Generative adversarial network-based molecule generation method, apparatus and device, medium and product

Abstract

The invention discloses a molecule generation method and device based on a generative adversarial network, equipment, a medium and a product, and relates to the technical field of computational chemistry and batteries. The method comprises the following steps: constructing a molecule database by adopting a matrix engineering method according to atom types and quantity required by target molecules; converting a molecular structure corresponding to the candidate molecular set in the molecular database into a graph structure to obtain a molecular graph database; inputting the molecular graph database into the trained generative adversarial network, and performing molecule generation and decoding to obtain newly generated molecules; based on a preset screening rule, performing molecular screening on the newly generated molecules to obtain screened qualified molecules; converting the screened qualified molecules into three-dimensional structures to obtain three-dimensional molecules; the three-dimensional molecules are used for computational simulation and experimental verification to provide a reference basis for electrolyte design and optimization. According to the invention, molecule generation and screening can be efficiently realized.

Description

Technical Field

[0001] This application relates to the fields of computational chemistry and battery technology, and in particular to a molecular generation method, apparatus, device, medium and product based on generative adversarial networks. Background Technology

[0002] Electrolytes play a crucial role in battery performance optimization, with their ionic conductivity, chemical stability, and interfacial compatibility directly impacting cycle life and energy efficiency. High-quality electrolytes not only require efficient ion transport but also the formation of a stable solid electrolyte interphase (SEI) film on the electrode surface, thereby enhancing the battery's long-term stability and safety. However, traditional electrolyte screening methods rely heavily on extensive experimental testing, which is not only inefficient but also limited by a finite pool of candidate molecules, hindering the rapid discovery of novel, high-performance molecules. Summary of the Invention

[0003] The purpose of this application is to provide a molecular generation method, apparatus, device, medium, and product based on generative adversarial networks, which can efficiently realize molecular generation and screening.

[0004] To achieve the above objectives, this application provides the following solution:

[0005] In a first aspect, this application provides a molecular generation method based on generative adversarial networks, comprising:

[0006] A molecular database is constructed using matrix engineering methods based on the required atom types and quantities for the target molecule;

[0007] The molecular structures corresponding to the candidate molecule sets in the molecular database are transformed into graph structures to obtain a molecular graph database;

[0008] The molecular graph database is input into the trained generative adversarial network (GAN) to generate and decode molecules, resulting in newly generated molecules. The GAN is trained using the Adam optimizer, which optimizes the objective function through adversarial game theory and incorporates a gradient penalty term. The GAN includes a generator and a discriminator.

[0009] Based on preset screening rules, the newly generated molecules are subjected to molecular screening to obtain qualified molecules;

[0010] Qualified molecules are transformed into three-dimensional structures to obtain three-dimensional molecules; these three-dimensional molecules are used for computational simulation and experimental verification to provide a reference for electrolyte design and optimization.

[0011] In one embodiment, a matrix engineering approach is used to construct a molecular database based on the required atom types and quantities for the target molecule, specifically including:

[0012] A matrix engineering method is used to generate a molecular set based on the required atom types and quantities for the target molecule. The molecular set is represented in matrix form. The first row and first column of the matrix are used to represent the atom types, and the remaining elements in the matrix are used to represent the bonding relationships. Among them, 0 represents no bonding, 1 represents a single bond, and 2 represents a double bond.

[0013] By traversing the elements of the upper triangular region of the matrix using a backtracking algorithm and combining it with valence bond constraint rules, a SMILES molecular structure that conforms to chemical laws is generated.

[0014] The molecular structures of SMILES were screened for connectivity and functional groups to determine a set of candidate molecules;

[0015] A molecular database is determined based on the set of candidate molecules.

[0016] In one embodiment, the molecular structures corresponding to the candidate molecule sets in the molecular database are transformed into graph structures to obtain a molecular graph database, specifically including:

[0017] The SMILES molecular structure is transformed into a graph structure, where atoms in the SMILES molecule serve as nodes in the graph structure, chemical bonds serve as edges, node features are used to represent atom type, electronegativity and hybridization state, and edge features are used to represent bond type and bond order.

[0018] In one embodiment, the qualified molecules are converted into three-dimensional structures to obtain three-dimensional molecules, specifically including:

[0019] The EmbedMolecule command from the RDKit library is used to convert qualified molecules into three-dimensional structures, and hydrogen atoms are automatically added using the Chem.AddHs() method to obtain three-dimensional molecules.

[0020] In one embodiment, the objective function includes: the objective function of the generator and the objective function of the discriminator;

[0021] The objective function of the discriminator, and its corresponding mathematical expression, is:

[0022] L D =-E x～pdata [logD(x)]-E z～pz [log(1-D(G(z)))];

[0023] The objective function of the generator, and its corresponding mathematical expression, is:

[0024] L G =-Ez～pz [logD(G(z))];

[0025] Among them, L D E is the objective function of the discriminator. x～pdata D(x) is the expectation of the true data distribution pdata(x); D(x) is the output of the discriminator for the true molecular sample x; E z～pz Let pz(z) be the expectation of the noise distribution; G(z) be the generator function that takes the noise vector z as input and outputs the generated molecular data; x is the real molecular sample; z is the noise vector; L G The objective function of the generator.

[0026] In one embodiment, the mathematical expression corresponding to the gradient penalty term is:

[0027]

[0028] Where L1 is the gradient penalty term; E x～pg D(x) is the expectation of the generated data distribution pg(x); D(x) is the output of the discriminator for the real molecular sample x; E x～pr Let be the expectation of the true data distribution pr(x); λ be the gradient penalty coefficient; E x～px Let be the expectation of the interpolation distribution px(x); Let ||x| be the gradient of the discriminator with respect to the input real molecular sample x; ||·||2 is the L2 norm.

[0029] Secondly, this application provides a molecular generation device based on a generative adversarial network, comprising:

[0030] The molecular database construction module is used to construct a molecular database based on the required atom types and numbers for the target molecule using matrix engineering methods.

[0031] The conversion module is used to convert the molecular structures corresponding to the candidate molecule sets in the molecular database into graph structures, thus obtaining a molecular graph database.

[0032] The molecule generation and decoding module is used to input the molecular graph database into the trained generative adversarial network (GAN) to generate and decode molecules, thereby obtaining newly generated molecules. The GAN is obtained by using the Adam optimizer, optimizing the objective function through adversarial game theory, and adding a gradient penalty term during training. The GAN includes a generator and a discriminator.

[0033] The molecular screening module is used to perform molecular screening on newly generated molecules based on preset screening rules to obtain qualified molecules.

[0034] The three-dimensional structure conversion module is used to convert qualified molecules into three-dimensional structures to obtain three-dimensional molecules. These three-dimensional molecules are used for computational simulation and experimental verification to provide a reference for electrolyte design and optimization.

[0035] Thirdly, this application provides a computer device, including: 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 above-described molecular generation method based on generative adversarial networks.

[0036] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned molecular generation method based on generative adversarial networks.

[0037] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned molecular generation method based on generative adversarial networks.

[0038] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0039] This application provides a molecular generation method, apparatus, device, medium, and product based on generative adversarial networks (GANs). It employs matrix engineering to construct a molecular database based on the required atom types and quantities for the target molecule. The molecular structures corresponding to candidate molecule sets in the molecular database are transformed into graph structures, resulting in a molecular graph database. This molecular graph database is then input into a trained GAN for molecular generation and decoding, yielding newly generated molecules. Based on preset screening rules, the newly generated molecules are screened to obtain qualified molecules. Finally, the qualified molecules are transformed into three-dimensional structures, resulting in three-dimensional molecules. This application, while utilizing matrix engineering to rapidly generate a molecular database, introduces a generative adversarial network. Through adversarial training between the generator and discriminator, molecular generation and screening can be achieved efficiently. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of a molecular generation method based on generative adversarial networks;

[0042] Figure 2 A schematic diagram illustrating the principle of adversarial networks;

[0043] Figure 3 This is a flowchart of the molecular generation process;

[0044] Figure 4 A schematic diagram illustrating the derivation of new molecules from basic molecules;

[0045] Figure 5 This is a structural diagram of a molecular generation device based on a generative adversarial network.

[0046] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] A matrix engineering-based molecular generation method maps atom types and bonding relationships to matrix elements, and uses backtracking and connectivity detection algorithms to systematically generate large-scale molecular sets that conform to chemical rules, thereby effectively expanding the molecular space and providing a new candidate molecular library for electrolyte design. While this method significantly improves the efficiency and completeness of molecular generation, it still relies heavily on combinatorial rules and lacks the ability to learn molecular distribution patterns, thus limiting its diversity and specificity. Furthermore, because this method requires global enumeration of the matrix during generation, the matrix size grows exponentially when the number of atoms exceeds 10, leading to a sharp increase in computational complexity and preventing the generation of molecular structures with more than 10 atoms.

[0049] Building upon matrix engineering methods, this paper further introduces Generative Adversarial Networks (GANs). Deep learning models are used to train and model the molecular database generated by the matrix, enabling the model to capture molecular topology and distribution characteristics, and generate new candidate molecules based on these characteristics. Through this combined approach, this application retains the advantages of matrix engineering in terms of generative completeness and rule constraints, while overcoming its shortcomings in molecular diversity and application adaptability, thus achieving more efficient and intelligent molecular generation and screening.

[0050] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] In one exemplary embodiment, such as Figure 1 As shown, a molecular generation method based on generative adversarial networks is provided, including:

[0052] Step 100: Construct a molecular database based on the required atom types and quantities for the target molecule using matrix engineering methods.

[0053] Specifically, a matrix engineering approach is used to construct a molecular database based on the required atom types and quantities for the target molecule. This includes:

[0054] A matrix engineering method is used to generate a molecular set based on the required atom types and quantities for the target molecule. The molecular set is represented in matrix form. The first row and first column of the matrix are used to represent the atom types, and the remaining elements in the matrix are used to represent the bonding relationships. Among them, 0 represents no bonding, 1 represents a single bond, and 2 represents a double bond.

[0055] By traversing the elements of the upper triangular region of the matrix using a backtracking algorithm and combining it with valence bond constraint rules, a SMILES molecular structure that conforms to chemical laws is generated.

[0056] The molecular structures of SMILES were screened for connectivity and functional groups to determine a set of candidate molecules; a molecular database was then established based on the set of candidate molecules.

[0057] Step 200: Convert the molecular structures corresponding to the candidate molecule sets in the molecular database into graph structures to obtain the molecular graph database.

[0058] This involves transforming the molecular structures corresponding to the candidate molecule sets in the molecular database into graph structures to obtain a molecular graph database, specifically including:

[0059] The SMILES molecular structure is transformed into a graph structure, where atoms in the SMILES molecule serve as nodes in the graph structure, chemical bonds serve as edges, node features are used to represent atom type, electronegativity and hybridization state, and edge features are used to represent bond type and bond order.

[0060] Step 300: Input the molecular graph database into the trained generative adversarial network (GAN) to generate and decode molecules, obtaining newly generated molecules. The GAN is trained using the Adam optimizer, which optimizes the objective function through adversarial game theory and incorporates a gradient penalty term; the GAN consists of a generator and a discriminator.

[0061] The objective functions include: the objective function of the generator and the objective function of the discriminator; the objective function of the discriminator, corresponding to the mathematical expression, is:

[0062] L D =-E x～pdata[logD(x)]-E z～pz [log(1-D(G(z)))].

[0063] The objective function of the generator, and its corresponding mathematical expression, is:

[0064] L G =-E z～pz [logD(G(z))].

[0065] Among them, L D E is the objective function of the discriminator. x～pdata D(x) is the expectation of the true data distribution pdata(x); D(x) is the output of the discriminator for the true molecular sample x; E z～pz G(z) represents the expectation of the noise distribution pz(z); G(z) is the generator function that takes the noise vector z as input and outputs the generated molecular data; x is the real molecular sample; z is the noise vector, usually a random variable; L G The objective function of the generator.

[0066] The mathematical expression for the gradient penalty term is:

[0067]

[0068] Where L1 is the gradient penalty term; E x～pg D(x) is the expectation of the generated data distribution pg(x); D(x) is the output of the discriminator for the real molecular sample x; E x～pr Let be the expectation of the true data distribution pr(x); λ be the gradient penalty coefficient; E x～px Let be the expectation of the interpolation distribution px(x); Let λ be the gradient of the discriminator with respect to the input real molecular sample x; ||·||2 is the L2 norm. λ is the gradient penalty coefficient, in other words, a hyperparameter used to control the weight of the gradient penalty term; ||·||2 is the L2 norm (Euclidean norm), i.e., the vector length.

[0069] Step 400: Based on the preset screening rules, perform molecular screening on the newly generated molecules to obtain qualified molecules.

[0070] Step 500: The selected molecules are converted into three-dimensional structures to obtain three-dimensional molecules. These three-dimensional molecules are used for computational simulations and experimental verification, providing a reference for electrolyte design and optimization.

[0071] The process of converting qualified molecules into three-dimensional structures to obtain three-dimensional molecules specifically includes:

[0072] The EmbedMolecule command from the RDKit library is used to convert qualified molecules into three-dimensional structures, and hydrogen atoms are automatically added using the Chem.AddHs() method to obtain three-dimensional molecules.

[0073] Specifically, the operation process of the method mentioned in this application is as follows:

[0074] First, a large-scale molecular dataset is generated using matrix engineering algorithms. By traversing the elements of the upper triangular region of the matrix and combining valence bond constraint rules, SMILES molecular structures conforming to chemical laws are obtained. These molecules are then further screened for connectivity and functional groups, ultimately constructing a high-quality initial molecular database. Next, this molecular database is transformed into a graph structure, where atoms in the molecules serve as nodes, chemical bonds as edges, node features represent atom type, electronegativity, hybridization state, etc., and edge features represent bond type and bond order. Using this as a training dataset, an adversarial generative neural network model is constructed and initialized. Through training on this database, the model learns the interaction patterns between atoms and the distribution patterns of molecular topologies. In the molecular generation stage, random noise or some known molecular fragments are input into the trained graph neural network generator, and new molecular graph structures are obtained through decoding. The generated molecular graphs are then screened according to chemical rules and application scenarios, eliminating molecules that do not meet valence bond constraints, connectivity requirements, or are undesirable in specific electrolyte applications. The retained molecular graphs are transformed into SMILES molecular representations and further generated into three-dimensional mol structures for subsequent calculations and analysis. This application achieves efficient, systematic and intelligent generation of electrolyte molecules through a combination of "matrix engineering + GAN". Compared with traditional methods that rely on experiments to discover superior electrolytes, this method is more efficient, comprehensive and has adaptive optimization capabilities.

[0075] S1. A candidate molecule set is generated using matrix engineering methods, where the atom type and bonding relationship are encoded by matrix elements. The backtracking and connectivity detection algorithms are combined to obtain SMILES molecules that conform to chemical rules and to build a high-quality molecular database.

[0076] S2. Transform the molecular database into a molecular graph structure, where nodes represent atoms with atomic features and edges represent chemical bonds with bond features.

[0077] S3. Construct and train a generative adversarial network (GAN) model based on a molecular graph database, enabling it to learn the distribution patterns of molecular structures and the connection patterns between atoms.

[0078] S4. In the molecular generation stage, random noise vectors or partial molecular fragments are input into the trained GAN model, and new molecular maps are obtained through decoding.

[0079] S5. Screen the newly generated molecular diagrams for effectiveness and application, including valence bond constraints, ring structure rationality, functional group detection, and electrolyte compatibility.

[0080] S6. Convert the qualified molecular diagrams into SMILES molecular representations and further generate the corresponding three-dimensional mol structures for subsequent calculations and analysis.

[0081] This application, based on the matrix algorithm for rapidly constructing a molecular database, introduces a generative adversarial network (GAN) model. Through adversarial training between the generator and discriminator, it achieves efficient generation and screening of battery electrolyte molecules, thereby improving the efficiency, diversity, and rationality of molecule generation. The process includes the following steps:

[0082] First, an initial set of molecules is constructed using matrix engineering methods. By defining the mapping rules between the rows and columns of the matrix and the atom types and bonding relationships, candidate molecules that conform to basic chemical rules are generated and transformed into a high-quality molecular database.

[0083] Secondly, the molecule is transformed into a graph structure, where nodes represent atoms and are accompanied by features such as atom type, electronegativity, and hybridization state, and edges represent chemical bonds and are accompanied by features such as bond type and bond order.

[0084] Then, the generative adversarial network is trained using a molecular graph database. The generator receives a random noise vector and outputs the adjacency matrix and node features of the molecules. The discriminator distinguishes between real and fake input molecular graphs, thereby guiding the generator to learn molecular topological rules and atomic connection patterns.

[0085] Finally, the new molecules output by the generator are screened for effectiveness and application scenarios. Molecules with unreasonable valence bond constraints, abnormal ring structures, or unsuitability as electrolytes are removed. The qualified molecules are then converted into SMILES representations and corresponding three-dimensional mol files are generated for subsequent calculations and analysis.

[0086] The generator employs a multi-layer fully connected network, combined with normalization, batch normalization, and LeakyReLU activation functions to enhance training stability and expressive power, and outputs the adjacency matrix and node features of the molecular graph.

[0087] The discriminator determines the authenticity of the molecular structure by flattening and splicing the adjacency matrix and node features, and then using a fully connected network to ensure the rationality of the generated molecule.

[0088] During training, an adversarial learning framework is used to alternately optimize the generator and discriminator, enabling the generator to gradually learn the distribution patterns of real molecules and generate new molecular structures with chemical rationality.

[0089] When screening the chemical rationality of the generated molecules, the following methods are used: valence bond constraint rules (e.g., the number of carbon bonds does not exceed 4 and the number of oxygen bonds does not exceed 2), ring structure rationality check, functional group screening, and electrolyte compatibility test.

[0090] Qualified molecules are then subjected to hydrogen atom completion and 3D coordinate optimization using Python's RDKit library to generate complete and usable 3D molecular structures.

[0091] The resulting SMILES molecules are stored in a database to avoid duplication and can be targeted for screening and optimization according to the needs of different battery systems.

[0092] The molecular generation method based on matrix engineering and generative adversarial networks provided in this application can be applied to battery electrolyte screening.

[0093] Compared with traditional experimental screening methods, this application can generate a large number of candidate molecules in a shorter time, which greatly improves the screening efficiency. Compared with molecular generation methods based solely on matrix engineering, this application, through the introduction of GAN, not only improves the diversity of molecular generation and enables the effective generation of molecular structures with more than 10 atoms, but also enhances the rationality and applicability of molecular structures, making the generated candidate molecules more in line with the performance requirements of actual electrolytes.

[0094] The results of the examples show that the organic molecules generated by the method of this application have advantages in terms of integrity, diversity and chemical rationality, and can provide a new and efficient solution for electrolyte screening, and have extremely high application value for battery performance optimization and new material discovery.

[0095] This application, building upon the rapid generation of an initial molecular database using matrix engineering methods, introduces a generative adversarial network (GAN) model. First, a large number of molecules conforming to basic chemical rules are automatically generated using matrices and Python scripts. Then, through steps such as valence bond constraints, functional group screening, and connectivity checks, unreasonable molecules are eliminated, constructing a preliminary molecular library. Subsequently, this database is used as training data to conduct adversarial training on the generator and discriminator. This allows the generator to gradually learn the distribution characteristics and connection patterns of molecular structures, thereby enabling the automatic generation of more structurally sound and high-performance new molecules.

[0096] The advantage of this method lies in combining the efficiency of matrix engineering with the intelligent learning capabilities of GANs. The matrix method ensures the integrity and regularity of molecule generation, while the introduction of GANs further enhances the diversity and innovation of molecule generation, avoiding the high repetition rate problem inherent in single matrix methods. These newly generated molecules can be directly used for material property simulation and electrolyte screening, thereby helping to discover potential high-performance candidate solvents.

[0097] This application not only provides more novel molecular candidates but also offers entirely new ideas for experimental research. By combining computation and generative models, a large number of candidate molecules can be rapidly obtained in a short time, and the simulation results can guide experimental design, accelerating the materials development process. This molecular generation method based on matrix engineering and GANs provides new tools and ideas for the design of electrolytes and other functional materials.

[0098] Example 1

[0099] S1: Matrix engineering to build a molecular database.

[0100] First, a matrix representation is constructed based on the required atom types and numbers for the target molecule. The first row and first column of the matrix represent the atom types, while the remaining elements represent bonding relationships, where 0 indicates no bonding, 1 indicates a single bond, and 2 indicates a double bond. A backtracking algorithm is used to traverse the upper triangular region of the matrix, combined with a connectivity detection algorithm and valence bond constraints, to generate SMILES molecules that conform to chemical rules. This process eliminates unconnected molecules, molecules with more than 4 carbon atom bonds, molecules with more than 2 oxygen atom bonds, and functional groups that do not meet the electrolyte design requirements, thus obtaining a high-quality set of candidate molecules and establishing an initial molecular database.

[0101] S2: Molecular diagram representation of transformation.

[0102] The SMILES molecules obtained in step S1 are transformed into a graph structure, where nodes represent atoms in the molecule and include atomic characteristics such as atom type and electronegativity; edges represent chemical bonds and include bond type information (single bond, double bond, etc.). This constructed molecular graph database can completely preserve molecular topological information, providing standardized input for subsequent model training.

[0103] The specific implementation is as follows:

[0104] Each molecule is analyzed as a molecular graph G = (V, E); the set of nodes V represents the atoms in the molecule, and each node contains features such as atom type, atomic number, electronegativity, hybridization type, and whether it is in a ring; the set of edges E represents the chemical bonds between atoms, and each edge contains features such as bond type (single bond, double bond, triple bond, aromatic bond), whether it is in a ring, and bond directionality.

[0105] The transformed molecular graph is stored in the form of a node feature matrix X and an adjacency matrix A.

[0106] This representation not only preserves the molecular topology completely, but also provides standardized input for subsequent GAN modeling.

[0107] S3: Training the generative adversarial network model.

[0108] Based on the molecular graph database constructed in step S2, this embodiment further constructs a generative adversarial network (GAN) model for the automatic generation and optimization of molecular structures. This model consists of a generator and a discriminator, which continuously optimize through adversarial game-like interaction during training, ultimately outputting novel molecules that conform to chemical rules and possess potential for electrolyte applications.

[0109] Generator architecture design:

[0110] The generator takes a random noise vector z∈Rd or a partial molecular fragment as input, and gradually maps it into the adjacency matrix and node feature matrix of the target molecule through a multi-layer fully connected network.

[0111] Input layer: The input noise vector has a dimension of 64 (LATENT_DIM = 64), which is transformed into a high-dimensional latent representation through linear mapping and activation functions.

[0112] Intermediate layers: Multiple linear layers and Batch Normalization are used, with LeakyReLU as the activation function, to enhance the non-linear expressive power of the network and avoid gradient vanishing.

[0113] Output layer: The output is divided into two parts: an adjacency matrix, which is averaged after adding the output matrix and its transpose to ensure the undirectedness of the molecular graph, forming a symmetric matrix; and a node feature matrix, which is normalized by the softmax function so that each node corresponds to only one type of atom.

[0114] The above design ensures that the molecular candidate graph output by the generator not only conforms to the basic topological constraints, but also can generate different structures through random noise.

[0115] Discriminator architecture design:

[0116] After receiving the input molecular graph (adjacency matrix and node feature matrix), the discriminator flattens it into a vector representation and then uses a multilayer perceptron to distinguish between true and false graphs.

[0117] Input processing: Expand and concatenate the adjacency matrix and the node feature matrix to obtain a one-dimensional vector input discriminator.

[0118] Hidden layers: Linear layers employing spectral normalization and the LeakyReLU activation function. Spectral normalization limits the maximum singular values ​​of the weight matrix, effectively improving training stability and preventing gradient explosion.

[0119] Output layer: Outputs a scalar representing the probability that the input molecule is a real sample.

[0120] The goal of the discriminator is to improve its ability to distinguish between real molecules and generated molecules, thereby driving the generator to continuously improve its molecular generation distribution.

[0121] Adversarial training mechanism:

[0122] During training, the generator and discriminator are optimized through adversarial game theory:

[0123] The objective function of discriminator D is:

[0124] L D =-E x～pdata [logD(x)]-E z～pz [log(1-D(G(z)))].

[0125] The objective function of generator G is:

[0126] L G =-E z～pz [logD(G(z))].

[0127] That is, by minimizing the loss function, the generator forces the discriminator to have difficulty distinguishing the generated molecules from the real molecules.

[0128] To further stabilize training, this embodiment introduces the Wasserstein GAN (WGAN) concept, using the scores output by the discriminator to approximate the Wasserstein distance between distributions, and adds a gradient penalty term, which takes the following form:

[0129]

[0130] Model optimization and convergence.

[0131] The Adam optimizer was used with a learning rate of 1e-4, parameter β1 = 0.9, and parameter β2 = 0.9.

[0132] The discriminator is trained 5 times, and the generator is trained once to ensure that the discriminator has sufficient discrimination ability.

[0133] During training, the validity ratio of generated molecules is continuously monitored. If the proportion of valid molecules in the generated results increases, it indicates that the model is gradually converging.

[0134] Through this mechanism, the generator can continuously capture the distribution patterns and interatomic connection patterns of real molecules, thereby generating new molecules that conform to chemical rules; the discriminator, on the other hand, promotes the improvement of the generator by enhancing its discrimination ability. When the two reach a balance in their competition, the model can be used for subsequent molecule generation and decoding.

[0135] S4: Molecular generation and decoding.

[0136] After the generative adversarial network (GAN) has completed training, it enters the molecule generation stage. In this embodiment, a random noise vector or a partial molecular fragment is first fed into the trained generator as input. The adjacency matrix and node feature matrix of the candidate molecules are obtained through multi-layer nonlinear mapping. The adjacency matrix represents the connection relationship between atoms in the molecule. To ensure the symmetry of chemical bonds, the output matrix needs to be averaged with its transpose and then rounded down, and normalized using softmax so that the elements in the matrix can represent different bond types such as "no bond, single bond, double bond". The node feature matrix represents the category and feature of each atom in the molecule. After softmax normalization, it is ensured that each node is uniquely mapped to a specific atom type, such as carbon, oxygen, nitrogen, etc.

[0137] S5: Molecular screening and application validation.

[0138] The newly generated molecules undergo effectiveness and application screening, including valence bond rationality checks, ring structure rationality tests, functional group type screening, and electrolyte compatibility assessments. Through these screening steps, molecules that do not conform to chemical principles or are unsuitable for electrolyte systems are eliminated, while qualified candidate molecules are retained.

[0139] S6: Generation and application of three-dimensional molecular structures.

[0140] For qualified molecules, they are converted into three-dimensional structures (mol or pdb files). In this embodiment, the EmbedMolecule command from the RDKit library is preferably used to generate the three-dimensional structure, and hydrogen atoms are automatically added using the Chem.AddHs() method. The obtained three-dimensional molecules can be further used for computational simulations (such as energy level calculations and solvation energy analysis) and experimental verification, providing an important reference for electrolyte design and optimization.

[0141] Example 2

[0142] like Figure 3 As shown, this embodiment provides an actual operation process for generating molecules using the molecular generation method based on matrix engineering and generative adversarial networks described in Embodiment 1.

[0143] First, import the necessary Python libraries, including RDKit for molecule processing and validity testing, PyTorch and PyTorch Geometric for building and training generative adversarial networks, Pandas for storing and retrieving molecular databases, NumPy and random for numerical computation and random number generation, and os and time for file management and runtime recording. Use time.perf_counter() to record the program's start time and initialize the molecule collection generated_smiles to store the generated SMILES molecules.

[0144] Next, auxiliary functions are constructed to implement molecular screening and efficacy detection. In this embodiment, several functions for chemical rule checking are written, including:

[0145] (1) Valence bond constraint check: Traverse each atom in the generated molecule and count the number of bonds formed. When the valence of carbon atom exceeds 4 or the valence of oxygen atom exceeds 2, the molecule is determined to be invalid.

[0146] (2) Ring structure rationality detection: Call the GetRingInfo() method provided by RDKit to check whether there are two double bonds attached to the carbon atoms in the ring at the same time. If so, it is determined to be an unstable molecule.

[0147] (3) Functional group detection: By traversing the types of molecular bonds, we can identify whether there are any unwanted functional groups, such as aldehyde groups (terminal structures of carbon-oxygen double bonds) and peroxides (OO single bonds). If they are detected, they are removed.

[0148] (4) Connectivity detection: The DFS (Depth-First Search) algorithm is used to traverse all bonds starting from a certain atom. If all atoms cannot be accessed, it means that the molecular structure is not connected and is deemed invalid.

[0149] After obtaining the initial matrix-generated molecular database, the SMILES molecules are converted into molecular graph representations. SMILES are parsed using RDKit, extracting atomic numbers, valence states, and electronegativity as node features, and chemical bond information (single, double, and aromatic bonds) as edge features. Graph objects are then constructed using torch_geometric.data.Data to form a standardized molecular graph database.

[0150] Subsequently, a generative adversarial network (model) is constructed, such as... Figure 2As shown, a Generative Adversarial Network (GAN) consists of a generator and a discriminator. The generator takes a random noise vector or a partial molecular fragment as input and progressively generates the adjacency matrix and node feature matrix of the molecule using multi-layer graph convolutions (GCNConv or GraphConv) and fully connected layers. During generation, symmetry constraints are imposed on the adjacency matrix to ensure it satisfies the symmetry properties of molecular bonds, and softmax normalization is applied to the node feature matrix to ensure that each node corresponds to a unique atom type. The discriminator takes the molecular graph as input, first extracting topological features through graph convolutional layers, then using global pooling (a combination of global mean pooling and max pooling) to obtain a graph-level vector, and finally outputting the discrimination result—the probability that the input molecule is a real molecule or a generated molecule—through a fully connected layer with spectral normalization.

[0151] During adversarial training, an alternating optimization strategy is employed: on one hand, the discriminator learns to maximize its ability to distinguish between real and generated molecules; on the other hand, the generator continuously adjusts its parameters to make the distribution of generated molecules as close as possible to the real molecule database. The loss function uses Wasserstein distance combined with gradient penalty (WGAN-GP) to improve training stability.

[0152] Once the model training converges, the process enters the molecular generation and decoding stage. By inputting a random noise vector z or a fragment of SMILES, the generator outputs a new candidate molecular map. The system decodes this molecular map: determining the atom types based on the node feature matrix, determining the bond types and connectivity based on the adjacency matrix, and reconstructing the complete molecular skeleton. Subsequently, the aforementioned screening function is called to perform validity checks, retaining only molecules with reasonable valence bonds, stable ring structures, and no unwanted functional groups.

[0153] For molecules that pass the screening, the molecular diagrams are converted into SMILES representations using RDKit, and then Chem.EmbedMolecule() is called to generate the three-dimensional structure. Simultaneously, hydrogen atoms are automatically added using Chem.AddHs() to ensure molecular saturation and structural integrity. Finally, all molecules that pass the screening are stored in the generated_smiles collection and saved in Excel or CSV format for subsequent property calculations and electrolyte screening.

[0154] The results of this embodiment demonstrate that this method not only generates a large number of new molecules that conform to chemical rules, but also significantly improves the diversity and structural novelty of the generated molecules due to the introduction of a GAN model. Compared with simple matrix enumeration methods, GANs can learn and reproduce the distribution characteristics of real molecules, making the generated molecules closer to practical application requirements in terms of structural complexity, functional group distribution, and electrolyte compatibility. Therefore, the molecule generation method based on matrix engineering and generative adversarial networks proposed in this embodiment has high practicality and innovation, providing an effective tool for the design and screening of novel electrolyte molecules.

[0155] Example 3

[0156] This embodiment provides an actual operation process for generating molecules using the molecular generation method based on matrix engineering and generative adversarial networks described in Embodiment 1, and focuses on comparing the impact of different discriminator structures on model performance.

[0157] First, import the necessary Python libraries, including RDKit (for molecular structure analysis and SMILES conversion), PyTorch (for building and training deep learning models), Pandas (for molecular database management), NumPy (for numerical computation), and time and os (for runtime recording and file storage). Use time.perf_counter() to record the total time spent on model training and molecule generation, and initialize a results storage folder to save the generated SMILES and their corresponding 3D molecular structures.

[0158] In this embodiment, the matrix method described in Example 1 is still used to generate the basic molecular database and convert it into a molecular graph structure. Nodes represent atoms and are accompanied by atomic features (atomic number, electronegativity, hybridization, etc.), and edges represent chemical bonds and are accompanied by bond features (single bond, double bond, aromatic bond, etc.). Finally, a molecular graph dataset for training is obtained. Figure 4 This is a schematic diagram of deriving new molecules from basic molecules.

[0159] The difference lies in the design of the Generative Adversarial Network (GAN), where this embodiment employs two different discriminator structures for comparison:

[0160] Discriminator A (standard structure in Example 1): Spectral Normalization is introduced into the fully connected layer to stabilize training and prevent the discriminator from becoming too strong and causing the generator gradient to vanish; at the same time, the LeakyReLU activation function is used to enhance the nonlinear expressive power.

[0161] Discriminator B (a version without SpectralNormalization): Maintains the same number of layers and parameter size, but removes spectral normalization and relies solely on BatchNorm and LeakyReLU for training stability.

[0162] Both take the molecular adjacency matrix and the vector resulting from the expansion of node features as input, and output a score after passing through multiple fully connected layers, representing the probability that the input molecule belongs to a real molecule or a generated molecule.

[0163] During training, the generator continues to receive random noise as input to z, mapping layer by layer to obtain the adjacency matrix and node feature matrix of the molecules, and obtaining candidate molecules through symmetry constraints and softmax normalization. The discriminator learns the molecular distribution features through binary classification game, using WGAN-GP as the loss function to ensure training stability.

[0164] In the experimental comparison, when using discriminator A (with SpectralNormalization), the model converged stably after 20 iterations, and the training process was relatively smooth, with the losses of both the discriminator and generator remaining within a reasonable range. However, when using discriminator B (without SpectralNormalization), the model exhibited instability during training. The generator loss fluctuated significantly in the early stages of iteration, and the discriminator was prone to overfitting, performing far worse than discriminator A. Furthermore, in terms of training time, discriminator B required more iterations to partially converge due to frequent oscillations, resulting in lower overall efficiency.

[0165] The comparative results show that introducing SpectralNormalization not only improves the stability of training but also significantly enhances the efficiency and diversity of molecule generation. Therefore, in molecule generation tasks based on matrix engineering and GANs, the choice of discriminator structure has a significant impact on the final performance.

[0166] In summary, this embodiment verifies that in the discriminator design of generative adversarial networks, reasonable regularization methods (such as SpectralNormalization) can effectively improve the quality and efficiency of molecule generation, providing more reliable data support for subsequent electrolyte molecule screening.

[0167] Example 4

[0168] This embodiment provides a comparative experiment based on matrix engineering and generative adversarial networks (GANs) to illustrate the differences between the two molecular generation strategies in terms of efficiency, molecular diversity, and application value.

[0169] In matrix engineering, a matrix is ​​first constructed where the first row and first column restrict the atom types, and the remaining elements represent the bonding relationships between atoms (0 for no bonding, 1 for a single bond, and 2 for a double bond). Connectivity and validity checks are performed using a backtracking traversal algorithm and a depth-first search (DFS) algorithm to filter out molecules with valence bonds exceeding the limits or containing unreasonable functional groups (such as peroxides and bridgehead carbons). Finally, all matrices meeting the conditions can be converted into unique SMILES representations, and 3D structures are generated using RDKit. The advantages of this method are strong controllability of the molecule generation process, and all constraints can be explicitly encoded to ensure that the generated molecules satisfy chemical rules. The disadvantages are high search complexity; when the number of atoms is large, the traversal space grows exponentially, severely limiting the generation efficiency, and the diversity of generated molecules is limited.

[0170] In the Generative Adversarial Network (GAN) approach, molecules generated by matrix engineering are first used as a high-quality database and transformed into molecular graph representations: nodes represent atoms with features such as atom type and electronegativity; edges represent chemical bonds with bond types. When constructing the GAN model, the generator takes a random noise vector as input and outputs an adjacency matrix and node feature matrix. Candidate molecular graphs are obtained through symmetry and softmax constraints. The discriminator maps the real and generated molecular graphs to a feature space and calculates their true / false probabilities using spectral normalization and a nonlinear activation function. During adversarial training, the generator continuously learns the distribution patterns of real molecules, while the discriminator improves its recognition capabilities. This game-like optimization allows the generator to efficiently generate new, plausible molecules. The advantages of this method are its fast molecular generation speed, producing a large number of new molecules in a short time, while also exhibiting higher diversity and structural novelty. The disadvantage is the limited explicit chemical constraints in the generation process, requiring a discriminator and post-processing to ensure chemical plausibility.

[0171] The comparison results show that:

[0172] Generation efficiency: Matrix methods rely on traversal and rule selection, which is inefficient, while GAN methods are based on learned distributions and generate data much faster.

[0173] Molecular diversity: Matrix methods are prone to generating regular molecules with a single pattern, while GAN methods can learn implicit distributions, generating molecules with greater diversity in functional group combinations and topological structures.

[0174] Constraint controllability: Matrix methods can explicitly control constraints such as the number of atoms and bond types, making them more suitable for generating specific target systems; although GAN methods are difficult to directly control the number of atoms and structure, combining matrix-generated databases as training data can enhance the exploration space of generated molecules while ensuring basic chemical rationality.

[0175] Based on the same inventive concept, this application also provides a generative adversarial network-based molecular generation apparatus for implementing the generative adversarial network-based molecular generation method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the generative adversarial network-based molecular generation apparatus provided below can be found in the limitations of the generative adversarial network-based molecular generation method described above, and will not be repeated here.

[0176] In one exemplary embodiment, such as Figure 5 As shown, a molecular generation device based on a generative adversarial network is provided, comprising:

[0177] The molecular database construction module is used to construct a molecular database based on the required atom types and numbers for the target molecule using matrix engineering methods.

[0178] The transformation module is used to transform the molecular structures corresponding to the candidate molecule sets in the molecular database into graph structures, thus obtaining a molecular graph database.

[0179] The molecule generation and decoding module is used to input the molecular graph database into the trained generative adversarial network (GAN) to generate and decode molecules, resulting in newly generated molecules. The GAN is trained using the Adam optimizer, which optimizes the objective function through adversarial game theory and incorporates a gradient penalty term. The GAN consists of a generator and a discriminator.

[0180] The molecular screening module is used to perform molecular screening on newly generated molecules based on preset screening rules to obtain qualified molecules.

[0181] The three-dimensional structure conversion module is used to convert qualified molecules into three-dimensional structures, resulting in three-dimensional molecules. These three-dimensional molecules are then used for computational simulations and experimental verification, providing a reference for electrolyte design and optimization.

[0182] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores molecular generation data based on generative adversarial networks (GANs). The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a molecular generation method based on generative adversarial networks.

[0183] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0184] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0185] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0186] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0187] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0188] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0189] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0190] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0191] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for molecule generation based on a generative adversarial network, characterized in that, The application relates to a method for generating a new molecule, and belongs to the technical field of molecular design. The method comprises the following steps: a matrix engineering method is used to construct a molecular database according to required atom types and quantities of a target molecule; molecular structures corresponding to a candidate molecule set in the molecular database are converted into graph structures to obtain a molecular graph database; the molecular graph database is input into a trained generative adversarial network to perform molecular generation and decoding, and a newly generated molecule is obtained; the generative adversarial network is obtained by optimizing an objective function through an adversarial game mode and adding a gradient penalty term after training; the generative adversarial network comprises a generator and a discriminator; a preset screening rule is used to perform molecular screening on the newly generated molecule to obtain a screened molecule; the screened molecule is converted into a three-dimensional structure to obtain a three-dimensional molecule; 2. The generative adversarial network-based molecule generation method according to claim 1, characterized in that, the three-dimensional molecule is used for performing calculation simulation and experimental verification to provide a reference basis for electrolyte design and optimization. The method comprises the following steps: a matrix engineering method is used to generate a molecule set according to required atom types and quantities of a target molecule; the molecule set is represented in a matrix form; the first row and the first column of the matrix are used to represent atom types, and the remaining elements in the matrix are used to represent bonding relationships; wherein 0 represents no bond, 1 represents a single bond, and 2 represents a double bond; a backtracking algorithm is used to traverse elements in an upper triangular region of the matrix, and a SMILES molecular structure conforming to chemical rules is generated by combining a valence bond constraint rule; connectivity and functional group screening are performed on the SMILES molecular structure to determine a candidate molecule set; 3. The generative adversarial network-based molecule generation method according to claim 2, characterized in that, a molecular database is determined based on the candidate molecule set. The method comprises the following steps:

4. The generative adversarial network-based molecule generation method according to claim 1, wherein, the SMILES molecular structure is converted into a graph structure; wherein atoms in the SMILES molecule are used as nodes of the graph structure, chemical bonds are used as edges, node features are used to represent atom types, electronegativity and hybridization states, and edge features are used to represent bond types and bond orders. The method comprises the following steps:

5. The generative adversarial network-based molecule generation method according to claim 1, wherein, an EmbedMolecule command of an RDKit library is used to convert the screened molecule into a three-dimensional structure, and a Chem.AddHs() method is used to automatically supplement hydrogen atoms to obtain a three-dimensional molecule. The objective function comprises a target function of the generator and a target function of the discriminator; L D = -E x～pdata [logD(x)] - E z～pz [log(l - D(G(z)))]; the target function of the discriminator has the following mathematical expression: L G = -E z～pz [logD(G(z))]; where L D is the objective function of the discriminator; E x～pdata is the expectation over the real data distribution pdata(x); D(x) is the output of the discriminator for a real molecular sample x; E z～pz is the expectation over the noise distribution pz(z); G(z) is the generator function that takes a noise vector z as input and outputs a generated molecular data; x is a real molecular sample; z is a noise vector; L G is the objective function of the generator.

6. The generative adversarial network-based molecule generation method according to claim 1, wherein, the target function of the generator has the following mathematical expression: where L1is a gradient penalty term; E x～pg is the expectation over the generated data distribution pg(x); D(x) is the output of the discriminator for a real molecular sample x; E x～pr is the expectation over the real data distribution pr(x); λ is a gradient penalty coefficient; E x～px is the expectation over the interpolated distribution px(x); is the gradient of the discriminator with respect to the input real molecular sample x; ||·||2is the L2norm.

7. A molecular generation device based on a generative adversarial network, characterized by, the gradient penalty term has the following mathematical expression: The application further provides a method for generating a new molecule. The method comprises the following steps: a molecular database construction module is used to adopt a matrix engineering method to construct a molecular database according to required atom types and quantities of a target molecule; a conversion module is used to convert molecular structures corresponding to a candidate molecule set in the molecular database into graph structures to obtain a molecular graph database; The molecule generation and decoding module is configured to input the molecule graph database into a trained generative adversarial network to perform molecule generation and decoding, and obtain a newly generated molecule; the generative adversarial network is obtained by training the generative adversarial network based on an objective function through an adversarial game manner and using an Adam optimizer and adding a gradient penalty term; the generative adversarial network includes a generator and a discriminator; The molecule screening module is configured to perform molecule screening on the newly generated molecule based on a preset screening rule, and obtain a molecule that passes the screening; The three-dimensional structure conversion module is configured to convert the molecule that passes the screening into a three-dimensional structure, and obtain a three-dimensional molecule; the three-dimensional molecule is used for performing computational simulation and experimental verification, and providing a reference basis for electrolyte design and optimization.

8. A computer device comprising: 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 molecule generation method based on the generative adversarial network in any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the molecule generation method based on the generative adversarial network in any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the molecule generation method based on the generative adversarial network in any one of claims 1-6.

Images