A molecular screening and design method for additives to optimize the performance of ester-based insulating oils

Abstract

The present invention discloses a method for screening and designing additive molecules for optimizing the performance of ester-based insulating oil, and relates to the technical field of electrical insulating materials. In view of the problem that ester-based insulating oil is prone to generate rapid streamers under high-intensity lightning strikes, resulting in insufficient breakdown voltage performance and that it is time-consuming and laborious to find suitable additives through experiments, the present invention proposes a method for designing additive molecules in combination with a deep molecular generation model. The method comprises: obtaining molecular data with target electronic properties; using fragments of the molecular structure as input to the model, and designing new additive molecular structures using a molecular generation model; performing molecular screening according to target characteristics; and verifying the performance of the screened molecules through streamer discharge simulation. The present invention utilizes a deep autoregressive neural network to realize the generative design of ester-based insulating oil additive molecules, significantly improving the efficiency of molecular screening and providing a new technical route for improving the breakdown voltage performance of ester-based insulating oil.

Description

Technical Field

[0001] The invention relates to a method for screening and designing additive molecules for optimizing the performance of ester-based insulating oil, and belongs to the technical field of electrical insulation materials. Background Art

[0002] With the rapid development of modern power systems toward ultra-high voltage (UHV) and extra-high voltage (EHV), power equipment is placing increasingly stringent demands on the electrical performance of insulation materials. As a crucial insulating medium within power transformers, the performance of insulating oil is directly related to the safe operation and stability of the transformer. Due to the advantages of ester-based insulating oil, such as easy degradation, high flash point, and oxidation resistance, the option of replacing the traditional insulating medium (mineral oil) in power transformers with ester-based insulating oil has been increasingly accepted by researchers, and this approach has achieved certain results in engineering applications.

[0003] However, ester-based insulating oils exhibit significant insulation deficiencies when exposed to high-intensity lightning impulse voltages and severe transient overvoltages. In particular, their breakdown voltage is significantly lower than that of mineral oil under conditions of non-uniform electric fields and long oil gaps, severely restricting their widespread adoption in large power transformers. To address this critical issue, it is first necessary to combine existing streamer discharge research on ester-based insulating oils to gain a deeper understanding of their lightning impulse breakdown mechanisms.

[0004] Existing experimental and simulation studies have shown that ester-based insulating oils are more susceptible to rapid streamers during streamer discharges under conditions of non-uniform electric fields and long oil gaps. These rapid streamers have been shown to be a key factor in power transformer insulation failure. To overcome this problem, it is crucial to identify methods and mechanisms to suppress the generation of rapid streamers in ester-based insulating oils.

[0005] On this basis, some studies have further found that the photoionization mechanism is an important reason for the transformation of slow streamers into fast streamers. Studies have found that the photoionization process can be effectively suppressed by adding molecular additives with low ionization energy (IP) or low first excited state (S1) energy. Such molecules can preferentially absorb photon energy, thereby suppressing the photoionization process of surrounding ester-based insulating oil molecules and ultimately curbing the formation of fast streamers. Early studies have confirmed that the addition of ultraviolet light absorbers can indeed improve the breakdown voltage performance of ester-based insulating oil, which points the way to solving the problem. Therefore, there is an urgent need to find more efficient ways to suppress the photoionization mechanism during streamer discharge, thereby improving the lightning impulse breakdown voltage of ester-based insulating oil.

[0006] However, among the many possible additives, how to efficiently find the optimal additive molecules and determine which molecular structural features and functional groups are most effective in suppressing the photoionization process remains a major challenge.

[0007] Traditional molecular screening methods rely primarily on trial and error, requiring researchers to synthesize and test different candidate molecules one by one. This approach is not only time-consuming, labor-intensive, and costly, but also makes it difficult to systematically explore the vast molecular space. Even experienced researchers struggle to accurately predict which molecular structures will exhibit desirable low ionization energy or low excited state properties based solely on intuition. With the rapid development of machine learning and artificial intelligence technologies in materials science, deep molecular generation models offer an innovative solution to this problem.

[0008] This deep learning-based molecular generation model has unique advantages: First, it can automatically learn and understand the complex relationship between molecular structure and performance, especially the relationship between the molecular geometry and electronic structure and its target properties. Second, the model can generate candidate molecules with low IP or low S1 energy based on the set target properties, greatly improving the efficiency of molecular design. More importantly, by analyzing the large amount of molecular data generated by the model, we can identify the structural features and functional groups that are key to the target performance, thereby providing more in-depth theoretical guidance for additive synthesis. This data-driven approach not only significantly accelerates the material development cycle, but also helps us establish a systematic understanding of the quantitative-structure property relationship of molecules, providing strong technical support for the development of a new generation of high-performance ester-based insulating oil additives. Summary of the Invention

[0009] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for screening and designing additive molecules for optimizing the performance of ester-based insulating oil. The method uses machine learning technology to realize the screening and design of ester-based insulating oil additive molecules, accelerate the material design process, and improve the breakdown voltage of ester-based insulating oil.

[0010] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0011] The present invention provides a method for screening and designing additive molecules for optimizing the performance of ester-based insulating oil, comprising:

[0012] Obtain molecular fragment data with target characteristics;

[0013] Taking molecular fragment data as input, multiple complete additive molecular structures are output based on the molecular generation model;

[0014] Screening multiple complete additive molecular structures according to target properties to obtain a screened additive molecular structure;

[0015] The molecular structure of the screened additives was verified by streamer discharge simulation analysis to obtain the molecular structure of the additive for optimizing the performance of ester-based insulating oil.

[0016] Furthermore, the target characteristics include low first excited state energy and low ionization energy; the molecular fragment data includes structural information and attribute information of the molecular fragment, the structural information includes the atom type and atomic position contained in the molecular fragment, and the attribute information includes structural attributes and electronic attributes.

[0017] Furthermore, the molecular generation model uses a deep autoregressive neural network, which includes a first processing route and a second processing route, wherein the first processing route is used to process the structural information to obtain an atomic feature vector, and the second processing route is used to process the attribute information to obtain an attribute feature vector;

[0018] The output ends of the first processing route and the second processing route are both connected to the input end of the atom type prediction layer, the atom type prediction layer is used to predict the next atom type according to the atom feature vector and the attribute feature vector to obtain the next atom type probability distribution function, the output end of the atom type prediction layer is connected to the atom type sampling layer and the atom type embedding layer in sequence, the atom type sampling layer is used to randomly sample the next atom type according to the next atom type probability distribution function, and the atom type embedding layer is used to convert the next atom type into an atom type feature vector;

[0019] The atomic feature vector and the atomic type feature vector are sequentially multiplied and added and then input into the atomic distance prediction layer. The atomic distance prediction layer is used to predict the next atomic position according to the atomic feature vector and the atomic type feature vector to obtain the next atomic distance probability distribution function. The output end of the atomic distance prediction layer is connected to the input end of the atomic position sampling layer. The atomic position sampling layer is used to randomly sample the next atomic position according to the next atomic distance probability distribution function to generate the next molecular fragment, and input the next molecular fragment into the first processing route and the second processing route for continuous iteration until the complete additive molecular structure is obtained.

[0020] Furthermore, the first processing route includes an embedding layer, an atom-pair distance weighting layer, and a feature processing layer connected in sequence, wherein the embedding layer is used to extract atomic feature vectors from structural information, the atom-pair distance weighting layer is used to perform weighted interaction on atomic feature vectors according to interatomic distances, and the feature processing layer is used to process the relationship between atoms and their surroundings, and includes multiple feature interaction layers and feature mixing layers connected in sequence;

[0021] The second processing route includes an attribute embedding module and a multi-layer perceptron connected in sequence, wherein the attribute embedding module is used to convert structural attributes and electronic attributes into structural attribute feature vectors and electronic attribute feature vectors respectively, and the multi-layer perceptron is used to integrate the structural attribute feature vectors and the electronic attribute feature vectors to obtain an attribute feature vector.

[0022] Furthermore, the expression of the atom pair distance weighted layer is:

[0023]

[0024] in, represents the atomic characteristics of the i-th atom, represents the atomic characteristics of the jth atom adjacent to the i-th atom, represents the distance between atom i and its neighboring atom j, represents the distance cutoff function, represents the radial basis function, Indicates network filtering;

[0025] The expression of the feature processing layer is:

[0026] ;

[0027] in, represents the atomic features of the i-th atom after feature processing, 、 are all weight matrices, 、 are all bias terms, represents the activation function, ;

[0028] The expression of the attribute embedding module is:

[0029]

[0030] ;

[0031] in, represents the structural attribute feature vector, represents multi-layer perceptron processing, represents the concentration weight of atom type a in the existing additive molecule fragment, Embedding vector representing atom type a; represents the electronic property eigenvector, represents electronic properties, represents the minimum value of electronic properties, represents the interval parameter of the radial basis function;

[0032] The expression of the multi-layer perceptron is:

[0033] ;

[0034] in, represents the attribute feature vector;

[0035] Furthermore, the expression of the atom type prediction layer is:

[0036]

[0037] ;

[0038] in, represents the weight factor of the j-th candidate atom type, represents the sum of atomic features in the generated molecular fragments, represents the probability distribution function of the next atom type;

[0039] The expression of the atomic position prediction layer is:

[0040]

[0041]

[0042] in, represents the weight factor of the j-th candidate atomic position, represents the atomic number of the next atom, express The embedding vector of represents the attribute feature vector, represents the probability distribution function of the next atom position.

[0043] Furthermore, the molecular generation model further includes an attention module, wherein the input end of the attention module is connected to the output end of the feature processing layer;

[0044] The attention module includes a multi-head attention unit, a self-attention unit, and a cross-attention unit connected in sequence. The multi-head attention unit is used to project the atomic feature head into multiple representation subspaces. The self-attention unit is used to enable each atom to obtain information about other atoms. The cross-attention unit is used to associate information from different feature spaces to achieve multi-dimensional fusion of atomic features.

[0045] The expression of the attention module is:

[0046] ;

[0047] in, represents the atomic features of the i-th atom after feature processing, represents the atomic features of the i-th atom after being processed by the attention module, represents the query matrix, represents the bond matrix, represents the value matrix, Represents the feature dimension.

[0048] Furthermore, the method further includes pre-training the molecular generation model, wherein the pre-training method includes:

[0049] a. Obtain multiple known complete molecular structures;

[0050] b. Split multiple known complete molecular structures into multiple molecular fragments and form the training set data;

[0051] c. Use the training set data as input to train the molecular generation model, output the predicted complete molecular structure, calculate the loss function, and use the loss function to adjust the model parameters;

[0052] d. Repeat steps c to d until the loss function drops to a preset range or reaches a preset number of iterations, obtaining a pre-trained molecular generation model.

[0053] The loss function adopts the cross entropy of type prediction and distance prediction for predicting the complete molecular structure.

[0054] Furthermore, the molecular structures of the additives are screened according to the target characteristics to obtain the screened molecular structures of the additives, including:

[0055] The energy prediction model is used to predict the first excited state energy and ionization energy of multiple additive molecular structures to obtain prediction results;

[0056] Based on the prediction results, multiple additive molecular structures are roughly screened to retain effective additive molecular structures that meet the target properties;

[0057] Perform quantum chemical calculations on the first excited state energy and ionization energy of the effective additive molecular structure that meets the target characteristics to obtain the calculation results;

[0058] The retained additive molecular structures are finely screened according to the calculation results, and the additive molecular structures with target properties within a preset range are retained and used as the screened additive molecular structures.

[0059] Furthermore, the streamer discharge simulation analysis and verification of the screened additive molecular structure is performed to obtain the additive molecular structure for optimizing the performance of the ester-based insulating oil, including:

[0060] The streamer discharge simulation analysis is performed on the molecular structure of the screened additives to obtain the streamer discharge simulation analysis results;

[0061] According to the results of streamer discharge simulation analysis, the molecular structure with the best effect of slowing down the streamer velocity is retained as the molecular structure of the additive for optimizing the performance of ester-based insulating oil.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] This invention uses machine learning technology to achieve automated screening of additive molecules, accelerating the material development process, avoiding the tedious process of traditional experimental screening, and reducing development costs. Through systematic, data-driven molecular design, it can effectively ensure that the resulting additives have excellent electrical properties, providing a green and efficient solution for material design.

[0064] The additive molecules screened by the method of the present invention can effectively improve the breakdown voltage of ester-based insulating oil, shorten the development cycle of ester-based insulating oil, and provide a more sustainable material option. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 Schematic diagram of a process for screening and designing additive molecules for optimizing the performance of ester-based insulating oil according to an embodiment of the present invention;

[0066] Figure 2 A schematic structural diagram of a molecular generation model in one embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of a partially complete molecular structure used for pre-training in Example 1 of the present invention;

[0068] Figure 4 This is a schematic diagram of a partially complete additive molecular structure output by the molecular generation model in Example 1 of the present invention;

[0069] Figure 5 This is a schematic diagram comparing the simulation results of COMSOL simulation to verify the development of streamer discharge in Example 1 of the present invention. The left side shows the simulation effect of the photoionization model after adding certain additive molecules, and the right side shows the simulation effect of the original photoionization model. DETAILED DESCRIPTION

[0070] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0071] Example 1

[0072] like Figure 1 As shown, an embodiment of the present invention provides an additive molecule screening and design method for optimizing the performance of ester-based insulating oil, comprising the following steps:

[0073] First, obtain the existing additive molecular fragment data

[0074] In this example, quantum chemical property data for the molecule were collected from databases such as QM8 and VERDE Materials DB. Published literature covering the latest research results related to the molecule's first excited state (S1) energy and ionization potential (IP) was obtained through platforms such as Web of Science, Google Scholar, and IEEE. Furthermore, quantum chemical computational tools such as Gaussian and ORCA were used for molecular structure optimization and energy calculations. Experimental methods such as ultraviolet-visible spectroscopy (UV-Vis) and photoelectron spectroscopy (PES) were used to obtain directly measured S1 and IP data.

[0075] The Atomistic Simulation Environment (ASE) database was established through a multi-channel data collection process. The ASE database is specifically designed to store key information about molecular structures, such as atom types, atomic coordinates, and related physical and electronic properties, providing basic input data for molecular model generation.

[0076] The existing additive molecular fragment data includes structural information and attribute information of the existing molecular fragments. The structural information includes the types and positions of atoms contained in the molecular fragments, and the attribute information includes structural attributes and electronic attributes.

[0077] Next, build a molecular generation model

[0078] In this embodiment, the molecular generation model adopts a deep autoregressive neural network, which can gradually predict the atomic type and atomic position of each atom in the molecule and continuously iterate to generate a complete molecular structure.

[0079] like Figure 2 As shown, the molecular generation model includes two routes. The first processing route is used to process the structural information to obtain the atomic feature vector, and the second processing route is used to process the attribute information to obtain the attribute feature vector. Specifically:

[0080] The first processing route includes an embedding layer, an atom-pair distance weighting layer, and a feature processing layer connected in sequence. The embedding layer is used to extract atomic feature vectors from structural information. Then, the atom-pair distance weighting layer performs weighted interaction on the atomic feature vectors. The processing expression is:

[0081]

[0082] in, represents the atomic characteristics of the i-th atom, represents the atomic characteristics of the jth atom adjacent to the i-th atom, represents the distance between atom i and its neighboring atom j, represents the distance cutoff function, represents the radial basis function, Indicates network filtering.

[0083] The feature processing layer is used to process the relationship between atoms and their surrounding environment. It includes multiple feature interaction layers and feature mixing layers connected in sequence. Its processing expression is:

[0084]

[0085] in, represents the atomic features of the i-th atom after feature processing, 、 are all weight matrices, 、 are biased, represents the activation function, .

[0086] The atomic feature of the i-th atom after feature processing is the final atomic feature of the i-th atom, and the atomic features of all atoms are combined to form .

[0087] The second processing route includes an attribute embedding module and a multi-layer perceptron (MLP) connected in sequence. The attribute embedding module is used to generate structural attribute feature vectors and electronic attribute feature vectors for structural attributes and electronic attributes respectively. The expressions for processing structural attributes and electronic attributes are:

[0088]

[0089]

[0090] in, represents the structural attribute feature vector, represents multi-layer perceptron processing, represents the concentration weight of atom type a in the existing additive molecule fragment, Embedding vector representing atom type a; represents the electronic property eigenvector, represents electronic properties, represents the minimum value of electronic properties, Represents the spacing parameter of the radial basis function.

[0091] Next, the multi-layer perceptron is used to integrate the structural attribute feature vector and the electronic attribute feature vector to obtain the attribute feature vector, and its processing expression is:

[0092]

[0093] in, Represents the attribute feature vector.

[0094] The output end of the first processing route is divided into two paths. In order to further improve the model's ability to characterize atomic features, an attention module is also provided in this embodiment. The attention module is provided after the output end of the first processing route, that is, both paths need to pass through the attention module before starting.

[0095] The attention module consists of a multi-head attention unit, a self-attention unit, and a cross-attention unit, all connected in sequence. The multi-head attention unit projects atomic features into multiple representation subspaces, enabling parallel attention to different aspects of the molecular structure. This mechanism is particularly well-suited for processing atomic environments with complex structural features. The self-attention unit constructs a global context-aware layer, enabling each atom to access information about all other atoms in the molecule, significantly enhancing the model's understanding of interatomic relationships. The cross-attention unit achieves multi-dimensional fusion of atomic feature representations by correlating information from different feature spaces, enhancing the model's ability to represent the local atomic environment.

[0096] The introduction of these attention mechanisms not only improves the model’s ability to capture the local chemical environment of atoms, but also maintains a global perception of the relationships between atoms during attention processing. The processing expression of the attention module is:

[0097]

[0098] in, represents the atomic features of the i-th atom after feature processing, represents the atomic features of the i-th atom after being processed by the attention module, represents the query matrix, represents the bond matrix, represents the value matrix, Represents the feature dimension.

[0099] The attention module calculates the query-key similarity and dynamically assigns feature importance weights, thus achieving selective enhancement of atomic features.

[0100] The output end of the first processing route inputs the atomic feature vector and the attribute feature vector obtained by the second processing route into the atom type prediction layer. The atom type prediction layer is used to predict the next atom type and obtain the next atom type probability distribution function. Its processing expression is:

[0101]

[0102]

[0103] in, represents the weight factor of the j-th candidate atom type, represents the sum of atomic features in the generated molecular fragments, Represents the next atom type probability distribution function.

[0104] The next atom type probability distribution function includes multiple candidate atoms and their probabilities. The output end of the atom type prediction layer is connected to the input end of the atom type sampling layer. The atom type sampling layer randomly samples an atom type from multiple candidate atoms as the next atom type according to the next atom type probability distribution function. In this embodiment, the random sampling is not necessarily the atom type corresponding to the maximum probability, but the atom type with the maximum probability is most likely to be sampled, which makes the selection of the next atom type more diverse.

[0105] After sampling by the atom type sampling layer, it is input into the atom type embedding layer. The atom type embedding layer is used to convert the next atom type obtained by sampling into an atom type feature vector to facilitate subsequent model processing.

[0106] At the output of the first processing route, the other path performs dot multiplication and addition on the atomic feature vector and the atomic type feature vector, and then inputs them into the atomic distance prediction layer. The atomic distance prediction layer is used to predict the next atomic position and obtain the next atomic distance probability distribution function. Its processing expression is:

[0107]

[0108]

[0109] in, represents the weight factor of the j-th candidate atomic position, represents the atomic number of the next atom, express The embedding vector of represents the attribute feature vector, represents the probability distribution function of the next atom position.

[0110] The output of the atomic distance prediction layer is connected to the input of the atomic position sampling layer. The atomic position sampling layer then randomly samples the next atomic position based on the next atomic distance probability distribution function. At this point, the atomic type and position of the next atom are determined, and the next molecular fragment with an additional atom is generated. This next molecular fragment is then fed back into the first and second processing routes, and the process continues iteratively until the complete molecular structure is generated.

[0111] Then, the constructed molecular generation model needs to be pre-trained. The pre-training method includes:

[0112] a. Obtain multiple known complete molecular structures;

[0113] b. Split multiple known complete molecular structures into multiple molecular fragments and form the training set data;

[0114] c. Use the training set data as input to train the molecular generation model, output the predicted complete molecular structure, and calculate the loss function;

[0115] d. Repeat steps c-d until the loss function drops to a preset range or reaches a preset number of iterations. The loss function uses the cross entropy of the type prediction and distance prediction of the complete molecular structure to obtain a pre-trained molecular generation model.

[0116] Next, the existing additive molecular fragment data is used as input, and multiple additive molecular structures are output based on the pre-trained molecular generation model.

[0117] Multiple additive molecular structures are screened according to target characteristics. In this embodiment, the molecular structure is an additive for optimizing the performance of ester-based insulating oil, so the target characteristics that the molecular structure needs to meet are low first excited state energy and low ionization energy.

[0118] Screening is divided into coarse and fine screening. In the coarse screening stage, an energy prediction model is used to predict the first excited state and ionization energy of multiple additive molecular structures. The energy prediction model can quickly predict the first excited state energy and ionization energy of each molecule based on information such as the molecule's structural characteristics, atomic type, and position. The prediction results are used to assess whether the additive's molecular structure meets the target properties and eliminate invalid molecules (an intact molecule is not necessarily a valid molecule, but a valid molecule is always intact). Coarse screening can greatly reduce the number of molecules requiring in-depth analysis, providing an efficient foundation for subsequent fine screening.

[0119] The introduction of predictive models makes the virtual screening process more automated and rapid, eliminating the reliance on extensive experimental measurements. Through the evaluation of predictive models, molecules with potentially ideal properties can be initially screened, improving the efficiency of the entire molecule generation and optimization process.

[0120] During the fine screening process, quantum chemical calculations are performed on the molecules that meet the target characteristics obtained through the coarse screening, and their properties are evaluated with higher precision, especially recalculating key electronic properties such as the first excited state energy and ionization energy. In this embodiment, the quantum chemical calculations use appropriate functionals and basis sets in Gaussian software. Through more precise quantitative methods, it is ensured that the molecules ultimately retained can meet the strict performance requirements in practical applications.

[0121] During the fine-scale screening process, quantum chemical calculations not only accurately measure the electronic properties of molecules but also reveal their stability and reactivity under different conditions, providing reliable data support for the final molecule selection. This rigorous screening approach provides practical validation for model-generated molecules with low S1 energy or low IP, helping to further narrow the candidate pool and lay a solid foundation for subsequent experimental verification and application.

[0122] Finally, the molecules retained after screening are subjected to streamer discharge simulation analysis and verification. In this embodiment, COMSOL simulation verification is used. Specifically, the photoionization model is used to simulate the performance of the molecule as an ester-based insulating oil additive in a streamer environment, and its effect on the streamer velocity is evaluated to verify the potential role of the additive molecule under actual discharge conditions to confirm whether it can effectively reduce the streamer velocity, thereby improving the electrical properties of the ester-based insulating oil. Finally, the molecular structure that is retained with the best effect of slowing down the streamer velocity is the additive molecular structure used to optimize the performance of the ester-based insulating oil.

[0123] After determining the specific molecular structure of the additive, we enter the synthesis and preparation stage of the additive. First, we select a suitable chemical synthesis route based on the structural characteristics of the molecule, and control the synthesis conditions (such as temperature, pressure, catalyst, etc.) to ensure the purity and structural integrity of the molecule.

[0124] After synthesis, the molecule undergoes necessary purification and verification, including conventional analytical methods such as nuclear magnetic resonance (NMR) and mass spectrometry (MS), to confirm the accuracy of the molecular structure. The resulting molecule will be used as an additive in subsequent ester-based insulating oil blending experiments, providing a foundation for improving the electrical properties of ester-based insulating oil.

[0125] The synthesized additives were added to the ester-based insulating oil matrix at varying concentrations and thoroughly mixed to ensure uniform distribution. The mixed samples were then preliminarily tested to observe their basic physical properties and stability, ensuring the mixture's uniformity and basic compatibility met expectations.

[0126] Then, using a standard breakdown voltage tester—an impulse voltage generator—the breakdown voltage of ester-based insulating oil samples mixed with different types and concentrations of additives was measured. During the experiment, the impulse voltage generator applied positive and negative voltage pulses to the samples. The samples were tested under extreme conditions, including high temperature, high voltage, and simulated lightning strikes, and the breakdown voltage of each sample was recorded.

[0127] After all tests are completed, detailed experimental data for various additives is recorded, including key parameters such as concentration, mixing ratio, and breakdown voltage. By comparing and analyzing the performance data of different additives, the additive combination that best improves the breakdown voltage of ester-based insulating oil is selected, with particular attention paid to its environmental performance, such as whether it is non-toxic, easily soluble in ester-based insulating oil, and cost-effective.

[0128] The following is a detailed explanation based on specific implementation data.

[0129] Figure 3 are some known complete molecular structures used to train the molecular generation model in this embodiment. Figure 4 The molecular structures of some additives were obtained through screening in this example. Figure 5 The simulation effect diagram is verified by COMSOL simulation. Figure 5 The left side shows the simulation results of the photoionization model after adding a specific additive molecule, and the right side shows the simulation results of the original model. From the comparison results, it can be clearly observed that the development length of the streamer branch is significantly shortened after adding the specific molecular additive, which further demonstrates that the addition of low ionization energy or low first excited state additives to ester-based insulating oil has a certain effect on suppressing the occurrence of fast streamers in ester-based insulating oil.

[0130] The effect of adding the finalized additive molecules to the ester-based insulating oil FR3 on improving the breakdown voltage of the ester-based insulating oil is shown in Table 1.

[0131]

[0132] In lightning impulse breakdown tests, the "gap" (also called the "oil gap" or "gap") typically refers to the distance between the test electrodes. In a standard pin-plate electrode configuration, this specifically refers to the vertical distance between the pin-tip electrode and the plate electrode. As shown in Table 1, the additive-containing insulating oil exhibits a higher breakdown voltage than the unadditive-containing insulating oil in most cases, effectively demonstrating that the final additive molecule can effectively improve the breakdown voltage performance of ester-based insulating oil.

[0133] Example 2:

[0134] Based on Example 1, the difference between this embodiment and Example 1 is that the attention module is not included.

[0135] Example 3:

[0136] Based on Example 1, the difference between this embodiment and Example 1 is that the attention module only includes multi-head attention units.

[0137] Example 4:

[0138] Based on Example 1, the difference between this embodiment and Example 1 is that the attention module only includes cross-attention units.

[0139] Example 5:

[0140] Based on Example 1, the difference between this embodiment and Example 1 is that the attention module only includes a self-attention unit.

[0141] The 1000 molecular structures generated by the molecular generation models of Examples 2 to 5 were compared, and the results are shown in Table 2.

[0142]

[0143] As can be seen from Table 1, after introducing different attention mechanisms, the effectiveness, uniqueness and true molecule generation rate of the molecular structures generated by the model have been improved to varying degrees.

[0144] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for screening and designing additive molecules for optimizing the performance of ester-based insulating oil, characterized in that: include: Acquire molecular fragment data with target characteristics; Taking molecular fragment data as input, multiple complete additive molecular structures are output based on the molecular generation model; Screening multiple complete additive molecular structures according to target properties to obtain a screened additive molecular structure; The molecular structure of the screened additives was verified by streamer discharge simulation analysis to obtain the molecular structure of the additive for optimizing the performance of ester-based insulating oil; The molecular generation model uses a deep autoregressive neural network, which includes a first processing route and a second processing route, wherein the first processing route is used to process the structural information to obtain an atomic feature vector, and the second processing route is used to process the attribute information to obtain an attribute feature vector; The output ends of the first processing route and the second processing route are both connected to the input end of the atom type prediction layer, the atom type prediction layer is used to predict the next atom type according to the atom feature vector and the attribute feature vector to obtain the next atom type probability distribution function, the output end of the atom type prediction layer is connected to the atom type sampling layer and the atom type embedding layer in sequence, the atom type sampling layer is used to randomly sample the next atom type according to the next atom type probability distribution function, and the atom type embedding layer is used to convert the next atom type into an atom type feature vector; The atomic feature vector and the atomic type feature vector are sequentially multiplied and added and then input into the atomic distance prediction layer. The atomic distance prediction layer is used to predict the next atomic position according to the atomic feature vector and the atomic type feature vector to obtain the next atomic distance probability distribution function. The output end of the atomic distance prediction layer is connected to the input end of the atomic position sampling layer. The atomic position sampling layer is used to randomly sample the next atomic position according to the next atomic distance probability distribution function to generate the next molecular fragment, and input the next molecular fragment into the first processing route and the second processing route for continuous iteration until the complete additive molecular structure is obtained.

2. The additive molecule screening and design method for optimizing the performance of ester-based insulating oil according to claim 1, characterized in that: The target properties include low first excited state energy and low ionization energy; The molecular fragment data includes structural information and attribute information of the molecular fragment, wherein the structural information includes the atomic type and atomic position of the atoms contained in the molecular fragment, and the attribute information includes structural attributes and electronic attributes.

3. The additive molecule screening and design method for optimizing the performance of ester-based insulating oil according to claim 1, characterized in that: The first processing route includes an embedding layer, an atom-pair distance weighting layer, and a feature processing layer connected in sequence, wherein the embedding layer is used to extract atomic feature vectors from structural information, the atom-pair distance weighting layer is used to perform weighted interaction on atomic feature vectors according to interatomic distances, and the feature processing layer is used to process the relationship between atoms and their surroundings, and includes multiple feature interaction layers and feature mixing layers connected in sequence; The second processing route includes an attribute embedding module and a multi-layer perceptron connected in sequence, wherein the attribute embedding module is used to convert structural attributes and electronic attributes into structural attribute feature vectors and electronic attribute feature vectors respectively, and the multi-layer perceptron is used to integrate the structural attribute feature vectors and the electronic attribute feature vectors to obtain an attribute feature vector.

4. The additive molecule screening and design method for optimizing the performance of ester-based insulating oil according to claim 3, characterized in that: The expression of the atom pair distance weighted layer is: ； in, represents the atomic characteristics of the i-th atom, represents the atomic characteristics of the jth atom adjacent to the i-th atom, represents the distance between atom i and its neighboring atom j, represents the distance cutoff function, represents the radial basis function, Indicates network filtering; The expression of the feature processing layer is: ； in, represents the atomic features of the i-th atom after feature processing, 、 are all weight matrices, 、 are all bias terms, represents the activation function, ; The expression of the attribute embedding module is: ； ； in, represents the structural attribute feature vector, represents multi-layer perceptron processing, represents the concentration weight of atom type a in the existing additive molecule fragment, Embedding vector representing atom type a; represents the electronic property eigenvector, represents electronic properties, represents the minimum value of electronic properties, represents the interval parameter of the radial basis function; The expression of the multi-layer perceptron is: ； in, Represents the attribute feature vector.

5. The additive molecule screening and design method for optimizing the performance of ester-based insulating oil according to claim 1, characterized in that: The expression of the atom type prediction layer is: ； ； in, represents the weight factor of the j-th candidate atom type, represents the sum of atomic features in the generated molecular fragments, represents the attribute feature vector, represents the probability distribution function of the next atom type; The expression of the atomic distance prediction layer is: ； ； in, represents the weight factor of the j-th candidate atomic position, represents the atomic number of the next atom, express The embedding vector of represents the attribute feature vector, represents the next atom distance probability distribution function.

6. The additive molecule screening and design method for optimizing the performance of ester-based insulating oil according to claim 3, characterized in that: The molecule generation model further includes an attention module, wherein an input end of the attention module is connected to an output end of the feature processing layer; The attention module includes a multi-head attention unit, a self-attention unit, and a cross-attention unit connected in sequence. The multi-head attention unit is used to project the atomic feature head into multiple representation subspaces. The self-attention unit is used to enable each atom to obtain information about other atoms. The cross-attention unit is used to associate information from different feature spaces to achieve multi-dimensional fusion of atomic features. The expression of the attention module is: ； in, represents the atomic features of the i-th atom after feature processing, represents the atomic features of the i-th atom after being processed by the attention module, represents the query matrix, represents the bond matrix, represents the value matrix, Represents the feature dimension.

7. The method for screening and designing additive molecules for optimizing the performance of ester-based insulating oil according to claim 1, characterized in that: The method also includes pre-training the molecular generation model, wherein the pre-training includes: a. Obtain multiple known complete molecular structures; b. Split multiple known complete molecular structures into multiple molecular fragments and form the training set data; c. Use the training set data as input to train the molecular generation model, output the predicted complete molecular structure, calculate the loss function, and use the loss function to adjust the model weight parameters; d. When the loss function drops to a preset range or reaches a preset number of iterations, a pre-trained molecular generation model is obtained; otherwise, step c is executed until the loss function drops to a preset range or reaches a preset number of iterations; The loss function adopts the cross entropy of type prediction and distance prediction for predicting the complete molecular structure.

8. The method for screening and designing additive molecules for optimizing the performance of ester-based insulating oil according to claim 1, characterized in that: The method further comprises screening a plurality of additive molecular structures according to target characteristics to obtain the screened additive molecular structures, including: The energy prediction model is used to predict the first excited state energy and ionization energy of multiple additive molecular structures to obtain prediction results; Based on the prediction results, multiple additive molecular structures are roughly screened to retain effective additive molecular structures that meet the target properties; Perform quantum chemical calculations on the first excited state energy and ionization energy of the effective additive molecular structure that meets the target characteristics to obtain the calculation results; The retained additive molecular structures are finely screened according to the calculation results, and the additive molecular structures with target properties within a preset range are retained and used as the screened additive molecular structures.

9. The additive molecule screening and design method for optimizing the performance of ester-based insulating oil according to claim 1, characterized in that: The streamer discharge simulation analysis and verification of the screened additive molecular structure is performed to obtain the additive molecular structure for optimizing the performance of the ester-based insulating oil, including: The streamer discharge simulation analysis is performed on the molecular structure of the screened additives to obtain the streamer discharge simulation analysis results; According to the results of streamer discharge simulation analysis, the molecular structure with the best effect of slowing down the streamer velocity is retained as the molecular structure of the additive for optimizing the performance of ester-based insulating oil.

Images