Prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning

Abstract

The invention discloses a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning. The prototype network DNTF eutectic prediction method comprises the following steps: obtaining a crystal compound meeting a preset DNTF related condition as a related crystal compound; constructing a training data set according to related crystal compounds; dividing the training data set into a support set and a query set, and performing comparative learning training on the constructed DNTF eutectic prediction network by using the support set and the query set; and predicting ligand molecules of the DNTF by using the trained DNTF eutectic prediction network model. According to the eutectic prediction method, a data set is constructed based on existing crystal structure data for learning, a reliable and effective eutectic prediction network model is established through a data driving method, DNTF crystals are helped to rapidly predict and recommend matched ligand molecules, the ligand molecule screening cost can be remarkably reduced, and the method is suitable for large-scale popularization and application. The application of the DNTF high-performance oxidizing agent in the field of solid propellants is promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of DNTF eutectic materials, and specifically relates to a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning. Background Art

[0002] The design, synthesis and application of high energy density materials are important means to improve the energy characteristics of propellants and explosives. Currently, CL-20 (hexanitrohexaazaisowurtzitane) is the most commonly used material in the field, but its high cost restricts its widespread application. Research in recent years has found that DNTF (3,4-dinitrofurazanyl furazan oxide) has the characteristics of high energy density, high detonation velocity and good stability. Most importantly, the performance of DNTF is close to that of CL-20, but its cost is only half of that of CL-20. Therefore, DNTF has a high research value and has great development potential in the field of high-performance oxidizers for solid propellants.

[0003] Studies have found that DNTF still has many problems, such as poor thermal stability, poor mechanical properties, crystallization, poor compatibility, etc., all of which can be solved by eutectics. In recent years, eutectics have been used as an effective way to improve material performance. The energetic eutectic technology can change the molecular structure composition (molecular arrangement) and crystal structure (crystal packing coefficient, density, etc.) of energetic materials without destroying the original molecular structure of energetic compounds, reduce the free volume of crystals, form a tight stack, and further reduce the probability of hot spot formation. As a result, physical properties such as density, melting point, sensitivity and explosion performance change, achieving a balance between energy and safety. However, due to the huge experimental cost and certain safety issues of eutectic technology, in the past few decades, although eutectic materials have gradually shown their due role in their respective application fields, such as electronic devices, semiconductor engineering, chemical engineering, etc., but in the final analysis, the eutectic system is too complex and the research on eutectic structure is relatively restricted, so it has not been widely used in the field of energetic materials.

[0004] At present, the field of artificial intelligence machine learning is developing rapidly. In the field of deep learning in artificial intelligence, as the performance of deep learning model prediction and classification is improved, deep learning methods have begun to be applied to the field of materials science. Compared with the traditional method of extracting simple underlying features through data mining, the method based on deep learning can learn a lot of high-level features that cannot be recognized by the naked eye from a large amount of data to represent information, which improves the prediction ability of complex data environments. At the same time, the deep learning method is more robust and adaptive through large-scale data training, and can adapt to different types of data forms, avoiding the problem of large differences in prediction effects of traditional methods due to the complexity of material data.

[0005] However, in the research on deep learning methods for predicting DNTF cocrystal ligands, the following problems still need to be solved: (i) The ratio of positive samples to negative samples in the existing database is unbalanced, and the difference in dimensions is not conducive to adjusting the weights of the neural network during training; (ii) There is currently no deep learning framework established for DNTF cocrystal ligand molecules, which fails to play the role of deep learning and requires the establishment of a corresponding deep learning model for the data model; (iii) Many current model predictions do not reflect the iterative nature of the cocrystal research process. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning. The technical problem to be solved by the present invention is achieved by the following technical solutions: The present invention provides a prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning, comprising: S1: obtaining a crystalline compound satisfying a preset DNTF-related condition as a related crystalline compound according to the chemical parameters of the DNTF; S2: constructing a training data set according to the obtained related crystalline compounds; S3: dividing the training data set into a support set and a query set, and performing comparative learning training on the constructed DNTF eutectic prediction network using the support set and the query set to obtain a trained DNTF eutectic prediction network model; S4: Predicting the ligand molecules of DNTF using the trained DNTF co-crystal prediction network model.

[0007] In one embodiment of the present invention, the S1 includes: According to the chemical parameters of DNTF, a crystalline compound that meets the preset DNTF-related conditions is searched in the Cambridge Crystal Library as a related crystalline compound, wherein the chemical parameters of DNTF include the molecular formula of DNTF, the structure of DNTF, the number of functional groups of DNTF and the type of functional groups of DNTF; the preset DNTF-related conditions include at least one of the following: having the same molecular formula as DNTF, being partially identical to the chemical structure of DNTF, having the same number of functional groups as DNTF and having partially identical functional groups as DNTF.

[0008] In one embodiment of the present invention, the S2 includes: S2.1: Obtaining crystal information of the relevant crystal compound and the selected corresponding ligand molecule; S2.2: obtaining multimodal data according to the crystal information of the relevant crystalline compound and the ligand molecule, and generating a multimodal data sample, wherein the multimodal data includes atomic coordinate data, chemical bond data, crystal package parameters, space group data, density data, melting point data, crystallographic lattice type and refractive index of the relevant crystalline compound and the ligand molecule; S2.3: All multimodal data samples are used to form a training data set, wherein the co-crystal and non-co-crystal situations of the relevant crystalline compounds and the corresponding ligand molecules are labeled separately, the multimodal data of the relevant crystalline compounds and the corresponding ligand molecules that can form co-crystals are labeled as positive samples, and the multimodal data of the relevant crystalline compounds and the corresponding ligand molecules that cannot form co-crystals are labeled as negative samples.

[0009] In one embodiment of the present invention, the S2.3 includes: All multimodal data are encoded in the form of feature encoding to obtain vector format data that can be recognized by the DNTF eutectic prediction network to form a data set; The data in the data set are randomly divided into a training data set and a test data set in a ratio of 4:1, and the multimodal data that the relevant crystalline compounds and the corresponding ligand molecules can form a co-crystal is set to label 1, and the multimodal data that the relevant crystalline compounds and the corresponding ligand molecules cannot form a co-crystal is set to label 0.

[0010] In one embodiment of the present invention, S3 includes: S3.1: Create N training tasks and the support set and query set of each training task. The support set of each training task includes 5 negative samples selected from the training data set, and the query set of each training task includes 1 positive sample and 1 negative sample selected from the training data set. S3.2: constructing a DNTF eutectic prediction network, wherein the DNTF eutectic prediction network includes a connected feature extractor and a classifier, wherein the feature extractor includes an embedding layer, a batch normalization layer, a principal component analysis layer, and a linear mapping layer connected in sequence; S3.3: Inputting the sample data in the support set and the query set into the feature extractor in sequence to obtain 5 negative sample features of the support set and 1 positive sample feature and 1 negative sample feature of the query set; S3.4: using the classifier to respectively calculate the Mahalanobis distances between the five negative sample features of the support set and the one positive sample feature of the query set, and the Mahalanobis distances between the five negative sample features of the support set and the one negative sample feature of the query set; S3.5: using the Mahalanobis distance to obtain a total contrastive learning loss, and performing gradient back propagation based on the total contrastive learning loss to update the network parameters of the DNTF eutectic prediction network; S3.6: Repeat steps S3.3 to S3.5 using the support set and query set in another training task until all training tasks are completed, and obtain the trained DNTF eutectic prediction network model.

[0011] In one embodiment of the present invention, the calculation formula of the Mahalanobis distance is: , , in, Represents the feature vector and The Mahalanobis distance, Represents the feature vector and The difference vector of Represents the difference vector The transpose of represents the covariance matrix, represents the covariance, Represents the covariance matrix The inverse matrix of , , Represents the feature vector and Dimension.

[0012] In one embodiment of the present invention, using the Mahalanobis distance to obtain the total loss of contrastive learning includes: The classifier adds or averages the Mahalanobis distances of 5 negative sample features of the support set of the current training task and 1 negative sample feature of the query set to obtain the negative-negative sample contrast learning loss; adds or averages the Mahalanobis distances of 5 negative sample features of the support set of the current training task and 1 positive sample feature of the query set to obtain the positive-negative sample contrast learning loss; subtracts the negative-negative sample contrast learning loss from the positive-negative sample contrast learning loss to obtain the total contrast learning loss of the current training task.

[0013] In one embodiment of the present invention, after S3, the method further includes: The positive samples and negative samples of the query set in the test data set are respectively input into the trained DNTF eutectic prediction network model, and the predicted labels are output. The predicted labels are compared with the true labels of the positive samples or negative samples in the query set to test the accuracy of the trained DNTF eutectic prediction network model.

[0014] In one embodiment of the present invention, the S4 includes: S4.1: Use the Cambridge Crystal Library to obtain multimodal data of DNTF and the ligand molecules to be predicted; S4.2: Input the multimodal data of DNTF and the ligand molecule to be predicted into the trained DNTF co-crystal prediction network model to obtain a prediction result of whether DNTF and the ligand molecule to be predicted can form a co-crystal or not.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention proposes a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning, which uses existing crystal structure data to construct a training data set for learning, and establishes a reliable and effective prototype neural network model through a data-driven method, which helps eutectic experiments to quickly predict and recommend matching ligand molecules for the molecular structure of DNTF crystals, which can significantly reduce the cost of ligand molecule screening, effectively reduce potential risks in the experiment, and promote the application of DNTF high-performance oxidizers in the field of solid propellants.

[0016] 2. The present invention establishes a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning. The deep learning used can learn deeper information in the data compared with ensemble learning, promotes the deep integration of the learning field and the material field, reduces the loss of resources and time caused by human experiments, and improves the screening efficiency of ligand molecules of energetic molecules.

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flowchart of a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning provided by an embodiment of the present invention; Figure 2 is a schematic diagram of a planar structure of a DNTF provided by an embodiment of the present invention; Figure 3 It is a schematic diagram of constructing a training data set and a test data set provided by an embodiment of the present invention; Figure 4 is a schematic diagram of a training strategy provided by an embodiment of the present invention; Figure 5 It is a schematic diagram of a framework of a DNTF eutectic prediction network based on positive and negative sample contrast learning provided by an embodiment of the present invention; Figure 6 It is a schematic diagram of the testing process of a DNTF eutectic prediction network model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning proposed in accordance with the present invention, in combination with the accompanying drawings and specific implementation methods.

[0020] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0021] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0022] Embodiment 1 See also Figure 1 , Figure 1 : is a flowchart of a prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning provided by an embodiment of the present invention. The DNTF eutectic prediction method comprises the following steps: S1: Acquire a crystalline compound satisfying preset DNTF related conditions as a related crystalline compound according to the chemical parameters of the DNTF.

[0023] This embodiment searches for crystalline compounds that meet preset DNTF-related conditions in the Cambridge Crystal Database (CCDC) according to the chemical parameters of DNTF as related crystalline compounds. In the current Cambridge Crystal Database, no materials that can form substantial eutectics with DNTF have been found, so it is impossible to directly obtain DNTF eutectic data through the existing database as training data for the subsequent DNTF eutectic prediction network. For this reason, this embodiment proposes a method for training a eutectic prediction network using related crystalline compounds similar to DNTF.

[0024] Optionally, the chemical parameters of DNTF include: molecular formula of DNTF, structure of DNTF, number of functional groups of DNTF and type of functional groups of DNTF, see Figure 2 , Figure 2: is a planar structure diagram of a DNTF provided in an embodiment of the present invention. In this embodiment, based on the molecular formula of DNTF, the structure of DNTF, the number of functional groups of DNTF and the type of functional groups of DNTF, a search is performed in the Cambridge Crystal Library to obtain related crystalline compounds similar to DNTF.

[0025] Optionally, the preset DNTF-related conditions include: having the same molecular formula as DNTF, having partially the same chemical structure as DNTF, having the same number of functional groups as DNTF, and having partially the same functional groups as DNTF. In other words, the relevant crystalline compound is a crystalline compound having at least one of the preset DNTF-related conditions.

[0026] like Figure 2 As shown, the DNTF molecule has a five-membered furazan group, and two nitro groups are connected as substituents to the 3rd and 4th carbon atoms of the furazan group. In this embodiment, the chemical structure partially identical to that of DNTF includes: the crystalline compound has a furazan ring and a nitro group is connected to any position of the furazan ring, or the crystalline compound has a benzene ring and corresponding nitro groups are connected to the 3rd and 4th positions of the benzene ring.

[0027] The functional groups of the DNTF molecule include a furazan ring, a nitro group, a carbonyl group and a CN bond. In this embodiment, the functional groups having some of the same functional groups as those of DNTF include: the crystalline compound having at least two functional groups that are the same as those of DNTF.

[0028] S2: Construct a training dataset based on the acquired related crystal compounds.

[0029] In an optional embodiment, S2 includes: S2.1: Obtaining crystal information of the relevant crystal compound and the selected corresponding ligand molecule.

[0030] Specifically, the relevant information of the relevant crystal compound and its ligand molecule can be called in the Cambridge Crystal Library through the code by using the CSD python API. In this embodiment, chemical structure information, crystallographic information and chemical physical properties are mainly used as crystal information for co-crystal research, that is, the crystal information includes chemical structure information, crystallographic information and chemical physical properties. Among them, the chemical structure information includes the coordinates of atoms and the information of chemical bonds, the crystallographic information includes the unit cell parameters, space group information and crystallographic lattice type, and the chemical physical properties include melting point, density and refractive index.

[0031] S2.2: Obtain multimodal data based on the crystal information of the relevant crystalline compound and the ligand molecule, and generate a multimodal data sample.

[0032] In eutectic studies, atomic coordinates are used to analyze the spatial arrangement and interactions between different components in the eutectic system. Information on chemical bonds helps to understand the chemical bonding between different components, which is crucial for the formation of eutectics. Crystallographic package parameters are used to determine the lattice matching and lattice spacing of the eutectic system, which has an important influence on the formation and stability of the eutectic. Space group information helps to understand the symmetry of the eutectic structure, which is helpful for predicting and explaining the physical properties of the eutectic. The crystallographic lattice type provides basic structural framework information of the eutectic system. The melting point is one of the characteristics of the eutectic system and can be used to confirm the formation of the eutectic. Density is related to the composition and structure of the eutectic and is an important parameter for studying the properties of the eutectic. The refractive index may be used to analyze the phase purity and optical properties of the eutectic system.

[0033] Based on this, in this embodiment, the multimodal data obtained based on the crystal information of the relevant crystalline compounds and ligand molecules include: atomic coordinate data, chemical bond data, crystal package parameters, space group data, density data, melting point data, crystallographic lattice type and refractive index.

[0034] S2.3: All multimodal data samples are used to form a training data set, wherein the co-crystal and non-co-crystal situations of the relevant crystalline compounds and the corresponding ligand molecules are labeled separately, the multimodal data of the relevant crystalline compounds and the corresponding ligand molecules that can form co-crystals are labeled as positive samples, and the multimodal data of the relevant crystalline compounds and the corresponding ligand molecules that cannot form co-crystals are labeled as negative samples.

[0035] Specifically, the multimodal data of all relevant crystalline compounds and corresponding ligand molecules are encoded in the form of feature encoding to obtain corresponding vector format data that can be recognized by the DNTF co-crystal prediction network to form a data set. In this embodiment, the data in the data set is randomly divided into a training data set and a test data set in a ratio of 4:1, and the data of the relevant crystalline compounds and their ligand molecules that can form co-crystals are set to label 1 (positive samples), and the data of the relevant crystalline compounds and their ligand molecules that cannot form co-crystals are set to label 0 (negative samples).

[0036] That is to say, when constructing the data set, a part of the related crystal compounds and ligand molecules that can form cocrystals are selected as positive samples, and a part of the related crystal compounds and ligand molecules that cannot form cocrystals are selected as negative samples.

[0037] S3: Divide the training data set into a support set and a query set, and use the support set and the query set to perform comparative learning training on the constructed DNTF eutectic prediction network to obtain the trained DNTF eutectic prediction network model.

[0038] In this embodiment, step S3 includes: S3.1: Create N training tasks and a support set and a query set for each training task. The support set of each training task includes 5 negative samples selected from the training data set, and the query set of each training task includes 1 positive sample and 1 negative sample selected from the training data set.

[0039] See also Figure 3 , Figure 3 : is a schematic diagram of constructing a training data set and a test data set provided by an embodiment of the present invention. As described above, the data in the data set is randomly divided into a training data set and a test data set in a ratio of 4:1. This step creates N training tasks (Task), and takes out 5 data with a label of 0 (negative samples) from the training data set as a support set, and then takes out 1 data with a label of 1 (positive sample) and 1 data with a label of 0 (negative sample) as a query set. The support set and the query set are used as the data sets to be used for a training task, such as Figure 4 As shown in Figure 1. That is, a training task includes a support set containing 5 negative samples and a query set containing 1 negative sample and 1 positive sample. The same is true for the test data set, that is, 5 samples with a label of 0 (negative samples) are taken from the test data set as a support set, and 1 sample with a label of 1 (positive sample) and 1 sample with a label of 0 (negative sample) are taken as a query set.

[0040] The strategy for training the DNTF eutectic prediction network using this training dataset is: use the support set for training as a prototype (i.e., an invariant reference point in the feature space) in each training task, and then use the data labeled 1 and 0 in the query set for contrastive learning, and then subtract the contrastive learning loss of the positive sample from the negative sample as the total contrastive learning loss, and back-propagate this total contrastive learning loss to optimize the weights in the DNTF eutectic prediction network. After the training is completed, the accuracy of the model is tested on the trained network model using the query set in the test dataset.

[0041] S3.2: Construct DNTF eutectic prediction network.

[0042] See also Figure 5 , Figure 5It is a schematic diagram of a framework of a prototype network based on positive and negative sample contrast learning provided by an embodiment of the present invention. The DNTF eutectic prediction network of this embodiment includes a connected feature extractor and a classifier, wherein the feature extractor includes an embedding layer (embedding layer), a batch normalization layer (batch normal layer), a principal component analysis (Principal Component Analysis, PCA) layer and a linear mapping (Linear) layer connected in sequence. The embedding layer is used to initially map the input data to the primary features of the data; the batch normalization layer is used to normalize the data dimensions of each batch (a batch refers to a training task) to facilitate processing in the same feature space. Principal component analysis is a multivariate statistical analysis method for unsupervised learning. The main principle of PCA analysis is to project high-dimensional data into a lower-dimensional space, extract the main factors of multivariate things, and reveal their essential characteristics. It can efficiently find the main part of the data and reduce the dimensionality of the original complex data. The principal component analysis layer of this embodiment is used to extract the main features in the data. The function of the linear mapping layer is to linearly map the main features extracted by the PCA layer, so as to facilitate input into the classifier for classification.

[0043] S3.3: Input the sample data in the support set and the query set into the feature extractor in sequence to obtain 5 negative sample features corresponding to the support set and 1 positive sample feature and 1 negative sample feature corresponding to the query set.

[0044] In this step, after the 5 negative samples in the support set are subjected to feature extraction by the feature extractor, 5 negative sample features are output, and after the 1 positive sample and 1 negative sample in the query set are subjected to feature extraction by the feature extractor, 1 positive sample feature and 1 negative sample feature are output respectively.

[0045] S3.4: Use the classifier to calculate the Mahalanobis distance between the 5 negative sample features of the support set and the 1 positive sample feature of the query set, as well as the Mahalanobis distance between the 5 negative sample features of the support set and the 1 negative sample feature of the query set.

[0046] Specifically, the calculation formula of Mahalanobis distance can be expressed as: , , in, Represents the feature vector and The Mahalanobis distance, Represents the feature vector and The difference vector of Represents the difference vector The transpose of represents the covariance matrix, represents the covariance, Represents the covariance matrix The inverse matrix of , , Represents the feature vector and Dimension.

[0047] Exemplarily, the feature vector x in the formula is the positive sample feature or negative sample feature corresponding to the query set, and the feature vector y is the negative sample feature of the support set.

[0048] The distance calculation process here uses the Mahalanobis distance because the Mahalanobis distance has many advantages. The Mahalanobis distance is not affected by the dimension, and the Mahalanobis distance between two points is independent of the measurement unit of the original data. The Mahalanobis distance between two points calculated by the standardized data and the centered data (that is, the difference between the original data and the mean) is the same. In addition, the Mahalanobis distance can also eliminate the interference of the correlation between variables.

[0049] S3.5: Use Mahalanobis distance to obtain the total loss of contrastive learning, and backpropagate the loss gradient to update the network parameters of the DNTF eutectic prediction network.

[0050] As shown above, the classifier calculates the Mahalanobis distances of the 5 negative sample features corresponding to the support set and the 1 positive sample feature corresponding to the query set respectively; and calculates the Mahalanobis distances of the 5 negative sample features corresponding to the support set and the 1 negative sample feature corresponding to the query set respectively. In this step, the classifier adds or averages the Mahalanobis distances of the 5 negative sample features of the support set of the current training task and the 1 negative sample feature of the query set to obtain the negative-negative sample contrast learning loss; adds or averages the Mahalanobis distances of the 5 negative sample features of the support set of the current training task and the 1 positive sample feature of the query set to obtain the positive-negative sample contrast learning loss; subtracts the negative-negative sample contrast learning loss from the positive-negative sample contrast learning loss to obtain the total contrast learning loss of the current training task.

[0051] Subsequently, the total loss is learned using this contrast, and the network parameters of the DNTF eutectic prediction network are updated based on the loss gradient backpropagation.

[0052] S3.6: Repeat steps S3.3 to S3.5 using the support set and query set in another training task until all training tasks are completed, and obtain the trained DNTF eutectic prediction network model.

[0053] In this embodiment, after the network training is completed in step S3, a testing phase is also included. Figure 6The testing process includes inputting the positive samples and negative samples of the query set in the test data set into the trained DNTF eutectic prediction network model respectively, thereby outputting the prediction results, which are label 1 (indicating eutectic) or label 0 (indicating non-eutectic), and comparing the predicted labels with the actual labels of the data in the query set to test the accuracy of the trained DNTF eutectic prediction network model.

[0054] S4: Use the trained DNTF co-crystal prediction network model to predict the ligand molecules of DNTF.

[0055] In an optional embodiment, S4 includes: S4.1: Use the Cambridge Crystallographic Library to obtain multimodal data of DNTF and the ligand molecules to be predicted.

[0056] In this embodiment, the multimodal data includes: atomic coordinate data, chemical bond data, crystal package parameters, space group data, density data, melting point data, crystallographic lattice type and refractive index of DNTF and the ligand molecule to be predicted.

[0057] S4.2: Input the multimodal data of DNTF and the ligand molecule to be predicted into the trained DNTF co-crystal prediction network model to obtain the prediction result of whether DNTF and the ligand molecule to be predicted can form a co-crystal or not. Specifically, the classifier of the DNTF co-crystal prediction network model outputs label 1 or label 0, label 1 represents that DNTF and the ligand molecule to be predicted can form a co-crystal, and label 0 represents that DNTF and the ligand molecule to be predicted cannot form a co-crystal.

[0058] The present invention proposes a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning, which utilizes existing crystal structure data to construct a training data set for learning, and establishes a reliable and effective prototype neural network model through a data-driven approach, so as to help eutectic experiments to quickly predict and recommend matching ligand molecules for the molecular structure of DNTF crystals, which can significantly reduce the cost of ligand molecule screening, effectively cut down on potential risks in the experiment, and promote the application of DNTF high-performance oxidizers in the field of solid propellants.

[0059] The present invention establishes a prototype network DNTF eutectic prediction method based on positive and negative sample comparative learning. The deep learning used can learn deeper information in the data compared with ensemble learning, promotes the deep integration of the learning field and the material field, reduces the loss of resources and time caused by human experiments, and improves the screening efficiency of ligand molecules of energetic molecules.

[0060] Another embodiment of the present invention provides a storage medium, in which a computer program is stored, and the computer program is used to execute the steps of the prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning described in the above embodiment. Another aspect of the present invention provides an electronic device, including a memory and a processor, in which a computer program is stored, and when the processor calls the computer program in the memory, the steps of the prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning as described in the above embodiment are implemented. Specifically, the above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including several instructions for an electronic device (which can be a personal computer, a server, or a network device, etc.) or a processor (processor) to perform some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0061] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning, characterized in that: include: S1: obtaining a crystalline compound satisfying a preset DNTF-related condition as a related crystalline compound according to the chemical parameters of the DNTF; S2: constructing a training data set according to the obtained related crystalline compounds; S3: dividing the training data set into a support set and a query set, and performing comparative learning training on the constructed DNTF eutectic prediction network using the support set and the query set to obtain a trained DNTF eutectic prediction network model; S4: Predicting the ligand molecules of DNTF using the trained DNTF co-crystal prediction network model.

2. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 1 is characterized in that: The S1 includes: According to the chemical parameters of DNTF, a crystalline compound that meets the preset DNTF-related conditions is searched in the Cambridge Crystal Library as a related crystalline compound, wherein the chemical parameters of DNTF include the molecular formula of DNTF, the structure of DNTF, the number of functional groups of DNTF and the type of functional groups of DNTF; the preset DNTF-related conditions include at least one of the following: having the same molecular formula as DNTF, being partially identical to the chemical structure of DNTF, having the same number of functional groups as DNTF and having partially identical functional groups as DNTF.

3. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 1 is characterized in that: The S2 includes: S2.1: Obtaining crystal information of the relevant crystal compound and the selected corresponding ligand molecule; S2.2: obtaining multimodal data according to the crystal information of the relevant crystalline compound and the ligand molecule, and generating a multimodal data sample, wherein the multimodal data includes atomic coordinate data, chemical bond data, crystal package parameters, space group data, density data, melting point data, crystallographic lattice type and refractive index of the relevant crystalline compound and the ligand molecule; S2.3: All multimodal data samples are used to form a training data set, wherein the co-crystal and non-co-crystal situations of the relevant crystalline compounds and the corresponding ligand molecules are labeled separately, the multimodal data of the relevant crystalline compounds and the corresponding ligand molecules that can form co-crystals are labeled as positive samples, and the multimodal data of the relevant crystalline compounds and the corresponding ligand molecules that cannot form co-crystals are labeled as negative samples.

4. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 3 is characterized in that: The S2.3 includes: All multimodal data are encoded in the form of feature encoding to obtain vector format data that can be recognized by the DNTF eutectic prediction network to form a data set; The data in the data set are randomly divided into a training data set and a test data set in a ratio of 4:1, and the multimodal data that the relevant crystalline compounds and the corresponding ligand molecules can form a co-crystal is set to label 1, and the multimodal data that the relevant crystalline compounds and the corresponding ligand molecules cannot form a co-crystal is set to label 0.

5. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 2 is characterized in that: The S3 includes: S3.1: Create N training tasks and the support set and query set of each training task. The support set of each training task includes 5 negative samples selected from the training data set, and the query set of each training task includes 1 positive sample and 1 negative sample selected from the training data set. S3.2: constructing a DNTF eutectic prediction network, wherein the DNTF eutectic prediction network includes a connected feature extractor and a classifier, wherein the feature extractor includes an embedding layer, a batch normalization layer, a principal component analysis layer, and a linear mapping layer connected in sequence; S3.3: Inputting the sample data in the support set and the query set into the feature extractor in sequence to obtain 5 negative sample features of the support set and 1 positive sample feature and 1 negative sample feature of the query set; S3.4: using the classifier to respectively calculate the Mahalanobis distances between the five negative sample features of the support set and the one positive sample feature of the query set, and the Mahalanobis distances between the five negative sample features of the support set and the one negative sample feature of the query set; S3.5: using the Mahalanobis distance to obtain a total contrastive learning loss, and performing gradient back propagation based on the total contrastive learning loss to update the network parameters of the DNTF eutectic prediction network; S3.6: Repeat steps S3.3 to S3.5 using the support set and query set in another training task until all training tasks are completed, and obtain the trained DNTF eutectic prediction network model.

6. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 5 is characterized in that: The calculation formula of the Mahalanobis distance is: ， ， in, Represents the feature vector and The Mahalanobis distance, Represents the feature vector and The difference vector of Represents the difference vector The transpose of represents the covariance matrix, represents the covariance, Represents the covariance matrix The inverse matrix of , , Represents the feature vector and Dimension.

7. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 5 is characterized in that: The total loss of contrastive learning is obtained using the Mahalanobis distance, including: The classifier adds or averages the Mahalanobis distances of 5 negative sample features of the support set of the current training task and 1 negative sample feature of the query set to obtain the negative-negative sample contrast learning loss; adds or averages the Mahalanobis distances of 5 negative sample features of the support set of the current training task and 1 positive sample feature of the query set to obtain the positive-negative sample contrast learning loss; subtracts the negative-negative sample contrast learning loss from the positive-negative sample contrast learning loss to obtain the total contrast learning loss of the current training task.

8. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 4 is characterized in that: After S3, the following is also included: The positive samples and negative samples of the query set in the test data set are respectively input into the trained DNTF eutectic prediction network model, and the predicted labels are output. The predicted labels are compared with the true labels of the positive samples or negative samples in the query set to test the accuracy of the trained DNTF eutectic prediction network model.

9. The prototype network DNTF eutectic prediction method based on positive and negative sample contrast learning according to claim 2 is characterized in that: The S4 includes: S4.1: Use the Cambridge Crystal Library to obtain multimodal data of DNTF and the ligand molecules to be predicted; S4.2: Input the multimodal data of DNTF and the ligand molecule to be predicted into the trained DNTF co-crystal prediction network model to obtain a prediction result of whether DNTF and the ligand molecule to be predicted can form a co-crystal or not.

Images