural deformation of the sensing material or the gas molecules themselves (Δl value must be within a reasonable range).

[0071] Furthermore, regarding the stability, the cluster energy E is calculated using the following formula. clus Assess the likelihood of aggregation of elements or atoms doped / embedded on the surface of the first candidate sensing material;

[0072] E clus =E bind -E coh ;

[0073] Among them, E bind E represents the binding energy between the doped / intercalated element or atom and the first candidate sensing material, indicating the energy at which a TM atom is adsorbed on the surface of the first candidate sensing material. coh Cohesive energy represents the average binding energy of each atom of TM atoms in the first candidate sensing material.

[0074] If the cluster energy E clus <0 indicates that the doped / embedded elements or atoms tend not to aggregate;

[0075] If the cluster energy E clus >0 indicates that the doped / embedded elements or atoms tend to aggregate to form clusters;

[0076] Furthermore, the thermal stability of the third candidate sensing material structure was verified by performing molecular dynamics AIMD simulations at predetermined temperatures (e.g., 300 K, 500 K, etc.) for a period of time (e.g., 5-10 ps).

[0077] Finally, the final list of candidate sensitive materials that simultaneously meet all the above performance requirements is output, along with the corresponding optimal detection gas. Key performance parameters, such as precise adsorption energy, sensitivity, selectivity (Δθ), recovery time, and stability evidence, are provided.

[0078] Compared with the prior art, the advantages of the present invention are:

[0079] 1. This invention uses machine learning prediction models to replace most of the expensive DFT calculations, enabling high-speed and low-cost traversal screening of a vast material space, shortening the research and development cycle from "years" to "weeks" or even "days".

[0080] 2. This invention is based on accurate DFT calculation data and incorporates profound physicochemical characteristics, ensuring high accuracy and reliability of machine learning predictions.

[0081] 3. This invention is a general method that is not only applicable to the screening of materials for detecting the decomposition products of CF3SO2F and C3F7CN, but can also be widely applied to the material design of other specific gas sensing materials with slight adjustments.

[0082] 4. This invention introduces a series of new quantitative descriptors such as molecular weight (MW), maximum protrusion distance (Δl), FET sensitivity (λ), and average selectivity difference (Δθ), and constructs a more refined and comprehensive performance evaluation system for gas sensing materials, which is especially suitable for the design of FET-type sensing materials.

[0083] 5. This invention can discover novel sensitive materials with excellent performance that have not been reported before, providing clear theoretical guidance and target for experimental synthesis, and greatly reducing the blindness of experiments. Attached Figure Description

[0084] Figure 1 This is a flowchart of the high-throughput screening method described in this invention.

[0085] Figure 2 A scatter plot comparing the machine learning predictions of the training set and the actual values ​​calculated by the DFT to the actual values ​​on the test set.

[0086] Figure 3 This is a heatmap of adsorption energy predicted for high throughput.

[0087] Figure 4 The band structure and density of electronic states (DOS) diagram for the final selected material.

[0088] Figure 5 This is the transfer characteristic curve (IV curve) of the FET device. Detailed Implementation

[0089] 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.

[0090] Example

[0091] The machine learning-based screening method for sensing materials of insulating gas decomposition products provided in this embodiment uses CF3SO2F as the environmentally friendly insulating gas to be detected, and screens sensitive materials for detecting SO2F2, the decomposition product of CF3SO2F. For example... Figure 1 As shown, the specific steps are as follows.

[0092] Step 1: Determine the target system and construct a candidate sensing material library.

[0093] 1. Define the target system

[0094] The environmentally friendly insulating gas to be tested was identified as CF3SO2F, and the corresponding decomposition products were identified as CF4, SO2, SO2F2, and HF.

[0095] 2. Construct a candidate sensor material library

[0096] Graphite-phase carbon nitride monolayer (g-C3N4) was selected as the substrate material (first candidate sensing material). g-C3N4 has abundant nitrogen coordination vacancies, which facilitates the stabilization of metal atoms.

[0097] Twenty-eight 3d, 4d, and 5d transition metals (such as Sc, Zn, Cd, Os, Pd, Nb, and Cr) were embedded into the central cavity of a g-C3N4 monolayer to construct a candidate sensing material library containing 28 TM / g-C3N4 (transition metal-doped graphite-phase carbon nitride monolayer materials, i.e., the second candidate sensing materials).

[0098] The candidate sensing material library constructed in this embodiment has high specific surface area, tunable electronic structure and good stability, providing an ideal platform for gas adsorption and sensing.

[0099] Step 2: Obtain a partial initial dataset by using first principles calculations.

[0100] 1. A total of 112 full combination spaces were formed by combining 28 types of TM / g-C3N4 and the decomposition product gases of 4 selected insulating gases (CF4, SO2, SO2F2, HF), namely 112 "TM / g-C3N4-gas" (i.e. the second candidate sensing material TM / g-C3N4 and the corresponding decomposition product gases).

[0101] Decomposition product gases refer to the main decomposition product gases produced during the discharge or thermal decomposition of insulating gases, namely CF4, SO2, SO2F2, and HF.

[0102] 2. 20% of the combinations (approximately 22) were randomly selected from the 112 “TM / g-C3N4-gas” samples as representative samples.

[0103] 3. For each selected "TM / g-C3N4-gas" combination, DFT calculations were performed using VASP, a first-principles calculation software based on density functional theory (DFT), to obtain the most stable adsorption configuration, charge transfer amount, and electronic structure information (including band structure, density of states, and band gap of the material). These parameters are important components of the characteristic descriptor.

[0104] The calculations employed a PAW pseudopotential and a PBE functional, and the van der Waals force was corrected using the DFT-D3 method. The cutoff energy was set to 520 eV, and the K-point mesh was optimized using a Γ-centered 3×3×1 Monkhorst-Pack mesh. The convergence criterion for the electronic self-consistent iteration was 10 eV. -5 eV, with the atomic force convergence threshold set to 0.02 eV / Å. A 20 Å vacuum layer was set vertically to eliminate periodic mirror interactions.

[0105] 4. Precise calculation: Based on the above structural optimization, calculate and determine the adsorption energy (E) for the most stable adsorption configuration of each selected "TM / g-C3N4-gas" combination. ads This data provides high-quality labeled data for machine learning, serving as target values ​​for training machine learning prediction models.

[0106] Adsorption energy E ads The calculation formula is:

[0107] E ads =E total -E material -E gas ;

[0108] Among them, E total E represents the total energy of the "TM / g-C3N4- gas" combination after adsorption. material E represents the energy of the second candidate sensing material itself. gas E represents the energy of the gas molecules in the isolated decomposition products. ads E is a negative number. ads The more negative the value, the stronger the adsorption.

[0109] The adsorption energies E corresponding to the above 22 "TM / g-C3N4-gas" combinations ads Data, including multidimensional feature vector data (i.e., the specific data corresponding to each feature descriptor), is used as part of the initial dataset for subsequent training and testing of the prediction model.

[0110] Step 3: Obtain feature descriptors.

[0111] For each "TM / g-C3N4-gas" in step two, a feature descriptor is constructed and used as another part of the initial training set. The feature descriptors are divided into material-side descriptors, gas-side descriptors, and structural deformation descriptors.

[0112] 1. Material-side descriptor

[0113] This applies to both the modifying atoms (transition metals™) and the material as a whole. It includes, but is not limited to, the following nine feature descriptors: atomic radius (r) of the modifying atoms. M ), relative atomic mass (M) M Pauli electronegativity (χ²) M ), the number of electrons in the outermost d / p orbitals (e d / e p ), first ionization energy (I M ), electron affinity (A M ), d / p band center (ε) d / ε p Enthalpy of formation of oxides, enthalpy of formation of fluorides, etc.

[0114] These characteristic descriptors determine the ability of TM atoms to donate or accept electrons, which is key to their interaction with decomposition product gas molecules.

[0115] 2. Gas-side descriptor

[0116] The gas-side descriptor includes the characteristic descriptor of the decomposition product gas molecules and the characteristic descriptor of the adsorbed atoms of the decomposition product gas molecules; it is also necessary to take into account the type and number of functional groups (such as -CN, -CF3) and the electronic contributions of key reactive atoms (such as Cα, Nα, i.e. C and N in -C≡N) to obtain the quantum chemical descriptor.

[0117] (1) Characteristic descriptors of decomposition product gas molecules

[0118] Including but not limited to the following seven characteristic descriptors: molecular weight (MW), average atomic radius (r) Mg ), polar surface area (PSA), average electronegativity (χ²) avg ), lipid-water partition coefficient (Log P), number of hydrogen bond acceptors (HB) a ), and a geometric descriptor based on Smooth Overlapping Atom Positions (SOAP).

[0119] (2) Characteristic descriptors of adsorbed atoms of decomposition product gas molecules

[0120] The characterization descriptors for the adsorbed atoms (e.g., O, F, N) of the decomposition product gas molecules include the following six characterization descriptors: atomic radius (r) g ), relative atomic mass (M) gPauli electronegativity (χ²) g ), the number of electrons in the outermost p orbital (e p ), first ionization energy (I g ), electron affinity (A g ).

[0121] (3) Quantum chemical descriptors

[0122] By specifically considering the type and number of functional groups (e.g., -CN, -CF3) and the electronic contributions of key reacting atoms (e.g., Cα, Nα, i.e., C and N in -C≡N), quantum chemical descriptors are obtained. The specific method is as follows:

[0123] ① Identify the characteristic functional groups contained in the gas molecules of the decomposition products based on their three-dimensional structure or linear identifier (such as SMILES).

[0124] For example, common decomposition products of insulating gases include cyano (-CN), trifluoromethyl (-CF3), and sulfonyl fluoride (-SO2F).

[0125] ② Multi-hot encoding is used to generate a functional group existence vector for each decomposition product gas molecule; functional group counting features are added on the basis of functional group existence vector to accurately reflect the type and quantity information of functional groups.

[0126] For example, the C2N2 molecule has a "-CN functional group count" characteristic value of 2, which strictly corresponds to its chemical fact of containing two cyano groups. This method can capture the linear or nonlinear influence of functional group abundance on adsorption strength.

[0127] ③ Based on atomic-level electronic structure analysis (including wave density of states analysis and frontier molecular orbital analysis), key functional groups were identified; the atomic-level electronic structure of the screened "TM / g-C3N4-gas" combination was determined. The specific method is as follows:

[0128] a. Partial density of states analysis

[0129] The density of states of the decomposition product gas molecules is decomposed into the contributions of different atomic orbitals (e.g., C-2p, N-2p, F-2p, etc.). By comparing the density of states contribution intensity of each atom in the bandgap defect state energy range, the atom that plays a dominant role in the formation of defect states is identified.

[0130] b. Frontier Molecular Orbital Analysis

[0131] Plotting the distribution of frontier molecular orbital electron clouds for the highest occupied orbitals and lowest unoccupied orbitals of decomposition product gas molecules can visually demonstrate the regions of concentrated electron density and reactive sites in the decomposition product gas.

[0132] c. Identification of key functional groups

[0133] Based on the combined results of partial wave density of states analysis and frontier orbit analysis, the functional group to which the atom with the largest contribution and the highest electron cloud density belongs is identified as the key functional group.

[0134] For example, if the results of partial density of states analysis show that the defect states are mainly contributed by C and N atoms in the functional group -CN, and the results of frontier orbit analysis show that the frontier orbit electron cloud is highly concentrated in -CN, then -CN is determined to be the key functional group affecting the sensing performance.

[0135] ④ Identify the key functional groups and then identify the key reaction atoms.

[0136] Key reactive atoms are the core atoms that directly participate in charge transfer, orbital hybridization, or chemical bonding within key functional groups. Identification of key reactive atoms relies on electronic structure analysis.

[0137] Taking molecules containing the -CN (-C≡N) functional group as an example, preliminary density functional theory calculations were used to analyze the distribution of the partial density of states (PDOS) near the Fermi level and the spatial localization of the frontier molecular orbital electron cloud. The calculation results show that the electronic states of the carbon atom (Cα) and nitrogen atom (Nα) constituting the -C≡N triple bond dominate the defect states in the band gap, and the frontier molecular orbital electron cloud is highly concentrated around these two atoms.

[0138] Therefore, carbon (Cα) and nitrogen (Nα) atoms are identified as primary key reactants. Atoms directly bonded to Cα (such as Cβ and Fβ in C2F5CN) contribute secondary electrons due to inductive effects, and are thus defined as secondary key reactants.

[0139] ⑤ For the identified key reaction atoms, quantum chemical descriptors are extracted to accurately quantify the electronic behavior of the key reaction atoms.

[0140] The final quantum chemical descriptor includes:

[0141] a. Charge distribution descriptor

[0142] Calculate the Mullicken population or Bader charge of key reaction atoms to quantify the magnitude and positivity of the partial charge (net charge value) carried by key reaction atoms, directly reflecting their tendency to participate in electrostatic interactions.

[0143] b. Frontline track descriptor

[0144] Calculate the HOMO and LUMO orbital energy levels (EHOMO, ELUMO) of the decomposition product gas molecules; calculate the percentage contribution of key reaction atomic orbitals to the HOMO and LUMO through orbital composition analysis.

[0145] The contribution weight percentage directly measures the degree of dominance of the key reaction atom in the most active electron-donating or accepting process of the decomposition product gas molecules.

[0146] c. Chemical reactivity descriptors

[0147] Calculate the Fukui function values ​​of key reactive atoms to obtain their electrophilicity (f... + ) and nucleophilic (f - The index is used to predict whether an electron will act as an electron donor or an electron acceptor during adsorption.

[0148] Finally, the characteristic descriptors of the decomposition product gas molecules, the characteristic descriptors of the adsorbed atoms of the decomposition product gas molecules, and the quantum chemical descriptors are combined to form a multi-level, high-information-density gas-side descriptor.

[0149] 3. Structural Deformation Descriptor

[0150] The maximum vertical displacement of the doped atom (TM) protruding from the original plane (h-AlN atom plane) of the first candidate sensing material is defined as the maximum protrusion distance (Δl), which is used to quantitatively describe the geometric deformation of the doped atom relative to the original plane of the first candidate sensing material after gas adsorption, in order to evaluate the repeatability of the sensing material.

[0151] The three feature descriptors described above can transform complex physicochemical interactions into machine-learnable numerical features. These three feature descriptors are also used as part of the initial training set.

[0152] Step 4: Training and optimization of the machine learning prediction model.

[0153] The following are the specific methods for training gradient boosting regression (GBR), random forest regression (RFR), support vector regression (SVR), k-nearest neighbor regression (KNR), kernel ridge regression (KRR), LASSO regression, artificial neural network (ANN), and Gaussian process regression (GPR) models using the Scikit-learn library.

[0154] 1. Normalize the complete initial dataset containing feature descriptors and adsorption energy labels obtained in steps 2 and 3, scaling each feature value to the [0,1] interval to eliminate dimensional differences and accelerate model convergence.

[0155] 2. Randomly divide the data into a training set (80%) and a test set (20%).

[0156] 3. Five-fold cross-validation is used to optimize the hyperparameters of the above models on the training set to minimize the root mean square error (RMSE) of prediction. The specific method is as follows:

[0157] (1) Calculate the Pearson correlation coefficient matrix among all feature descriptors in the training set and remove redundant features that are highly correlated (e.g., correlation coefficient > 0.85);

[0158] (2) Calculate the Spearman correlation coefficient between the remaining features and the adsorption energy target, retain the features with the highest correlation to the target in each highly correlated feature group, and finally obtain a concise and effective feature subset.

[0159] The feature subset includes, for example, d-band centers (ε-bands) of transition metals. d Electronegativity of gas adsorbed atoms (χ²) g ), average atomic radius of gas molecules (r) avg Core descriptors with clear physicochemical significance, such as the count of key functional groups, are used as the final input to the machine learning model.

[0160] (3) Evaluate model performance through cross-validation (e.g., 5-fold cross-validation), and fine-tune model hyperparameters using grid search or Bayesian optimization. Measure root mean square error (RMSE) and coefficient of determination (R²). 2 Using this as the main evaluation metric, the model with the highest prediction accuracy and the strongest generalization ability was selected for subsequent testing.

[0161] 4. Retrain the model using the optimized hyperparameters; evaluate performance on an independent test set. The specific method for evaluating performance on an independent test set is as follows:

[0162] (1) Calculate the core evaluation indicators

[0163] The root mean square error (RMSE) and coefficient of determination (R²) between the predicted adsorption energy and the actual DFT value calculated by the computational model are as follows: 2 ).

[0164] (2) Conduct error visualization analysis

[0165] Plot a scatter plot of predicted and actual values ​​to visually assess the consistency and dispersion of the predictions. Figure 2 As shown.

[0166] (3) Assess statistical stability

[0167] To mitigate the impact of a single random partitioning of the initial dataset, the model training and testing process was repeated hundreds of times, with the initial dataset being re-randomized each time. The final report focuses on RMSE and R...2 The mean and standard deviation were used to verify the reliability of the model performance.

[0168] After 5-fold cross-validation, it was found that the GPR model performed well on the training set but poorly on the test set; KRR and ANN performed poorly on both sets; GBR and RFR performed well on the training set but had poor stability on the test set. SVR was the only model that performed excellently and stably on both the training and test sets.

[0169] Therefore, Support Vector Regression (SVR) was determined to be the optimal machine learning prediction model; its root mean square error (RMSE) between the predicted results and the true DFT values ​​was as low as 0.223 eV, and its coefficient of determination (R²) was also low. 2 The accuracy is as high as 0.911, demonstrating excellent prediction accuracy and generalization ability.

[0170] Step 5: High-throughput prediction and preliminary screening.

[0171] This step takes the "TM / g-C3N4-SO2F2" combination (i.e., the second candidate sensing material TM / g-C3N4 and the decomposition product gas molecules SO2F2) as an example, and inputs its corresponding feature descriptor into the optimal machine learning prediction model.

[0172] 1. Set an acceptable recovery time range based on actual application requirements, such as a few seconds to about 1000 seconds.

[0173] 2. Derive the corresponding ideal adsorption energy range. The deduction method is based on the adsorption energy E. ads The Arrhenius-type relationship formula between the recovery time τ and the recovery time τ is explained in the following steps.

[0174] (1) Set parameters according to the actual application scenario, including setting the temperature and target recovery time range.

[0175] ① Temperature setting: Considering that the internal temperature of the equipment is usually higher than room temperature, the temperature is set to 328 K (approximately 55°C). For a higher standard, the temperature at 388 K (approximately 115°C) can also be calculated.

[0176] ② Target recovery time range: Set to a few seconds to approximately 1000 seconds, depending on the actual application requirements of the sensor.

[0177] (2) Back-calculate the adsorption energy range: Substitute the upper and lower limits of the target recovery time τ and the set temperature T into the following formula for calculating the recovery time τ to solve for the corresponding ideal adsorption energy range.

[0178] τ=ν -1 exp(-E ads / (k B T ));

[0179] Where ν is the frequency of attempts, usually taken as 10. 12 s -1 E ads k is the absolute value of the adsorption energy. B Where is Boltzmann's constant, and T is the operating temperature, such as the internal temperature of electrical equipment at 328 K, or the maximum allowable temperature at 388 K.

[0180] 3. High-throughput predictions were performed using the output of the optimal machine learning prediction model (optimal SVR model) for the predicted adsorption energies of SO2F2 for all 28 types of TM / g-C3N4 materials, such as... Figure 3 As shown.

[0181] Specifically, the predicted adsorption energies output by the optimal machine learning prediction model are sorted from strongest to weakest to determine whether the predicted adsorption energies fall within the aforementioned ideal adsorption energy range.

[0182] After the screening in step five, the predicted adsorption energies of materials such as Cd / g-C3N4 fall within this range, and they are selected as candidate materials for the next round of fine-tuning, i.e., the third candidate sensing materials. This step efficiently reduces the number of candidate sensing materials requiring in-depth calculations from hundreds to just a few.

[0183] Step 6: Multi-indicator collaborative optimization screening.

[0184] For the several third-candidate sensing materials obtained in step five, rigorous DFT calculations and multi-performance index evaluations were performed again. The specific methods are as follows.

[0185] 1. Verification of Adsorption Energy and Recovery Time

[0186] The predicted adsorption energy of the third candidate sensing material is accurately calculated using DFT, and this value is compared with the machine learning prediction value to evaluate the prediction error. The accurately calculated adsorption energy is then substituted into the formula for calculating the recovery time τ to verify whether it is still within the ideal recovery time range.

[0187] 2. Sensitivity (S or λ)

[0188] (1) For resistive sensing materials, the sensitivity can be evaluated by calculating the change in the material's band gap (Eg) before and after the adsorption of decomposition product gases, and the sensitivity S is required to be higher than a certain threshold. For example:

[0189] S = |(G-G0) / G0|;

[0190] Among them, G0 and G are the electrical conductances of the third candidate sensing material in the clean state and after adsorbing gas, respectively.

[0191] (2) For FET-type sensing materials, a simulation model of the FET-type sensing material can be constructed.

[0192] The simulation model of the FET-type sensing material is a theoretical simulation device built to evaluate the field-effect transistor sensing performance of the third candidate sensing material. The simulation model uses a monolayer of the third candidate sensing material as the conductive channel, constructs highly doped source and drain electrodes at both ends, and adopts a back-gate structure (including gate dielectric and gate electrode) to simulate gate voltage regulation.

[0193] Based on a calculation method coupled with density functional theory and non-equilibrium Green's function, the electronic transport characteristics of this simulation model are simulated to obtain its drain current-gate voltage transfer characteristic curve, such as... Figure 5 As shown, the key parameter of subthreshold slope is then extracted.

[0194] By comparing the changes in the subthreshold slope before and after the adsorption of decomposition product gases, the sensitivity (λ) of the material as a FET-type gas sensor can be quantitatively calculated.

[0195] Sensitivity λ is defined as the relative rate of change of the subthreshold slope (SS) of the device before and after adsorption of the decomposition product gas. It requires that λ for the decomposition product gas be higher than a set threshold, i.e.:

[0196] λ=|SS * -SS0| / SS0;

[0197] Among them, SS0 and SS * These represent the subthreshold slopes of the device in a clean state and after adsorbing gas, respectively.

[0198] 3. Selectivity (θ and Δθ)

[0199] (1) Calculate the sensitivity ratio of the decomposition product gas relative to other coexisting interfering gases (including other decomposition products and background gases), i.e.:

[0200] θ=λ target / λ interfering ;

[0201] Where, λ target λ represents the sensitivity of the third candidate sensing material to the decomposition product gases. interfering The sensitivity of the third candidate sensing material to a certain interfering gas.

[0202] (2) To further quantify the overall selectivity advantage, the average selectivity difference (Δθ) is introduced, which is the ratio of the average sensitivity of the decomposition product gas to the average sensitivity of the interfering gas. The larger Δθ is, the stronger the selectivity of the material for the functional groups of the decomposition product gas. The formula for calculating Δθ is:

[0203] ;

[0204] in, The average sensitivity of the third candidate sensing material to the decomposition product gases. The average sensitivity of the third candidate sensing material to a certain interfering gas.

[0205] 4. Repeatability

[0206] Repeatability was assessed using both maximum protrusion distance (Δl) and artificial inductively coupled molecular dynamics (AIMD) simulations. The requirement was that gas adsorption should not cause irreversible and significant structural deformation of the sensing material or the gas molecules themselves; that is, the Δl value must be within a reasonable range. The joint assessment method was as follows:

[0207] (1) Static structural distortion analysis

[0208] Calculate the maximum local deformation (Δl) of the substrate material caused by gas adsorption. This parameter quantifies the change in the protrusion distance of the doped atom or active site at its vertical position before and after adsorption. Δl must be less than a set threshold (e.g., 1.9 Å) to ensure that adsorption does not cause irreversible damage to the substrate structure.

[0209] (2) Verification of dynamic thermodynamic stability

[0210] Ab initio molecular dynamics (AIMD) simulations were performed on the candidate “TM / g-C3N4- gas” adsorption system. The simulations were run at the application temperature (e.g., 300-500 K) for a sufficient duration (e.g., 5 picoseconds) to monitor energy and temperature fluctuations. The key examinations focused on whether the chemical bonds of the decomposition product gas molecules broke and whether the geometry of the substrate material remained intact. After the simulation, the final-state structure was geometrically optimized to verify whether it could be restored to the initial configuration.

[0211] Ultimately, the third candidate sensing material that simultaneously satisfies (1) Δl within a reasonable threshold and maintains structural integrity and stability in AIMD simulations is considered to have good reusability and is suitable for gas sensors that require long-term stable operation.

[0212] 5. Stability

[0213] (1) Calculate the cluster energy E using the following formula. clus This is used to assess the likelihood of transition metal (TM) atoms agglomerating on the surface of the first candidate sensing material, preventing metal atom aggregation and deactivation.

[0214] E clus =E bind -E coh ;

[0215] Among them, E bind E represents the binding energy between the doped / embedded atoms and the first candidate sensing material, and the energy at which a TM atom is adsorbed on the substrate surface.bind =E TM / g-C3N4 -E g-C3N4 -E TM-single E TM / g-C3N4 E is the total energy of the system. g-C3N4 As the base energy, E TM-single These parameters, representing the energy of an isolated atom, are calculated using DFT; E coh E represents the cohesive energy, which is the average binding energy of each atom of TM in its substrate material. coh =E TM-bulk / N−E TM-single N is the number of atoms, E TM-bulk This represents the total energy of a bulk transition metal crystal.

[0216] If E clus <0 indicates that TM atoms tend to disperse and adsorb on the surface and are not easy to aggregate.

[0217] If E clus A value >0 indicates that TM atoms tend to aggregate to form clusters.

[0218] (2) The thermal stability of the structure is verified by performing molecular dynamics (AIMD) simulations at specific temperatures (e.g., 300 K, 500 K, etc.) for a period of time (e.g., 5-10 ps). The specific method for simulating thermal stability using molecular dynamics (AIMD) is as follows.

[0219] ① Simulation settings

[0220] Using the geometrically optimized stable adsorption configuration as the initial structure, ab initio molecular dynamics simulations were performed in the NVT ensemble. Target temperatures were set (e.g., 300 K and 500 K); a Nosé-Hoover thermostat was used for temperature control; the time step was set to 1.0 femtosecond, and the total simulation time was 5 picoseconds (i.e., 5000 steps).

[0221] ② Calculation parameters and process monitoring

[0222] During the simulation, the electronic structure calculation at each step uses parameters consistent with those of static density functional theory (such as projected fused wave pseudopotential, PBE functional, etc.) and ensures self-consistent iterative convergence of the electronic structure.

[0223] During the simulation, the total energy, instantaneous temperature, and the motion trajectories of all atoms in the system are recorded in real time.

[0224] ③ After the simulation is completed, the thermal stability is evaluated and analyzed through the following steps.

[0225] a. Analyze the curves of total energy versus temperature over time to confirm that they fluctuate steadily and within a bounded range around the average value, indicating that the system has reached thermal equilibrium.

[0226] b. Visually inspect the entire simulation trajectory to confirm that the key chemical bonds of the decomposition product gas molecules have not been broken and that the decomposition product gas molecules have not dissociated or decomposed from the material surface; at the same time, confirm that the lattice structure of the substrate material is intact and that the doped atoms have not migrated or detached.

[0227] c. Extract the structure at the end of the simulation and perform static geometric re-optimization. Compare the optimized structure with the initial adsorption configuration, ensuring they are essentially identical, to demonstrate that the adsorption system has good structural recoverability under thermal disturbance.

[0228] Final precise calculations show that the actual adsorption energy of Cd / g-C3N4 for SO2F2 is -0.871 eV, which is in high agreement with the prediction value of the optimal machine learning prediction model. This adsorption energy value ensures that it has a reasonable recovery time at the operating temperature (~400 K), meeting the requirements of reversible sensing.

[0229] The detection sensitivity (S) of Cd / g-C3N4 is as high as 2.224, indicating that adsorption can cause significant changes in electronic structure (such as changes in band gap).

[0230] Selective evaluation confirmed that Cd / g-C3N4 responded significantly more to the target gas SO2F2 than to other coexisting interfering gases (such as CF4, SO2, HF).

[0231] Binding energy calculations and AIMD simulations confirmed that the doped structure of Cd atoms on the g-C3N4 substrate has excellent thermodynamic stability and no tendency to agglomerate or dissociate.

[0232] Step 7: Output the final candidate sensing materials.

[0233] Conclusion: Cd / g-C3N4 is a high-performance, sensitive sensing material for detecting SO2 and F2 gases. Its band structure and density of electronic states (DOS) diagram are shown below. Figure 4 As shown.

[0234] 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 screening sensing materials for insulating gas decomposition products based on machine learning, characterized in that, Includes the following steps: Select the insulating gas to be detected and the corresponding decomposition product gas, as well as the first candidate sensing material; Several second candidate sensing materials are obtained by doping with several elements or embedding atoms into the central cavity of the first candidate sensing material. Several second candidate sensing materials and decomposition product gases are combined to form several sensing material-gas combinations; For each of the aforementioned sensing material-gas combinations, the most stable adsorption configuration of the decomposition product gas molecules is calculated, along with the corresponding charge transfer amount, electronic structure information, and feature descriptors, forming an initial dataset; the initial dataset is then divided into a training set and a test set. Select several initial machine learning prediction models, train them using the training set, evaluate them using the test set, and select the optimal machine learning prediction model. Using the optimal machine learning prediction model, high-throughput prediction and screening of the sensing material-gas combination are performed to obtain a third candidate sensing material; The third candidate sensing material is verified by DFT calculation and its performance index is evaluated to output the final candidate sensing material.

2. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 1, characterized in that, For each of the aforementioned sensing material-gas combinations, the method for calculating and obtaining the most stable adsorption configuration and adsorption energy of the decomposition product gas molecules is as follows: Twenty percent of the sensor material-gas combinations were randomly selected, and DFT calculations were performed using VASP. The calculations employed PAW pseudopotentials and PBE functionals, and the DFT-D3 method was introduced to correct for van der Waals forces. The cutoff energy was set to 520 eV, and the K-point mesh was optimized using a Γ-centered 3×3×1 Monkhorst-Pack mesh. The electronic self-consistent iteration energy convergence criterion was 10 eV. -5 eV, atomic force convergence threshold set to 0.02 eV / Å, 20 Å vacuum layer set in the vertical direction; For each sensing material-gas combination, calculate the most stable adsorption configuration of the decomposition product gas molecules and its adsorption energy E. ads Adsorption energy E ads The calculation formula is: AND ads =E total -AND material -AND gas ; Among them, E total E represents the total energy of the sensor material-gas combination after adsorption. material E represents the energy of the second candidate sensing material itself. gas The energy of the isolated decomposition product gas molecules.

3. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 1, characterized in that, The feature descriptors include material-side descriptors, gas-side descriptors, and structural deformation descriptors; The material side descriptors include: atomic radius, relative atomic mass, Pauli electronegativity, number of electrons in the outermost d / p orbitals, first ionization energy, electron affinity, d / p band center, oxide formation enthalpy, and fluoride formation enthalpy; The structural deformation descriptor includes a maximum protrusion distance, which is the maximum vertical displacement that protrudes the doped atom from the original plane of the first candidate sensing material. The gas-side descriptor includes a feature descriptor for the decomposition product gas molecules, a feature descriptor for the adsorbed atoms of the decomposition product gas molecules, a global physical descriptor, and a quantum chemical descriptor. The descriptors of the decomposition product gas molecules include molecular weight, average atomic radius, polar surface area, average electronegativity, lipid-water partition coefficient, number of hydrogen bond acceptors, and geometric descriptors based on the smooth overlapping atomic positions. The characteristic descriptors of the adsorbed atoms of the decomposition product gas molecules include atomic radius, relative atomic mass, Pauli electronegativity, number of outermost p orbital electrons, first ionization energy, and electron affinity. The quantum chemical descriptor is obtained based on the type and number of functional groups in the decomposition product gas molecules and the electronic contributions of key reacting atoms; the quantum chemical descriptor includes charge distribution descriptors, frontier orbital descriptors, and chemical reactivity descriptors.

4. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 3, characterized in that, The method for obtaining the quantum chemical descriptor is as follows: Based on the three-dimensional structure or linear identifier of the gas molecules in the decomposition products, specific functional groups can be identified. Multithermal coding is used to generate a functional group existence vector for each decomposition product gas molecule; a functional group counting feature is added to the functional group existence vector; Key functional groups were identified based on partial density of states analysis and frontline trajectory analysis. Atomic-level electronic structure analysis was performed on the screened sensitive gas adsorption systems to further identify key reaction atoms; For the identified key reaction atoms, quantum chemical descriptors are extracted to quantify the electronic behavior of the key reaction atoms.

5. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 1, characterized in that, The initial machine learning prediction model includes, but is not limited to, any one of: gradient boosting regression, random forest regression, support vector regression, k-nearest neighbor regression, kernel ridge regression, LASSO regression, artificial neural network, and Gaussian process regression.

6. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 1, characterized in that, The optimal machine learning prediction model is selected by training the model using the training set and evaluating it using the test set. The initial dataset is randomly divided into a training set and a test set at a ratio of 80% and 20% respectively; The hyperparameters of the initial machine learning prediction model are optimized on the training set using 5-fold cross-validation to minimize the root mean square error of the prediction. The model was retrained using optimized hyperparameters, and its performance was evaluated on an independent test set.

7. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 6, characterized in that, The method for optimizing the hyperparameters of the initial machine learning prediction model on the training set using 5-fold cross-validation is as follows: Calculate the Pearson correlation coefficient matrix among all feature descriptors in the training set, and remove highly correlated redundant features; Calculate the Spearman correlation coefficient between the remaining features and the adsorption energy target, and retain the features with the highest correlation to the target within each highly correlated feature group to obtain a feature subset; Model performance was evaluated through cross-validation, and hyperparameters were tuned using grid search or Bayesian optimization. Root mean square error and coefficient of determination were used as the main evaluation metrics to select models that met the requirements. The method for evaluating performance on independent test sets is as follows: The root mean square error and coefficient of determination between the adsorption energy predicted by the computational model and the actual DFT value; Plot a scatter plot of the predicted and actual values ​​to visually assess the consistency and dispersion of the predictions; The training and testing process of the model was repeated hundreds of times, with the initial dataset being randomly re-split each time. The optimal machine learning prediction model was determined by evaluating the mean and standard deviation of the root mean square error and the coefficient of determination.

8. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 1, characterized in that, The method for obtaining a third candidate sensing material by using the optimal machine learning prediction model to perform high-throughput prediction and screening of the sensing material-gas combination is as follows: Determine the set temperature and target recovery time range based on actual application requirements; Based on the upper and lower limits of the target recovery time range and the set temperature, substitute them into the following calculation formula to solve for the corresponding ideal adsorption energy range; t=n -1 exp(-E ads / (k B T )); Where ν is the frequency of attempts, E ads k is the absolute value of the adsorption energy. B Where is Boltzmann's constant, and T is the operating temperature; The predicted adsorption energies output by the optimal machine learning prediction model are sorted from strong to weak. It is determined whether the predicted adsorption energies fall within the ideal adsorption energy range. The second candidate sensing material corresponding to the predicted adsorption energy falling within the range of the ideal adsorption energy is selected as the third candidate sensing material.

9. The method for screening sensing materials for insulating gas decomposition products based on machine learning according to claim 1, characterized in that, The method for performing DFT calculations and performance evaluations on the third candidate sensing material to output the final candidate sensing material is as follows: The predicted adsorption energy of the third candidate sensing material was accurately calculated using DFT to verify whether it was still within the ideal recovery time range and to evaluate the prediction error. The performance indicators include sensitivity assessment, selectivity assessment, repeatability assessment, and stability assessment. Furthermore, for resistive sensing materials, the sensitivity is evaluated by calculating the change in the material's band gap Eg before and after the adsorption of decomposition product gases, and the sensitivity S is required to be higher than a predetermined threshold. The formula for calculating the sensitivity S is: S = |(G-G0) / G0|; Wherein, G0 and G are the electrical conductances of the third candidate sensing material in a clean state and after adsorbing and decomposing the gas products, respectively. Furthermore, for FET-type sensing materials, the electron transport characteristics are simulated using density functional theory coupled with a non-equilibrium Green's function to obtain the drain current-gate voltage transfer characteristic curve and extract the subthreshold slope. By comparing the change in the subthreshold slope before and after the adsorption of decomposition product gases, the sensitivity λ is calculated, and it is required that λ for the decomposition product gases be higher than a set threshold. The formula for calculating the sensitivity λ is: λ=|SS * -SS0| / SS0; Among them, SS0 and SS * These represent the subthreshold slopes of the device in a clean state and after adsorbing gas, respectively. Furthermore, for the selectivity, the sensitivity ratio θ of the decomposition product gas relative to other coexisting interfering gases is calculated according to the following formula; θ=λ target / l interfering ; Where, λ target λ represents the sensitivity of the sensing material to the target gas. interfering The sensitivity of the sensing material to a certain interfering gas; Furthermore, we introduce the average selectivity difference Δθ, which is calculated using the following formula: ； in, The average sensitivity of the third candidate sensing material to the decomposition product gases. The average sensitivity of the third candidate sensing material to a certain interfering gas; Furthermore, the repeatability is jointly evaluated using the maximum protrusion distance and molecular dynamics AIMD simulations; Furthermore, regarding the stability, the cluster energy E is calculated using the following formula. clus Assess the likelihood of aggregation of elements or atoms doped / embedded on the surface of the first candidate sensing material; AND clus =E bind -AND coh ; Among them, E bind E represents the binding energy between the doped / intercalated element or atom and the first candidate sensing material. coh It is cohesive energy; If the cluster energy E clus <0 indicates that the doped / embedded elements or atoms tend not to aggregate; If the cluster energy E clus >0 indicates that the doped / embedded elements or atoms tend to aggregate to form clusters; The thermal stability of the third candidate sensing material structure at a predetermined temperature was simulated using molecular dynamics AIMD.

Images