Design method of molecular working medium and related device

Abstract

The invention discloses a molecular type working medium design method and a related device, and the method comprises the steps: obtaining a design target of a target molecular type working medium, abstracting atoms and chemical bonds in molecules into a graph structure composed of nodes and edges; according to the graph structure, taking an E (n) isovariant graph neural network with an attention mechanism as a basic structure, constructing an adversarial coding graph generation model, and performing training; and generating a model based on the trained adversarial coding graph, and generating a target molecular type working medium by using an adversarial regularization constraint mechanism and a dynamic priority asynchronous interpretation strategy. By constructing an adversarial regularization constraint mechanism, high-efficiency conversion between prior distribution and sample distribution is realized, and the boundary capacity of an adversarial coding graph generation model is remarkably expanded. Meanwhile, based on a dynamic priority asynchronous interpretation strategy, it is ensured that all atoms in output molecules strictly meet chemical valence constraints, and therefore the effectiveness of generating the molecular type working medium is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided molecular design technology, and relates to a design method and related apparatus for molecular working fluids. Background Technology

[0002] Lithium metal batteries, with their extremely high theoretical specific capacity of lithium metal anodes, exhibit energy density advantages far exceeding those of traditional lithium-ion batteries, providing a way to overcome range bottlenecks in electric vehicles, portable electronic devices, and energy storage systems. However, their commercialization is hampered by the core defect of uncontrolled lithium dendrite growth. During cycling, the porous interface formed by continuous side reactions between metallic lithium and the liquid electrolyte exacerbates dendrite penetration of the separator, leading to thermal runaway or even explosion. Meanwhile, low coulombic efficiency and short cycle life severely restrict practical applications. Furthermore, traditional carbonate electrolytes struggle to construct stable solid electrolyte mesophases on the lithium metal surface and stable electrochemical interfaces on the cathode surface. Therefore, it is necessary to design novel electrolytes that can suppress dendrite growth while maintaining high ionic conductivity, thereby achieving a synergistic optimization of high energy density and intrinsic safety.

[0003] For the design of novel electrolytes, using molecular generation models based on big data is an effective solution. After training with sufficient molecular structures and their properties, the molecular generation model can quickly find or design corresponding molecules not present in the dataset based on the target properties. Traditional molecular generation models use basic graph neural networks to operate on molecular graphs, requiring the molecular structure to be reversibly and as losslessly as possible converted into a digital representation. To achieve this, a step-by-step conversion strategy is usually adopted: first, the molecular structure is mapped to molecular identifiers, and then the identifiers are vectorized through mathematical encoding. As the core method for describing the structural features of compounds, molecular identifiers directly affect the generative model's ability to capture molecular information. However, they lack three-dimensional spatial information of molecular topology, thus the generative model's ability to analyze structural features is insufficient. In the development of molecular generation models, the original generation framework often exhibits random distribution characteristics in its output samples due to the lack of a directional guidance mechanism, making it difficult to meet the specific requirements for molecular physicochemical properties in actual research and development. To overcome this fundamental challenge, previous research has added extra structures to the basic framework of generative models, using ensemble constraint mechanisms to regulate the generation process, resulting in two main technical routes—property optimization-based generation and property-conditional generation. The difference between the two lies in the expression of the optimization objective and the algorithm implementation path. However, ensemble constraint mechanisms lack constraints on atomic valence state rules, leading to irrationalities in the generated molecular structures. Furthermore, generative models consist of multiple neural network layers, each with its own function to achieve the final goal of working fluid generation. Different types of neural networks accept different data structures and focus on different aspects of the information within those structures. This makes it difficult to achieve the distributed transformation of the generative model in a single step using a single neural network, resulting in performance limitations. Summary of the Invention

[0004] The purpose of this invention is to provide a design method and related apparatus for molecular working fluids, thereby solving the problems in the prior art.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A design method for molecular working fluids includes:

[0007] To obtain the design target of the target molecular working fluid, the atoms and chemical bonds in the molecule are abstracted into a graph structure composed of nodes and edges;

[0008] Based on the graph structure, an adversarial coding graph generation model is constructed and trained using an E(n) variable graph neural network with attention mechanism as the basic structure.

[0009] Based on a trained adversarial coding graph generation model, the target molecular working fluid is generated by utilizing an adversarial regularization constraint mechanism and a dynamic priority asynchronous interpretation strategy.

[0010] Furthermore, in the graph structure, nodes represent the individual atoms in a molecule, and edges represent the chemical bonds between the atoms; each node has its own unique thermal code for its atom type, its relative atomic mass, and the number of implicit hydrogen atoms connected; each chemical bond has its own unique thermal code for its bond type and its bond length.

[0011] Furthermore, the adversarial coding graph generation model includes an encoder, a decoder, a generator, and a discriminator. The encoder includes an embedding layer, an AEGNN layer, a mean output layer, a variance output layer, and a reparameterization layer.

[0012] The decoder includes a node reconstruction layer, a coordinate reconstruction layer, an atom number reconstruction layer, an AEGNN layer, and an embedding layer;

[0013] Both the generator and the discriminator are multi-layer neural networks, and the generator and the discriminator jointly constrain the latent space distribution.

[0014] Furthermore, the training method for the adversarial coding graph generation model is as follows:

[0015] In the adversarial coding graph generation model, the generator takes samples of the prior distribution as input and outputs a series of latent space coordinates. The latent space coordinates are input together with the coordinates output by the encoder of the molecular structure in the adversarial coding graph generation model. The discriminator is responsible for determining whether the series of received coordinates are sampled in the true latent space formed by the encoder or in the false latent space generated by the generator.

[0016] To achieve adversarial constraints, the generator and encoder try to maintain similar distributions so that the discriminator makes incorrect judgments, while the discriminator tries to distinguish the source of its input as accurately as possible.

[0017] Furthermore, dynamic priority asynchronous interpretation strategies include:

[0018] Determine all atom types and the number of implicit hydrogens, and determine the maximum valence state using an atom valence state algorithm. Based on the maximum valence state and the existing number of implicit hydrogens, set the probability of all impossible chemical bonds to 0.

[0019] Sample a single bond for all atoms with a remaining valence of 1, sample a chemical bond for atoms with a maximum valence of 2 that have already been bonded, and sample a chemical bond for atoms that have not yet been bonded.

[0020] Sample chemical bonds for all atoms whose connected chemical bonds have not reached their maximum valence state until all atoms have reached their maximum valence state.

[0021] Furthermore, the atomic valence state algorithm is as follows:

[0022] For an atom with multiple common valence states, the chemical bond connections are calculated based on each valence state, and the mean of the probabilities of each chemical bond sampled under that valence state is calculated. The mean probabilities of each chemical bond under each valence state are compared, and the valence state with the largest mean is selected as the valence state of that atom in the molecule.

[0023] Furthermore, the process of generating the target molecular working fluid using an adversarial coding graph generation model includes:

[0024] Sampling is performed from the prior distribution, and the sampled data is input into the generator in the adversarial coding graph generation model. The output of the generator and the generation conditions are then processed by the feature reconstruction layer of the decoder in the adversarial coding graph generation model to reconstruct the various matrices required by the decoder. The output of the feature reconstruction layer is then decoded by the decoder and interpreted using a dynamic priority asynchronous interpretation strategy to obtain the molecular structure.

[0025] A design system for molecular working fluids, comprising:

[0026] The abstract module is used to obtain the design target of the target molecular working fluid, and abstracts the atoms and chemical bonds in the molecule into a graph structure composed of nodes and edges;

[0027] The modeling module is used to construct and train an adversarial coding graph generation model based on the graph structure, using an E(n) variable graph neural network with attention mechanism as the basic structure.

[0028] The generation module is used to generate the target molecular working fluid based on the trained adversarial coding graph generation model, utilizing adversarial regularization constraint mechanism and dynamic priority asynchronous interpretation strategy.

[0029] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method.

[0030] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] This invention provides a design method for molecular working fluids. By obtaining the design target of the target molecular working fluid, the atoms and chemical bonds in the molecule are abstracted into a graph structure composed of nodes and edges. Based on this graph structure, an E(n) variable graph neural network with an attention mechanism is used as the basic structure to construct and train an adversarial coding graph generation model. Addressing the inherent deficiency of insufficient topological representation ability of molecular graphs, this invention innovatively constructs a molecular graph containing node coordinate information and uses AEGNN as the underlying structure of the adversarial coding graph generation model, establishing a novel adversarial coding graph generation model for molecular working fluids. This enhanced molecular graph, as the input identifier for the adversarial coding graph generation model, significantly strengthens the model's ability to extract information from the molecular structure. Finally, based on the trained adversarial coding graph generation model, an adversarial regularization constraint mechanism and a dynamic priority asynchronous interpretation strategy are proposed on the basis of a variational autoencoder to generate the target molecular working fluid. This invention, by constructing an adversarial regularization constraint mechanism, replaces the manually defined prior distribution, achieving efficient transformation between the prior distribution and the sample distribution, and significantly expanding the boundary capacity of the adversarial coding graph generation model. Meanwhile, in terms of decoding strategy, a dynamic priority asynchronous interpretation strategy is proposed based on the dynamic priority sorting mechanism of molecular graph node features. This strategy ensures that all atoms in the output molecule strictly satisfy the chemical valence state constraint by intelligently scheduling the order of atom generation, thereby greatly improving the effectiveness of generating molecular working fluids.

[0033] Furthermore, this invention uses the QM9 and ZINC databases to train an adversarial coding graph generation model. The rationality, uniqueness, and novelty of the adversarial coding graph generation model trained on the QM9 database reached 0.984, 0.537, and 0.771, respectively; the rationality, uniqueness, and novelty of the adversarial coding graph generation model trained on the ZINC database reached 1.000, 0.688, and 1.000, respectively, surpassing the current state-of-the-art GraphVAE, MolGAN, and NAGVAE models, demonstrating a powerful ability to generate molecular working fluids. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating the design method of the molecular working fluid of the present invention.

[0036] Figure 2 This is a schematic diagram of the overall structure of the adversarial coding graph generation model of the present invention.

[0037] Figure 3 This is a schematic diagram illustrating the application process of the adversarial coding graph generation model of the present invention.

[0038] Figure 4 This is a comparison diagram of the synchronous method, asynchronous method, and dynamic priority asynchronous interpretation strategy of the present invention.

[0039] Figure 5 The diagram shows the performance of various common electrolyte solvents used in this invention.

[0040] Figure 6 This is a comparison chart of the molecular structure, formation conditions, and overall performance of the DFEPN of the present invention.

[0041] Figure 7 This is a schematic diagram of the design system structure of a molecular working fluid according to a preferred embodiment of the present invention.

[0042] Figure 8 This is a schematic diagram of the electronic device structure according to a preferred embodiment of the present invention. Detailed Implementation

[0043] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0044] Obviously, the described embodiments are only some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0045] It should be noted that the terminals involved in the embodiments of this application may include, but are not limited to, mobile phones, personal digital assistants (PDAs), wireless handheld devices, tablet computers, personal computers (PCs), MP3 players, MP4 players, wearable devices (e.g., smart glasses, smartwatches, smart bracelets, etc.), smart home devices, and other smart devices.

[0046] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0047] The present invention will now be described in further detail with reference to the accompanying drawings:

[0048] See Figure 1 This invention provides a design method for molecular working fluids, specifically including the following steps:

[0049] Step 1: Obtain the design target of the target molecular working fluid, and abstract the atoms and chemical bonds in the molecule into a graph structure composed of nodes and edges.

[0050] To obtain the design target of the target molecular working fluid, the molecule is abstracted as a graph structure composed of atoms (nodes) and chemical bonds (edges). By preserving the interatomic connections and spatial topological information, the chemical properties of the molecule can be characterized relatively completely, and it has good algorithm adaptability. The graph consists of nodes and edges connecting the nodes: G = {V, E}. When used to represent the molecular structure, the node set V = {v...} i {i = 1, ..., n} represent the atoms in a molecule; the edge set E = {ε} ij},= <v i ,v j > represents the chemical bonds between atoms, and ε in the edge set. ij Represents v i and v j Chemical bonds between nodes (atoms). Each node v i With ε ij Each has its own characteristic h i and e ij h i The unique thermal encoding refers to the unique thermal encoding of the atom type, the relative atomic mass, and the number of implicit hydrogen atoms linked, e ij One-hot encoding and bond length of the bond type (including single bond, double bond, triple bond and no bond).

[0051] Furthermore, since the underlying network structure used in the adversarial coding graph generation model of this invention allows Euclidean space coordinate information to be used as node features, the coordinates of each node (atom) X = {x i It also participates in the description of molecular structural features as a special node feature.

[0052] Step 2: Based on the graph structure, use the E(n) variable graph neural network with attention mechanism as the basic structure to build an adversarial coding graph generation model and train it.

[0053] Adversarial coding graph generation models use E(n) equivariant graph neural networks with attention mechanisms as their basic structure. The E(n) equivariance of E(n) equivariant graph neural networks allows for the simultaneous acceptance and processing of the graph's topology and the Euclidean space coordinates of the nodes, enabling the adversarial coding graph generation model to extract information from the molecular structure more comprehensively and enhancing its multifaceted generation capabilities.

[0054] Based on the basic variational autoencoder, adversarial constraint encoding is used as a transition between the prior distribution and the sample distribution, simplifying the model's learning process and improving its generative capabilities. For example... Figure 2 As shown, the adversarial coding graph generation model has a similar structure to the variational autoencoder in the region covered by the yellow arrow. However, in addition to the original encoder and decoder, the adversarial coding graph generation model introduces a generator and a discriminator, and uses the generator and discriminator to jointly constrain the latent space distribution. Both the generator and the discriminator are multi-layer neural networks. The generator uses g... θ This indicates that a discriminator is used with parameter θ; the discriminator is controlled by f. w This indicates that it has the parameter w.

[0055] During the training of an adversarial coding graph generation model, the generator receives samples from a prior distribution as input and outputs a series of latent space coordinates. These latent space coordinates, along with the coordinates output by the encoder from the molecular structure, are input to the discriminator. The discriminator is responsible for determining whether the received series of coordinate samples originates from the latent space constructed by the encoder (also known as the true latent space) or from the latent space generated by the generator (also known as the false latent space). During this process, to achieve the adversarial constraint, the generator and encoder strive to maintain similar distributions to cause the discriminator to make incorrect judgments, while the discriminator tries to distinguish the source of its input as accurately as possible. Since the generator and discriminator have opposite goals, an adversarial relationship exists between them, and the convergence between the generator and encoder constitutes a constraint on the latent space distribution.

[0056] Considering node features v, edge features ε, atom coordinates x, and property conditions c, the encoding process of this substance can be represented as p φ (z|v,ε,x,c), the decoding process can be represented as: Similar to variational autoencoders, adversarial coding graph generation models require adjusting encoder and decoder parameters to minimize the reconstruction error L. rec ;

[0057] The generator takes a sample s from the prior distribution as input and outputs an adversarial constraint distribution. To make the distribution of multiple samples after encoder transformation converge with that of the generator, it is necessary to minimize the latent space distribution of the encoder output by adjusting various parameters of the encoder. With generator output distribution Wasserstein distance between The Wasserstein distance, calculated by the discriminator, is the difference between the discriminator's judgment of the true latent space distribution output by the encoder and the false latent space distribution generated by the generator. This process is also known as minimizing the coding error L. enc On the other hand, in order to overcome the constraints, the generator and encoder adjust their own parameters together to deceive the discriminator as much as possible, making it unable to distinguish whether the received input comes from the encoder or the generator. This process is called minimizing the generation error. In contrast, the discriminator optimizes its own parameters as much as possible to distinguish whether the input comes from the encoder or the generator; this process is called minimizing the discrimination error L. cri .

[0058] For performance evaluation of adversarial coding graph generation models, the QM9 and ZINC databases are commonly used to train QM9-based and ZINC-based adversarial coding graph generation models, respectively. All 133,885 molecules in the QM9 database were used to train the QM9-based model. Molecules in the QM9 dataset consist of C, H, O, N, and F atoms, with a maximum of 9 heavy atoms (non-hydrogen atoms). 708,521 molecules in the ZINC dataset, composed of C, H, O, N, F, S, and P atoms, with 7–18 heavy atoms, were selected for training the ZINC-based model. RDKit was used to read and process the database to obtain bond types and other information, resulting in the data structure used to train the adversarial coding graph generation model. Some hyperparameters of the two generative models trained on different databases are shown in Table 1. The latent feature dimension in Table 1 is the node feature dimension in AEGNN.

[0059] Table 1. Hyperparameters of the adversarial coding graph generation model

[0060] parameter QM9-based generative model Generative models based on ZINC Maximum number of atoms 9 18 Dimension expansion factor 3 5 Hidden feature dimension 20 40 Latent vector dimension 6 12 AEGNN layers 3 2

[0061] The network structures of the generative models trained using QM9 and ZINC are shown in Tables 2 and 3, respectively. The "Structure" column in Tables 2 and 3 describes the substructure of each adversarial coding graph generative model component. "FC(y→z)" represents a fully connected layer (FC) that transforms y input features into z output features. "BN" refers to a batch normalization layer (BN) used to stabilize the training of the deep neural network. All AEGNN layers share the same structure; except for the final FC layer, each layer is connected to subsequent layers via the ELU activation function.

[0062] Table 2. Network Structure of the Adversarial Coding Graph Generation Model Based on QM9

[0063]

[0064]

[0065] Table 3. Network Structure of the ZINC-based Adversarial Coding Graph Generation Model

[0066]

[0067]

[0068] In the encoder, V represents the original node features, and the initial embedding layer is responsible for changing the dimensions of the node features to adapt to the feature extraction of subsequent embedding layers. (i) e (i) x (i) These represent the hidden node features, hidden edge features, and hidden coordinates in the feature extraction process, respectively. Multiple AEGNNs then extract the structural information. The node features output by the AEGNNs are passed through a mean output layer and a variance output layer, and a reparameterization process to obtain the latent vector z of that node. The latent vector z is then input to the decoder. In the decoder, the feature reconstruction layer is responsible for processing the latent space coordinates into a data structure acceptable to the embedding layer. It consists of three sub-networks, responsible for reconstructing the number of atoms, node features, and node coordinates, respectively. Similar to the encoder, the node features received by the AEGNNs in the decoder are concatenated with molecular performance indicators, and then passed through multiple AEGNNs and the final embedding layer to reconstruct the molecular structure. The remaining components have relatively simple structures, all being multi-layer fully connected (FC) structures.

[0069] The node features of the input adversarial coding graph generation model are vectors of the number of atom types × dimension expansion + 3 (one-heat encoding of implicit hydrogen quantity) + 1 (relative atomic mass). The input to an edge operation is the edge feature and the features of the two nodes connected to it; therefore, the input dimension is the sum of the dimensions of the two connected nodes and the edge feature dimension. The output dimension can be customized and remains the same as the node feature dimension. The input to a node operation is its own feature dimension and the mean of the operation outputs of the edges connecting to it; therefore, the input dimension is the sum of the node feature dimension and the edge operation output dimension, and the output dimension is the node feature dimension. The input to the attention mechanism model is the output of the edge operation; therefore, the input dimension equals the edge operation output dimension, and the output is a scalar (dimension 1) representing the attention coefficient for that edge. The mean output layer and variance output layer, as reparameterization processes, calculate the latent vector for each node; therefore, the final latent vector of a molecule is a matrix, where the number of rows is the latent vector dimension and the number of columns is the maximum number of atoms.

[0070] When using an adversarial coding graph generation model for property conditional generation, the structural information received by the mean and variance output layers involves concatenating each node feature with the molecule's performance index. Taking a five-dimensional array as an example, the node feature dimension input to the mean and variance output layers is 5 dimensions greater than the node's embedding vector dimension, while the output dimension remains the latent space dimension. Before inputting the entire molecule's latent vector into the atom number reconstruction model in the feature reconstruction layer, the latent vector needs to be transformed from a matrix into a vector, resulting in a unique-hot encoded atom vector (a vector with the dimension of the maximum atom number). The node feature and node coordinate reconstruction submodules in the feature reconstruction layer also need to receive the molecule's performance index as input, thus the actual input dimension is 5 times the maximum atom number dimension greater than the latent vector dimension. Similarly, the generator and discriminator also need to process the molecule's latent vector as fully connected layers, therefore the input and output dimensions of the generator and discriminator need to be modified according to the actual input dimensions. The discriminator's output is processed using the Sigmoid function to approximate a 0 / 1 variable, consistent with its function of determining the data source.

[0071] During training, the dataset is divided into training, testing, and validation sets, with the training set comprising 70% of the total dataset, the validation set 20%, and the testing set 10%. The model is trained using the training set. Model parameters are saved when the loss on the validation set reaches its minimum. Training stops when the loss on the validation set fails to decrease for three consecutive iterations. The model's generalization ability to new data is then observed using the testing set. The batch size is set to 10, and the training process iterates across the entire dataset 500 times. The ADAM algorithm is used to optimize model parameters, with a learning rate of 1×10⁻³ for the encoder and decoder, and 1×10⁻⁶ for the generator and discriminator, with weight decay of 1×10⁻⁶ for each batch. The generator is trained twice per batch, the discriminator four times, and the remaining models once. The learning rate decays every four iterations, reducing it to 80% of its current value. Model parameters are initialized using the Xavier method, with the 'gain' parameter set to 0.7.

[0072] Step 3: Based on the trained adversarial coding graph generation model, generate the target molecular working fluid by using the adversarial regularization constraint mechanism and the dynamic priority asynchronous interpretation strategy.

[0073] The application process of the adversarial coding graph generation model is as follows: Figure 3 As shown, when generating property conditional expressions, sampling is first performed from the prior distribution. The sampled data is then input into the generator. The generator's output and the generation conditions are then processed by the feature reconstruction layer to reconstruct the matrices required by the decoder. The output of the feature reconstruction layer is then decoded by the decoder and interpreted using a dynamic priority asynchronous interpretation strategy to obtain the molecular structure.

[0074] like Figure 4As shown, this invention proposes a dynamic priority asynchronous interpretation strategy based on the asynchronous method. The core improvements of this method include (1) sampling the valence state of the atom through a probability distribution before sampling the chemical bonds connected to the atom; and (2) when sampling the chemical bond, considering not only the atom currently being sampled, but also the valence state of the atom at the other end of the chemical bond. The dynamic priority asynchronous interpretation strategy is carried out step by step, striving to ensure the connectivity of the final sampled molecular map and avoid the appearance of isolated atoms or isolated groups. The specific steps are as follows:

[0075] Step 1: Determine all atom types and the number of implicit hydrogen atoms, and determine the maximum valence state using the atomic valence state algorithm. Based on the maximum valence state and the existing number of implicit hydrogen atoms, set the probability of all impossible bonds to 0. The core idea of ​​the atomic valence state algorithm is that erroneous nodes and edges have higher cross-entropy.

[0076] The atomic valence state algorithm is as follows: For an atom with multiple common valence states, the chemical bond connection is calculated based on each valence state, and the mean of the probability of each chemical bond sampled under that valence state is calculated. The mean of the probability of each chemical bond under each valence state is compared, and the valence state with the largest mean is selected as the valence state of the atom in the molecule.

[0077] Step 2: Sample a single bond for all atoms with a remaining valence of 1.

[0078] Step 3: Sample a bond for an atom with a maximum valence state of 2 that has already been bonded in step 2.

[0079] Step 4: Sample a bond for an atom that is not yet connected to any chemical bond.

[0080] Step 5: Sample the chemical bonds of all connected atoms that have not reached their maximum valence state until all atoms have reached their maximum valence state, thus completing the interpretation.

[0081] When sampling chemical bonds, only bonds that conform to chemical rules are sampled according to the atom types on both sides of the chemical bond. After sampling, the probability of all chemical bonds between the two atoms connected by the chemical bond is set to 0, and the current valence state of the atoms on both sides of the chemical bond is updated according to the type of chemical bond.

[0082] The present invention will be further described in detail below through specific embodiments:

[0083] Example 1:

[0084] To design molecular electrolytes for lithium metal batteries using adversarial coding graph generation models, a structure-performance dataset of electrolytes needs to be constructed for training the adversarial coding graph generation model. Inspired by the group contribution method, this invention proposes a molecular classification derivation method. For the five most critical performance indicators of electrolytes—conductivity, flame retardancy, anodic / cathode stability, and wettability—a series of common electrolyte solvents, such as… Figure 5 As shown, the performance of potential electrolyte additives is predicted using molecular classification derivation based on the performance of several electrolyte solvents. To construct the dataset, this invention selected 14,019 molecules from the ZINC database according to the following criteria: containing fewer than 18 heavy atoms, containing one or more atoms of F, P, or S, and having at least one characteristic functional group. Molecules containing phosphorus atoms underwent 40 data augmentations, and molecules containing sulfur atoms underwent 30 data augmentations. After augmentation, the final dataset consists of 154,649 molecules. In addition to molecular structural information, this dataset also includes hypothetical performance predicted by molecular classification derivation.

[0085] When generating property conditional formulas, this invention starts from a baseline total score and gradually increases the scores of each property until further improvement of any property cannot yield a feasible molecular structure. For example... Figure 6 As shown, the formation conditions of 2,4-bis(2-fluoroethoxy)tetrafluorocyclotriphosphazene (DFEPN) are specifically designed to address the core issues of lithium metal batteries. This molecule should first have excellent flame retardancy (flame retardancy is scored 5 points), and at the same time, it should have excellent electrode stability (each is scored 4.5 points). As an additive, DFEPN does not need to participate in the energy conversion reaction during battery charge-discharge cycles, but it should also avoid adversely affecting battery performance as much as possible. Therefore, its wettability and conductivity are scored 4.5 points, and the total score for the properties is 23 points.

[0086] Example 2:

[0087] This invention also provides a design system for molecular working fluids, such as Figure 7 As shown, the system includes: an abstraction module, a modeling module, and a generation module.

[0088] The abstract module is used to obtain the design target of the target molecular working fluid, and abstracts the atoms and chemical bonds in the molecule into a graph structure composed of nodes and edges;

[0089] The modeling module is used to construct and train an adversarial coding graph generation model based on the graph structure, using an E(n) variable graph neural network with attention mechanism as the basic structure.

[0090] The generation module is used to generate the target molecular working fluid based on the trained adversarial coding graph generation model, utilizing adversarial regularization constraint mechanism and dynamic priority asynchronous interpretation strategy.

[0091] It is understood that the molecular working fluid design system provided by the present invention corresponds to the molecular working fluid design method provided in the foregoing embodiments. The relevant technical features of the molecular working fluid design system can be referred to the relevant technical features of the molecular working fluid design method, and will not be repeated here.

[0092] Another object of the present invention is to provide an electronic device, such as... Figure 8 As shown, it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor performing the steps of the design method for the molecular working fluid.

[0093] The design method for the molecular working fluid includes the following steps:

[0094] To obtain the design target of the target molecular working fluid, the atoms and chemical bonds in the molecule are abstracted into a graph structure composed of nodes and edges;

[0095] Based on the graph structure, an adversarial coding graph generation model is constructed and trained using an E(n) variable graph neural network with attention mechanism as the basic structure.

[0096] Based on a trained adversarial coding graph generation model, the target molecular working fluid is generated by utilizing an adversarial regularization constraint mechanism and a dynamic priority asynchronous interpretation strategy.

[0097] A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the design method for the molecular working fluid.

[0098] The design method for the molecular working fluid includes the following steps:

[0099] To obtain the design target of the target molecular working fluid, the atoms and chemical bonds in the molecule are abstracted into a graph structure composed of nodes and edges;

[0100] Based on the graph structure, an adversarial coding graph generation model is constructed and trained using an E(n) variable graph neural network with attention mechanism as the basic structure.

[0101] Based on a trained adversarial coding graph generation model, the target molecular working fluid is generated by utilizing an adversarial regularization constraint mechanism and a dynamic priority asynchronous interpretation strategy.

[0102] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0103] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0104] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0105] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for designing molecular working fluids, characterized in that, include: To obtain the design target of the target molecular working fluid, the atoms and chemical bonds in the molecule are abstracted into a graph structure composed of nodes and edges; Based on the graph structure, an adversarial coding graph generation model is constructed and trained using an E(n) variable graph neural network with attention mechanism as the basic structure. Based on a trained adversarial coding graph generation model, the target molecular working fluid is generated by utilizing an adversarial regularization constraint mechanism and a dynamic priority asynchronous interpretation strategy.

2. The design method for a molecular working fluid according to claim 1, characterized in that, In the graph structure, nodes represent the atoms in a molecule, and edges represent the chemical bonds between the atoms. Each node has its own unique thermal code for its atom type, its relative atomic mass, and the number of implicit hydrogen atoms it connects. Each chemical bond has its own unique thermal code for its bond type and its bond length.

3. The design method for a molecular working fluid according to claim 1, characterized in that, The adversarial coding graph generation model includes an encoder, a decoder, a generator, and a discriminator. The encoder includes an embedding layer, an AEGNN layer, a mean output layer, a variance output layer, and a reparameterization layer. The decoder includes a node reconstruction layer, a coordinate reconstruction layer, an atom number reconstruction layer, an AEGNN layer, and an embedding layer; Both the generator and the discriminator are multi-layer neural networks, and the generator and the discriminator jointly constrain the latent space distribution.

4. The design method for a molecular working fluid according to claim 1, characterized in that, The training method for the adversarial coding graph generation model is as follows: In the adversarial coding graph generation model, the generator takes samples of the prior distribution as input and outputs a series of latent space coordinates. The latent space coordinates are input together with the coordinates output by the encoder of the molecular structure in the adversarial coding graph generation model. The discriminator is responsible for determining whether the series of received coordinates are sampled in the true latent space formed by the encoder or in the false latent space generated by the generator. To achieve adversarial constraints, the generator and encoder try to maintain similar distributions so that the discriminator makes incorrect judgments, while the discriminator tries to distinguish the source of its input as accurately as possible.

5. The method for designing a molecular working fluid according to claim 1, characterized in that, Dynamic priority asynchronous interpretation strategies include: Determine all atom types and the number of implicit hydrogens, and determine the maximum valence state using an atom valence state algorithm. Based on the maximum valence state and the existing number of implicit hydrogens, set the probability of all impossible chemical bonds to 0. Sample a single bond for all atoms with a remaining valence of 1, sample a chemical bond for atoms with a maximum valence of 2 that have already been bonded, and sample a chemical bond for atoms that have not yet been bonded. Sample chemical bonds for all atoms whose connected chemical bonds have not reached their maximum valence state until all atoms have reached their maximum valence state.

6. The method for designing a molecular working fluid according to claim 5, characterized in that, The algorithm for atomic valence states is as follows: For an atom with multiple common valence states, the chemical bond connections are calculated based on each valence state, and the mean of the probabilities of each chemical bond sampled under that valence state is calculated. The mean probabilities of each chemical bond under each valence state are compared, and the valence state with the largest mean is selected as the valence state of that atom in the molecule.

7. The method for designing a molecular working fluid according to claim 1, characterized in that, The process of generating the target molecular working fluid using an adversarial coding graph generation model includes: Sampling is performed from the prior distribution, and the sampled data is input into the generator in the adversarial coding graph generation model. The output of the generator and the generation conditions are then processed by the feature reconstruction layer of the decoder in the adversarial coding graph generation model to reconstruct the various matrices required by the decoder. The output of the feature reconstruction layer is then decoded by the decoder and interpreted using a dynamic priority asynchronous interpretation strategy to obtain the molecular structure.

8. A design system for molecular working fluids, characterized in that, include: The abstract module is used to obtain the design target of the target molecular working fluid, and abstracts the atoms and chemical bonds in the molecule into a graph structure composed of nodes and edges; The modeling module is used to construct and train an adversarial coding graph generation model based on the graph structure, using an E(n) variable graph neural network with attention mechanism as the basic structure. The generation module is used to generate the target molecular working fluid based on the trained adversarial coding graph generation model, utilizing adversarial regularization constraint mechanism and dynamic priority asynchronous interpretation strategy.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images