Method, system and equipment for generating molecular structural formula of polymer electrolyte and medium

Abstract

The invention discloses a molecular structural formula generation method, system and equipment of a polymer electrolyte and a medium, and belongs to the technical field of lithium battery technologies. The molecular structural formula generation method of the polymer electrolyte comprises the following steps: acquiring a performance data set and a molecular structural formula data set of the polymer electrolyte; constructing a discriminant model by using the performance data set; constructing a generative model by using the molecular structural formula data set; training the generative model based on the discriminant model; and if a molecular structure design task is received, generating a plurality of new molecular structural formulas of the polymer electrolyte by using the trained generative model, and screening the new molecular structural formulas of the polymer electrolyte by using the discriminant model to obtain a molecular structural formula conforming to the molecular structure design task. The molecular structural formula of the polymer electrolyte can be automatically generated, and the generation efficiency and effectiveness of the molecular structural formula are improved.

Description

Technical Field

[0001] The present application relates to the technical field of lithium batteries, and in particular to a method, system, device and medium for generating a molecular structure of a polymer electrolyte. Background Art

[0002] Lithium metal batteries have been widely studied as a battery with a very high theoretical energy density. However, due to the uncontrollable growth of lithium dendrites, the cycle performance and safety of the battery are reduced. Replacing liquid electrolytes with polymer solid electrolytes is expected to solve this problem. However, polymer electrolytes still face problems such as low conductivity and low lithium ion mobility, which hinder their application in lithium metal batteries. Therefore, it is urgent to develop new high-performance polymer electrolytes to improve the performance of lithium metal batteries.

[0003] In the related technologies, the design method of polymer electrolyte mainly relies on laboratory experiments and experience accumulation, which is time-consuming and labor-intensive, resulting in a long R&D cycle and high costs. In addition, due to the complex molecular structure of polymer electrolytes, their performance is affected by many factors, and a single experimental method is difficult to fully characterize and optimize.

[0004] Therefore, how to automatically generate the molecular structure formula of the polymer electrolyte and improve the efficiency and effectiveness of the generation of the molecular structure formula are technical problems that those skilled in the art currently need to solve. Summary of the invention

[0005] The purpose of the present application is to provide a method, system, device and medium for generating a molecular structural formula of a polymer electrolyte, which can automatically generate a molecular structural formula of a polymer electrolyte and improve the efficiency and effectiveness of generating the molecular structural formula.

[0006] In order to solve the above technical problems, the present application provides a method for generating a molecular structural formula of a polymer electrolyte, comprising:

[0007] Obtain performance data sets and molecular structure data sets of polymer electrolytes;

[0008] Constructing a discriminant model using the performance data set; wherein the discriminant model is used to calculate performance indicators of the molecular structure formula, the performance indicators including any one or a combination of any several of conductivity, lithium ion migration number, glass transition temperature and thermal decomposition temperature;

[0009] Constructing a generative model using the molecular structural formula data set; wherein the generative model is used to generate a molecular structural formula of a polymer electrolyte;

[0010] Training the generative model based on the discriminative model;

[0011] If a molecular structure design task is received, the trained generative model is used to generate a plurality of new molecular structural formulas of polymer electrolytes, and the discriminant model is used to screen the molecular structural formulas of the new polymer electrolytes to obtain molecular structural formulas that meet the molecular structure design task.

[0012] Optionally, after obtaining the performance data set and molecular structure data set of the polymer electrolyte, the following is further included:

[0013] Data cleaning and data visualization operations are performed on the performance data set and the molecular structure formula data set.

[0014] Optionally, constructing a discriminant model using the performance data set includes:

[0015] Extracting data features from the performance data set, and standardizing and normalizing the data features;

[0016] Filtering the data features to obtain target data features;

[0017] Machine learning modeling is performed based on the target data features and corresponding labels in the performance data set to obtain the discriminant model.

[0018] Optionally, the performance data set includes any one or a combination of any several of a conductivity data set, a lithium ion migration number data set, a glass transition temperature data set and a thermal decomposition temperature data set; the conductivity data set includes polymer structure, polymer molecular weight, lithium salt structure, lithium salt concentration, temperature and conductivity label; the lithium ion migration number data set includes polymer structure, lithium salt structure, lithium salt concentration, temperature, additive type, additive content and lithium ion migration number label; the glass transition temperature data set includes polymer structure and glass transition temperature label; the thermal decomposition temperature data set includes polymer structure and thermal decomposition temperature label;

[0019] Accordingly, machine learning modeling is performed based on the target data features and corresponding labels in the performance data set to obtain the discriminant model, including:

[0020] Performing machine learning modeling based on target data features and conductivity labels in the conductivity data set to obtain a conductivity discriminant model;

[0021] And / or, performing machine learning modeling based on the target data features and lithium ion migration number labels in the lithium ion migration number data set to obtain a lithium ion migration number discriminant model;

[0022] And / or, performing machine learning modeling based on target data features and glass transition temperature labels in the glass transition temperature data set to obtain a glass transition temperature discriminant model;

[0023] And / or, machine learning modeling is performed based on the target data features and thermal decomposition temperature labels in the thermal decomposition temperature data set to obtain a thermal decomposition temperature discriminant model.

[0024] Optionally, constructing a generative model using the molecular structure formula dataset includes:

[0025] Converting the molecular structures in the molecular structure formula data set into a graph structure; wherein the nodes in the graph structure represent atoms and the edges represent chemical bonds;

[0026] A corresponding graph grammar is generated according to the graph structure, and the generative model is constructed based on the graph grammar.

[0027] Optionally, training the generative model based on the discriminant model includes:

[0028] Constructing an objective function of the generative model; wherein the value of the objective function is a weighted calculation result of multiple evaluation indicators, and the evaluation indicators include scores corresponding to performance indicators calculated by the discriminant model;

[0029] The generative model is iteratively trained based on the objective function.

[0030] Optionally, iteratively training the generative model based on the objective function includes:

[0031] Controlling the generative model to generate an i-th round of output results;

[0032] Calculate the value of the objective function corresponding to the output result of the i-th round;

[0033] Determine whether the value of the objective function calculated most recently is within a preset range;

[0034] If yes, it is determined that the generative model training is complete;

[0035] If not, adjust the parameters of the generative model, add 1 to the value of i and enter the operation of controlling the generative model to generate the i-th round output result;

[0036] Among them, the evaluation indicators also include effectiveness index, uniqueness index, novelty index, diversity index, retrosynthesis score index and membership index; the effectiveness index is used to describe the proportion of valid molecular structures in the output results of the generative model; the uniqueness index is used to describe the proportion of non-repeated molecular structures in the output results of the generative model; the novelty index is used to describe the proportion of molecular structures in the output results of the generative model that are not in the molecular structure formula data set; the diversity index is used to describe the structural similarity between the output results of the generative model; the retrosynthesis score index is used to describe the synthesis success rate of the output results of the generative model; the membership index is used to describe the proportion of molecular structures in the output results of the generative model that belong to the molecular structure formula data set.

[0037] The present application also provides a system for generating a molecular structure formula of a polymer electrolyte, the system comprising:

[0038] A data set acquisition module, used to acquire performance data sets and molecular structure data sets of polymer electrolytes;

[0039] A first model building module is used to build a discriminant model using the performance data set; wherein the discriminant model is used to calculate the performance index of the molecular structure formula, and the performance index includes any one or a combination of any several of the following: conductivity, lithium ion migration number, glass transition temperature and thermal decomposition temperature;

[0040] A second model building module is used to build a generative model using the molecular structure formula data set; wherein the generative model is used to generate a molecular structure formula of a polymer electrolyte;

[0041] A model training module, used for training the generative model based on the discriminant model;

[0042] The molecular generation module is used to generate a plurality of new molecular structural formulas of polymer electrolytes using the trained generative model if a molecular structure design task is received, and to screen the molecular structural formulas of the new polymer electrolytes using the discriminant model to obtain molecular structural formulas that meet the molecular structure design task.

[0043] The present application also provides a storage medium on which a computer program is stored. When the computer program is executed, the steps of the method for generating the molecular structure formula of the polymer electrolyte are implemented.

[0044] The present application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor calls the computer program in the memory, the steps of the method for generating the molecular structure formula of the polymer electrolyte are implemented.

[0045] The present application constructs a discriminant model through a performance data set of a polymer electrolyte. The discriminant model can learn the relationship between different molecular structures and performances. The discriminant model can be used to evaluate whether the newly generated molecular structure meets specific performance requirements. The present application constructs a generative model through a molecular structural formula data set of a polymer electrolyte. The generative model can automatically generate a new molecular structural formula of a polymer electrolyte based on a known molecular structure. The present application trains the generative model based on the discriminant model so that the generative model can generate a molecular structural formula whose performance indicators meet the requirements. After receiving a molecular structure design task, the present application can use the generative model to generate a plurality of new molecular structural formulas of polymer electrolytes, and screen them through the discriminant model to obtain a molecular structural formula that meets the molecular structure design task. The above process uses a discriminant model to guide the generative model to generate a molecular structural formula without manual participation. Therefore, the present application can automatically generate a molecular structural formula of a polymer electrolyte, thereby improving the efficiency and effectiveness of generating the molecular structural formula. The present application also provides a molecular structural formula generation system, a storage medium and an electronic device for a polymer electrolyte, which have the above beneficial effects and are not repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 A flow chart of a method for generating a molecular structure formula of a polymer electrolyte provided in an embodiment of the present application;

[0048] Figure 2 A schematic diagram of a method for representing a molecular graph provided in an embodiment of the present application;

[0049] Figure 3 A schematic diagram of a graph grammar generation rule and a molecular generation process using the grammar provided in an embodiment of the present application;

[0050] Figure 4 A probability density distribution diagram of the conductivity after logarithmization provided in an embodiment of the present application;

[0051] Figure 5 A schematic diagram of a feature correlation matrix after conductivity data feature screening provided in an embodiment of the present application;

[0052] Figure 6A schematic diagram of a high conductivity molecule generated by a trained graph grammar generation model provided in an embodiment of the present application;

[0053] Figure 7 A probability density distribution diagram of glass transition temperature data provided in an embodiment of the present application;

[0054] Figure 8 A probability density distribution diagram of thermal decomposition temperature data provided in an embodiment of the present application;

[0055] Fig. 9 A schematic diagram of a correlation matrix of characteristics of glass transition temperature data provided in an embodiment of the present application;

[0056] Fig.10 A schematic diagram of a correlation matrix of characteristics of thermal decomposition temperature data provided in an embodiment of the present application;

[0057] Fig.11 A schematic diagram of a molecule with excellent thermal properties provided in an embodiment of the present application;

[0058] Fig.12 A probability density distribution diagram of lithium ion migration number data provided in an embodiment of the present application;

[0059] Fig.13 A schematic diagram of a correlation matrix of characteristics of lithium ion migration number data provided in an embodiment of the present application;

[0060] Fig.14 A schematic diagram of a molecule with excellent comprehensive performance generated according to an embodiment of the present application;

[0061] Fig.15 A flow chart of a method for designing a lithium battery polymer electrolyte by coupling a discriminant and a generative model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0063] See below Figure 1 , Figure 1 A flow chart of a method for generating a molecular structure formula of a polymer electrolyte provided in an embodiment of the present application.

[0064] Specific steps may include:

[0065] S101: Obtain performance data set and molecular structure data set of polymer electrolyte;

[0066] Among them, this embodiment can be applied to electronic devices with data processing and modeling functions, and realizes the automatic design of lithium battery polymer electrolytes by establishing discriminant models and generative models. The performance data set of polymer electrolytes includes relevant data of performance indicators such as conductivity, lithium ion migration number, glass transition temperature and thermal decomposition temperature; the molecular structure data set includes molecular structure formulas of multiple polymer electrolytes.

[0067] S102: constructing a discriminant model using the performance data set;

[0068] The discriminant model is used to calculate the performance index of the molecular structure, and the performance index includes any one or a combination of any several of the following: conductivity, lithium ion migration number, glass transition temperature, and thermal decomposition temperature. Each performance index can have its corresponding performance data set, and then a discriminant model corresponding to each performance index can be constructed, and the above discriminant model can be a supervised learning model.

[0069] Specifically, this embodiment can split the performance data set into a training set and a test set, and use the data in the training set to train the discriminant model so that the discriminant model learns the relationship between the molecular structure and the performance index. This embodiment can also use a test machine to verify the trained discriminant model to ensure the predictive ability of the discriminant model for new data.

[0070] S103: constructing a generative model using the molecular structure formula dataset;

[0071] Among them, this step uses the molecular structure formula data set to build a generative model. The generative model can learn the patterns and features in the molecular structure formula data set, and then generate a new molecular structure formula of the polymer electrolyte.

[0072] S104: Training the generative model based on the discriminant model;

[0073] Among them, based on the constructed discriminant model and generative model, this embodiment can use the discriminant model as a guide to evaluate the quality of the molecular structure created by the generative model; through the feedback mechanism, the discriminant model provides performance prediction and optimization direction, so that the generative model can learn and generate new molecular structures that better meet the expected performance indicators.

[0074] S105: If a molecular structure design task is received, the trained generative model is used to generate a plurality of new molecular structural formulas of polymer electrolytes, and the discriminant model is used to screen the molecular structural formulas of the new polymer electrolytes to obtain molecular structural formulas that meet the molecular structure design task.

[0075] After receiving the molecular structure design task, the trained generative model can be used to quickly generate multiple new molecular structures of polymer electrolytes, and then the discriminant model can be used to evaluate and screen these newly generated molecular structures to ensure that the selected molecular structures meet the predetermined performance requirements and design goals, and finally obtain molecular structures that meet the task requirements. The above molecular structure design tasks can include performance requirements, quantity requirements, etc.

[0076] This embodiment constructs a discriminant model through a performance data set of a polymer electrolyte. The discriminant model can learn the relationship between different molecular structures and performance. The discriminant model can be used to evaluate whether the newly generated molecular structure meets specific performance requirements. This embodiment constructs a generative model through a molecular structural formula data set of a polymer electrolyte. The generative model can automatically generate a new molecular structural formula of a polymer electrolyte based on a known molecular structure. This embodiment trains the generative model based on the discriminant model so that the generative model can generate a molecular structural formula whose performance indicators meet the requirements. After receiving a molecular structure design task, this embodiment can use the generative model to generate a plurality of new molecular structural formulas of polymer electrolytes, and screen them through the discriminant model to obtain a molecular structural formula that meets the molecular structure design task. The above process uses a discriminant model to guide the generative model to generate a molecular structural formula without manual participation. Therefore, this embodiment can automatically generate the molecular structural formula of the polymer electrolyte, thereby improving the efficiency and effectiveness of generating the molecular structural formula.

[0077] As for Figure 1 As further described in the corresponding embodiment, after obtaining the performance data set and the molecular structure data set of the polymer electrolyte, data cleaning and data visualization operations may be performed on the performance data set and the molecular structure data set.

[0078] The above-mentioned data cleaning may include operations such as unit conversion, removal of duplicate data, filling in missing values ​​and correcting outliers; the above-mentioned data visualization operation may use data visualization methods to generate distribution diagrams and feature association diagrams of various performance indicators.

[0079] As for Figure 1For further introduction to the corresponding embodiment, a discriminant model can be constructed in the following manner: extracting data features from the performance data set, and standardizing and normalizing the data features; filtering the data features to obtain target data features; and performing machine learning modeling based on the target data features and corresponding labels in the performance data set to obtain the discriminant model.

[0080] Furthermore, the above-mentioned performance data set may include any one or a combination of any several of the conductivity data set, lithium ion migration number data set, glass transition temperature data set and thermal decomposition temperature data set. The conductivity data set includes polymer structure, polymer molecular weight, lithium salt structure, lithium salt concentration, temperature (i.e., the temperature corresponding to the conductivity label) and conductivity label; the lithium ion migration number data set includes polymer structure, lithium salt structure, lithium salt concentration, temperature (i.e., the temperature corresponding to the lithium ion migration number label) and lithium ion migration number label; the glass transition temperature data set includes polymer structure and glass transition temperature label; the thermal decomposition temperature data set includes polymer structure and thermal decomposition temperature label. The conductivity label is the conductivity corresponding to the set of data of polymer structure, polymer molecular weight, lithium salt structure, lithium salt concentration and temperature; the lithium ion migration number label is the lithium ion migration number corresponding to the set of data of polymer structure, lithium salt structure, lithium salt concentration and temperature; the glass transition temperature label is the glass transition temperature corresponding to the polymer structure; and the thermal decomposition temperature label is the thermal decomposition temperature corresponding to the polymer structure.

[0081] If the performance data set includes a conductivity data set, machine learning modeling of the discriminant model may be performed in the following manner: machine learning modeling is performed based on target data features and conductivity labels in the conductivity data set to obtain a conductivity discriminant model.

[0082] If the performance data set includes a lithium ion migration number data set, machine learning modeling of the discriminant model can be performed in the following manner: machine learning modeling is performed based on the target data features and lithium ion migration number labels in the lithium ion migration number data set to obtain a lithium ion migration number discriminant model.

[0083] If the performance data set includes a glass transition temperature data set, machine learning modeling of the discriminant model can be performed in the following manner: machine learning modeling is performed based on the target data features and glass transition temperature labels in the glass transition temperature data set to obtain a glass transition temperature discriminant model.

[0084] If the performance data set includes a thermal decomposition temperature data set, machine learning modeling of the discriminant model can be performed in the following manner: machine learning modeling is performed based on the target data features and thermal decomposition temperature labels in the thermal decomposition temperature data set to obtain a thermal decomposition temperature discriminant model.

[0085] As for Figure 1 Further introduction to the corresponding embodiment, the process of constructing a generative model using the molecular structure formula data set includes: converting the molecular structure in the molecular structure formula data set into a graph structure; wherein the nodes in the graph structure represent atoms and the edges represent chemical bonds; generating a corresponding graph grammar according to the graph structure, and constructing the generative model based on the graph grammar.

[0086] This embodiment generates a corresponding graph grammar based on the converted graph structure, and defines how to construct a new molecular structure through rule operations (such as adding, deleting or modifying nodes and edges). Based on the above graph grammar, a generative model is constructed, which can learn and follow these rules to automatically generate a molecular structure of a new polymer electrolyte that conforms to chemical logic.

[0087] On the premise of constructing a generative model based on graph grammar, the process of training the generative model using the discriminant model includes: constructing an objective function of the generative model; wherein the value of the objective function is a weighted calculation result of multiple evaluation indicators, and the evaluation indicators include scores corresponding to performance indicators calculated by the discriminant model; and iteratively training the generative model based on the objective function.

[0088] The above-mentioned process of iteratively training the generative model based on the objective function includes: controlling the generative model to generate the i-th round output result; calculating the value of the objective function corresponding to the i-th round output result; judging whether the value of the objective function calculated most recently is within a preset range; if so, determining that the generative model training is completed; if not, adjusting the parameters of the generative model, adding 1 to the value of i and entering the operation of controlling the generative model to generate the i-th round output result.

[0089] In addition to the scores corresponding to the performance indicators calculated by the discriminant model, the above evaluation indicators may also include: effectiveness indicators, uniqueness indicators, novelty indicators, diversity indicators, retrosynthesis score indicators and membership indicators.

[0090] The validity index is used to describe the proportion of valid molecular structures in the output of the generative model; the above-mentioned valid molecular structure is a molecular structure that conforms to chemical rules. The output of the generative model can be any number of newly generated structural formulas.

[0091] The uniqueness index is used to describe the proportion of non-repeated molecular structures in the output results of the generative model, that is, the proportion of unique molecules in the output results.

[0092] The novelty index is used to describe the proportion of molecular structures in the output results of the generative model that are not in the molecular structure formula dataset, that is, the proportion of molecular structure formulas in the output results that are not in the molecular structure formula dataset.

[0093] The diversity index is used to describe the structural similarity between the output results of the generative model, that is, the average pairwise molecular distance between molecular structural formulas.

[0094] The retrosynthesis score indicator is used to describe the synthesis success rate of the output results of the generative model.

[0095] The membership index is used to describe the proportion of molecular structures in the output results of the generative model that belong to the molecular structure formula data set, that is, the proportion of molecular structure formulas in the output results that belong to the molecular structure formula data set.

[0096] The process described in the above embodiment is explained below through an embodiment in actual application.

[0097] This embodiment provides a scheme for designing polymer electrolytes for lithium metal batteries by combining a discriminant model with a generative model of inverse design. In the case of limited data, this scheme can predict the performance of polymer electrolytes in the early stages of material design and generate polymers with excellent performance. This scheme can not only accelerate the development process of polymer electrolytes, but also provide effective guidance for the design of new polymers, reduce experimental costs, and improve research and development efficiency. In addition, based on the guiding role of the discriminant model, it is possible to optimize the polymer synthesis route, improve the overall performance of the electrolyte, and promote the development and application of high-performance lithium metal batteries.

[0098] The above-mentioned method of using the coupled discriminant model and the generative model of inverse design to design a polymer solid electrolyte includes the following steps A1-A4:

[0099] Step A1: Collect and organize the performance data set and molecular structure data set of polymer electrolytes.

[0100] The data sets used in this embodiment are divided into two categories: one is the performance data set of polymer electrolytes, which is used to train and optimize the discriminant model; the other is the molecular structure data set containing polymer monomers for lithium metal batteries, which is used to train the generative model.

[0101] The performance dataset of polymer electrolytes is used to construct a discriminant model to predict the key properties of polymer electrolytes in lithium metal batteries. The key properties of polymer electrolytes include conductivity (σ, S cm -1 ), lithium ion migration number (t Li+ ), glass transition temperature (T g , ℃) and thermal decomposition temperature (Td , ℃), these properties have an important impact on the overall performance of lithium metal batteries. The following is an introduction to each property:

[0102] Conductivity is a parameter used to describe the ease with which electric charges flow in a substance, usually expressed in Siemens per centimeter (S cm -1 ) as a unit. In batteries, conductivity is the most important performance indicator of polymer electrolytes, because it directly affects the ion transfer efficiency inside the battery, thus affecting the overall performance of the battery. The calculation formula is usually based on an expanded form of Ohm's law, taking into account the resistance and geometry of the electrolyte, as follows:

[0103] ;

[0104] Where, d is the thickness of the polymer electrolyte (cm), R is the impedance of the polymer electrolyte (Ω), and S is the area of ​​the polymer electrolyte (cm -2 ).

[0105] In terms of data set collection, the conductivity in polymer electrolytes is affected by both polymers and lithium salts. Therefore, the data in the performance data set corresponding to conductivity (i.e., conductivity data set) includes polymer structure, polymer molecular weight, lithium salt structure, lithium salt concentration, and temperature corresponding to conductivity, ensuring that the effects of different factors on conductivity are fully considered. Specifically, the performance data set corresponding to conductivity includes multiple data in the following format: {polymer structure, polymer molecular weight, lithium salt structure, lithium salt concentration, test temperature, label (i.e., conductivity)}.

[0106] Lithium ion migration number (t Li+ ) refers to the ability of lithium ions to move in the electrolyte and is an important parameter for measuring electrolyte performance. A high lithium ion migration number means that lithium ions can move more efficiently during battery charging and discharging, thereby improving battery performance. For the lithium ion migration number of polymer electrolytes, it is usually measured at 60 ° C using the Bruce-Vincent-Evans method. A symmetrical Li|Li battery is subjected to a 30 mV voltage for 90 minutes (chronoamperometric measurement) and the impedance spectrum of the battery before and after polarization is collected. The amplitude of the alternating voltage bias measured by electrochemical impedance spectroscopy (EIS) is 10 mV, and the research frequency range is 500 kHz to 100 mHz. The specific calculation formula is as follows:

[0107] ;

[0108] Where ΔV is the chronoamperometric voltage (i.e. 30 mV), i0 is the initial current during polarization, and iss is the final current at polarization, R0 and R ss are the interface resistances of the electrode and electrolyte before and after polarization measured by EIS, respectively.

[0109] In terms of data collection, the data in the performance data set corresponding to the lithium ion migration number (i.e., the lithium ion migration number data set) includes polymer structure, lithium salt structure and concentration, test temperature corresponding to the lithium ion migration number, and the type and content of additives in the polymer electrolyte (such as fillers and plasticizers) to ensure that the influence of additives on it can be reflected. Specifically, the performance data set corresponding to the lithium ion migration number includes multiple data in the following format: {polymer structure, lithium salt structure, lithium salt concentration, test temperature, additive type, additive content, label (i.e., lithium ion migration number)}.

[0110] The thermal properties of polymer electrolytes also have a certain impact on the performance of batteries, among which the most important is the glass transition temperature T g and thermal decomposition temperature T d .

[0111] The glass transition temperature refers to the temperature at which a polymer changes from a glassy state to a highly elastic state, and is generally measured by a differential scanning calorimeter (DSC). Its importance lies in that since the migration of lithium ions in polymer electrolytes mainly depends on the coordination between lithium ions and polymers and the movement of polymer chain segments, a lower glass transition temperature can enhance the flexibility of polymer chains and promote the migration of lithium ions, thereby improving the conductivity of the electrolyte. However, on the other hand, a lower glass transition temperature may also reduce the mechanical strength and stability of the polymer. Therefore, a suitable glass transition temperature is very important for polymer electrolytes.

[0112] Thermal decomposition refers to the process of decomposing a compound by heating. Polymers will also undergo thermal decomposition under the action of heat. The thermal decomposition temperature of polymer electrolytes generally refers to the temperature at which the mass of the material decreases by 5% when it is thermally decomposed, also known as T d5 The thermal decomposition temperature is usually measured by thermogravimetric analysis (TGA) and is a key parameter for measuring the thermal stability of polymer electrolytes.

[0113] In terms of data collection, since the thermal properties of polymer electrolytes are mainly related to the structure of the polymer, the influence of other components can be approximately ignored. Therefore, the features in the collected thermal performance data set only include the features of the polymer. In the data set, one polymer molecular structure corresponds to one T g or T d .

[0114] The performance data set corresponding to the glass transition temperature (i.e., the glass transition temperature data set) includes the polymer structure and the corresponding label (i.e., the glass transition temperature), and the performance data set corresponding to the thermal decomposition temperature (i.e., the thermal decomposition temperature data set) includes the polymer structure and the corresponding label (i.e., the thermal decomposition temperature).

[0115] The molecular structural formulas of polymer monomers used in lithium metal battery polymer electrolytes are collected as training data for the generative model (i.e., molecular structural formula dataset) to build a molecular generative model for generating new polymers.

[0116] To ensure the quality and consistency of the data set, this embodiment can systematically clean and visualize the collected performance data set and molecular structure data set to facilitate subsequent discriminant model performance prediction and generative model training and optimization.

[0117] This embodiment performs systematic data cleaning on the collected performance data set and molecular structure data set to ensure the uniformity and accuracy of the data. The cleaning steps include unit conversion, removal of duplicate data, filling in missing values ​​and correcting outliers.

[0118] The following is an explanation of the unit conversion: Since there are many units for lithium salt concentration used in the original data, such as mass fraction (denoted by ω, in wt%), amount of substance fraction (denoted by x, in mol%, where the amount of substance of the polymer is the amount of substance of the repeating unit), mass molar concentration (denoted by m, defined as the amount of lithium salt contained per unit mass of the polymer, in mol kg -1 ), the concentration of the substance (denoted by c, the unit is mol L -1 ), and before machine learning modeling, it is necessary to unify the data units. In this embodiment, the lithium salt concentration unit can be unified as mass molar concentration.

[0119] The lithium salt mass fraction ω s The formula for obtaining the mass molar concentration m is as follows:

[0120] ;

[0121] Among them, M s is the molar mass of the lithium salt in g mol -1 .

[0122] The formula for obtaining the mass molar concentration m from the amount fraction x of the lithium salt is as follows:

[0123] ;

[0124] Among them, M p is the molar mass of the polymer repeating unit, in g mol -1.

[0125] For in-situ polymerized polymer electrolytes, the polymer is used as a solvent for the lithium salt in the form of a monomer before preparation, using a molar concentration of c p As the unit of lithium salt concentration, it indicates the amount of lithium salt contained per unit volume, and the unit is mol L -1 . From the substance concentration c p The formula for obtaining the mass molar concentration m is as follows:

[0126] ;

[0127] where ρ p is the density of the polymer monomer, in g mL -1 .

[0128] The above formula takes into account the situation that the polymer electrolyte contains two components, polymer and lithium salt. The polymer electrolyte may also contain other additives.

[0129] Plasticizers, one of the additives, can be expressed in different units. The preparation process of gel polymer electrolytes is to soak dry electrolytes in plasticizers containing lithium salts to form an electrolyte system in which the plasticizer is fixed in a polymer matrix. Uptake% (liquid absorption rate) is an important parameter to measure the performance of gel polymer electrolytes. It indicates the ability of the polymer matrix to absorb liquid electrolytes and can also reflect the content of lithium salts and plasticizers in the polymer electrolyte. The calculation formula for Uptake% is as follows:

[0130] ;

[0131] Among them, M1 and M0 represent the mass of electrolyte after immersion and the mass of dry electrolyte before immersion, respectively.

[0132] The formula for obtaining the mass molar concentration m of lithium salt in the gel polymer solid electrolyte from the liquid absorption rate Uptake% is as follows:

[0133] ;

[0134] Among them, ρ l is the density of the plasticizer, in g mL -1 ;c l As the molar concentration of lithium salt in plasticizer, the unit is mol L -1 .

[0135] The unit of plasticizer content also needs to be unified, and the plasticizer mass percentage ω is calculated from Uptake% l The formula is as follows:

[0136] ;

[0137] The following are instructions for filling missing values: It is inevitable to encounter missing data in the process of collecting data. If the missing data is not particularly important, the missing values ​​can be filled. For the case where the temperature for measuring the lithium ion migration number is missing, since the lithium ion migration number of most polymer electrolytes is measured at 60°C, 60°C is used to fill the missing value; for a small number of samples with missing weight average molecular weight, the number average molecular weight or viscosity average molecular weight is used for approximate filling; if none of the three molecular weight data are given, only the source of the reagent is given, and the molecular weight of the reagent from that source is given in other data, the data is filled with other data.

[0138] The instructions for outlier processing are as follows: data points that do not match the actual situation are eliminated, and long-tailed data points that deviate significantly from other data are deleted to eliminate possible extreme errors.

[0139] The above cleaning steps improve the consistency and reliability of the data set, providing a reliable data foundation for accurate training and prediction of subsequent models.

[0140] To further analyze the data distribution, this embodiment uses data visualization methods to generate distribution diagrams and feature association diagrams of various performance indicators. For example, a probability density distribution diagram is used to display the distribution of various performance indicators, and a heat map is used to display the correlation between different features, which provides a basis for the next step of feature engineering and model construction.

[0141] Step A2: construct an efficient discriminant model based on the performance data set of the polymer electrolyte.

[0142] The discriminant model in this embodiment is a supervised learning model, which aims to predict the key properties (such as conductivity, lithium ion migration number, glass transition temperature and thermal decomposition temperature) of the polymer electrolyte by inputting its characteristics. Each performance has a corresponding discriminant model, but the construction process of each model is similar. The discriminant model can effectively support the prediction of polymer electrolyte performance and screen out molecules with excellent performance for the generation model. Steps A21-A27 are the specific steps for constructing the discriminant model:

[0143] Step A21: feature extraction;

[0144] In the dataset, the molecular structures of polymers and lithium salts are expressed in SMILES (Simplified Molecular Input Line Entry System). To meet the needs of machine learning modeling, Mordred descriptors are used to extract structural features of SMILES. This step can extract multiple features including the number of different atoms, molecular weight, polarization properties, acidity and alkalinity, hydrogen bond-related properties, etc. These features can be used to describe the molecular structure.

[0145] Step A22: data splitting;

[0146] To ensure the training effect and evaluation accuracy of the model, the data is divided into input (X) and output (Y), where X is the characteristic of the polymer electrolyte and Y is the performance to be predicted. At the same time, the data set is randomly divided into a training set and a test set. The training set is used to train the model; the test set is used to evaluate the performance and generalization ability of the model to ensure the model's predictive ability for new data.

[0147] Step A23: data standardization and normalization;

[0148] In order to improve the convergence speed, stability and prediction effect of the model, the input feature data is standardized and normalized to maintain the consistency of different feature scales. The standardized features are achieved by subtracting the mean and scaling to unit variance. The calculation formula is as follows:

[0149] ;

[0150] Among them, μ is the mean of the feature x, σ is the standard deviation of the feature x, and z represents the feature vector of the sample.

[0151] Normalization can scale the feature vector of each sample to unit norm (usually unit length), and the calculation formula is as follows:

[0152] ;

[0153] ;

[0154] Where ‖x‖ is the norm of the sample (usually the L2 norm, i.e. the Euclidean norm); n is the dimension of the feature vector, i.e. the number of features.

[0155] Standardization and normalization can eliminate the dimensional differences of each feature, ensuring that the model can more efficiently capture the relationship between features and outputs.

[0156] Step A24: feature screening;

[0157] In order to further improve the prediction accuracy of the model, reduce the computational complexity, and reduce the risk of overfitting, it is necessary to screen the extracted features. Specifically, variance filtering, correlation filtering, and RFECV (Recursive Feature Elimination with Cross-Validation), a feature selection method based on recursive feature elimination and cross-validation, are used.

[0158] The following are the instructions for variance filtering:

[0159] The basic idea of ​​variance filtering is to calculate the variance of each feature. Only features whose variance is greater than the set threshold will be retained, and the rest will be deleted. The calculation formula is as follows:

[0160] ;

[0161] Where X is the feature, N is the number of samples, and x i is the i-th sample value in the feature, and μ is the mean of the feature. Some features have very small variances (for example, all values ​​are similar), and these features may not contribute much to the predictive power of the model, so these low-variance features need to be removed.

[0162] The following are instructions for relevance filtering:

[0163] Correlation filtering uses the Pearson correlation coefficient to calculate the correlation r between each feature. The formula is as follows:

[0164] ;

[0165] Where X i and Y i are the i-th observation values ​​of the two features respectively, and are the means of the two features respectively. After obtaining the correlation between the features, a threshold is set and the features with correlation higher than this threshold are deleted.

[0166] The description of Recursive Feature Elimination Cross Validation (RFECV) is as follows:

[0167] RFECV is a feature selection method that combines recursive feature elimination (RFE) and cross validation (CV). It automatically determines the optimal number of features by gradually removing the features that have the least impact on model performance and using cross validation to evaluate the performance of feature subsets in each iteration.

[0168] Specifically, RFECV starts with all features and performs K-fold cross validation using the current feature set to evaluate the performance of the model. Then, the importance of each feature is evaluated based on the output of the model, and the features with the lowest importance are removed. This process is repeated until the predetermined number of features is reached or other stopping conditions are met. Finally, the number of features with the best performance is selected by comparing the average performance under different numbers of features.

[0169] The advantage of RFECV is that it can automatically optimize the number of features, reduce the risk of overfitting, and improve the performance and interpretability of the model.

[0170] The above three methods can reduce the dimension of the feature space and improve the generalization ability of the model.

[0171] Step A25: algorithm selection;

[0172] According to the requirements of polymer electrolyte performance prediction and data structure characteristics, three machine learning regression algorithms were selected, including random forest, gradient boosting tree, and XGBoost.

[0173] The description of random forest is as follows:

[0174] Random forest is an ensemble learning method based on decision trees. It improves prediction accuracy and stability by generating multiple decision trees and integrating the results of these trees. Each tree randomly extracts some samples and features from the data set during training to reduce the variance of the model.

[0175] Specifically, the random forest algorithm first creates multiple decision trees, creates multiple subsets by randomly sampling from the training data, and each tree is trained using these different subsets to form a diverse decision tree. When training each tree, features are randomly selected to determine the split of the node, which can avoid the excessive influence of certain features on the model.

[0176] Then, the test set is predicted, and the random forest will input the test set samples into all decision trees. Each tree will output a prediction value, and finally, the random forest will average these prediction values ​​to get the final prediction result, the formula is as follows:

[0177] ;

[0178] in, is the final prediction result, B is the number of trees, is the predicted value of the b-th tree.

[0179] The description of Gradient Boosting Tree (GBDT) is as follows:

[0180] GBDT is a decision tree algorithm based on gradient boosting. It gradually builds decision trees and adjusts the error of each tree to ultimately achieve better results. GBDT uses an additive model to combine multiple weak learners (usually decision trees) and uses gradient descent to minimize the loss function.

[0181] Assume that the model's prediction is , in The update of the step iteration is:

[0182] ;

[0183] Among them, F m-1 (x) is the prediction result of the first m-1 rounds; h m(x) is the value of the mth decision tree; η is the learning rate, which controls the update step size. The training goal of each step is to minimize the residual. In the regression task, the loss function L(y, F(x)) uses the mean square error loss function, y represents the true value, Represents partial derivative, negative gradient g m The calculation formula is as follows:

[0184] ;

[0185] When the loss function uses the mean square loss function, the negative gradient of the loss function g m The residual is the actual value minus the value predicted by the current model, so the training objective is to minimize the negative gradient of the current loss function. If the loss function is not the mean square loss function, the method of using the negative gradient of the loss function to replace the residual is also applicable.

[0186] The description of XGBoost is as follows:

[0187] XGBoost is an improved gradient boosting tree algorithm. Compared with the traditional GBDT, XGBoost introduces many optimizations, such as regularization, parallelization, missing value processing, and split search optimization, which makes the model training faster and has better performance. Unlike GBDT, whose objective function only contains the loss function, XGBoost's objective function obj consists of the loss function and the regularization term:

[0188] ;

[0189] in, is the loss function; is the regularization term, Indicates the actual output result. Represents the output of XGBoost. In order to optimize the objective function, the Taylor second-order expansion formula is used to approximate the objective function obj:

[0190] ;

[0191] Among them, g i and h i are the first and second derivatives of the loss function, respectively, f t represents the tree model added in the tth iteration (i.e., the tth tree), f t (x i ) represents the t-th tree for the i-th sample x i The predicted value of .

[0192] XGBoost then selects the optimal split point by maximizing the gain of the objective function. For each possible split point, the difference between the objective function values ​​before and after the split is calculated, and the split point with the maximum gain is selected. When the loop stop condition is reached, it will no longer iterate to generate new trees, and an optimal model will be obtained.

[0193] The above three algorithms have good performance in processing complex structure data and can effectively capture the nonlinear relationship between features and performance. In the subsequent steps, these three algorithms are used for machine learning modeling. By comparing the performance of different models in prediction tasks, the most suitable model is finally determined to meet the requirements of prediction accuracy and computational efficiency.

[0194] Step A26: Hyperparameter tuning;

[0195] To further improve the prediction accuracy of the model, hyperparameter tuning is required. The tool used for hyperparameter tuning can be Optuna. The tuned hyperparameters include the number of trees, tree depth, and learning rate, and Bayesian optimization is used for efficient parameter search. Through this optimization process, the generalization ability and prediction effect of the model are improved to ensure its reliability under different electrolyte structures and performance combinations.

[0196] Step A27: Model evaluation;

[0197] The regression model trained on the training set is applied to the test set to obtain the predicted value, and then a variety of indicators are used to comprehensively evaluate the prediction performance of the model. This step can use mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) or determination coefficient (R 2 ) to evaluate the predictive performance of the model.

[0198] The explanation of mean square error (MSE) is as follows:

[0199] MSE measures the average of the squared errors between the predicted values ​​and the actual values, and is calculated as follows:

[0200] ;

[0201] Where n is the number of samples, y i is the actual value of the ith sample, is the predicted value of the ith sample.

[0202] The explanation of root mean square error (RMSE) is as follows:

[0203] RMSE is the square root of MSE and represents the standard deviation of the prediction error. The formula is as follows:

[0204] .

[0205] The explanation of mean absolute error (MAE) is as follows:

[0206] MAE represents the average absolute error between the predicted value and the actual value. The formula is as follows:

[0207] .

[0208] About the coefficient of determination R 2 The instructions are as follows:

[0209] R 2 It indicates the proportion of the model explaining the data variance. The value is less than 1. The closer it is to 1, the stronger the explanatory power of the model. The calculation formula is as follows:

[0210] ;

[0211] in, is the mean of the actual values.

[0212] These evaluation indicators can comprehensively measure the prediction accuracy and stability of the model and verify the generalization ability of the model on the test set. By analyzing the evaluation results, the actual application effect of the model in the prediction of polymer electrolyte performance can be further determined.

[0213] Step A3: construct a molecular generation model based on graph grammar.

[0214] The molecular generative model algorithm used in this embodiment has the remarkable feature of effectively learning and generating high-quality, synthesizable molecules on a small data set. This feature is very suitable for research fields with limited data, such as the polymer electrolyte research in this embodiment. The construction and optimization process of the generative model includes the following steps A3.1-A3.4:

[0215] Step A3.1: Molecular graph representation;

[0216] The molecular structure can be represented as a graph structure, where nodes represent atoms and edges represent chemical bonds. For a molecule M, its graph H M = (V, E H ) contains a set of nodes V and a set of edges E H , where node V is the set of all atoms and edge E H This includes adding an edge for each bond that connects only two nodes, and adding an edge for all nodes in each ring (including aromatic rings). Figure 2 , Figure 2A schematic diagram of a molecular graph representation method provided in an embodiment of the present application, H represents hydrogen, C represents carbon, O represents oxygen, S represents sulfur, F represents fluorine, Carbon Atom Node represents carbon atom node, Oxygen AtomNode represents oxygen atom node, Sulphur Atom Node represents sulfur atom node, Fluorine Atom Node represents fluorine atom node, Single Bond represents single bond, Double Bond represents double bond, Aromatic Bond represents aromatic bond, Hyperedge represents hyperedge. Figure 2 As shown, the four benzene rings in the molecule correspond to the four edges in the figure.

[0217] Step A3.2: Generate graph grammar rules;

[0218] The following is an introduction to formal grammar: For a grammar G = (N, Σ, P, X) including an initial symbol X, a finite set of non-terminal symbols N, and a finite set of terminal symbols Σ. The grammar G describes how to use a set of production rules P = {p i |i = 1, …, k} constructs a string from the alphabet of the language, of the form p i : LHS→RHS, where LHS is the left symbol and RHS is the right symbol. Based on such a grammar, the rule whose left side matches the non-terminal symbol N in the current string is iteratively selected and replaced with the right side of the rule until the string does not contain any non-terminal symbol N.

[0219] Graph Grammar is an extension of formal grammar, see Figure 3 , Figure 3 A schematic diagram of a graph grammar generation rule and a molecular generation process using the grammar provided in an embodiment of the present application, wherein Production rules represent generation rules, Generation process represents generation process, p1, p2, p3 and p4 represent specific generation rules, X represents a string, and N represents a non-terminal symbol set. Figure 3 As shown in , the left and right sides of each generation rule are graphs. These graphs contain the initial node X, non-terminal nodes N, and terminal nodes representing atoms. The nodes corresponding to the hollow circles (i.e., white nodes) are anchor nodes, which do not change from the left to the right. The molecular graph generation process based on graph grammar is also Figure 3 Displayed in.

[0220] Graph grammar construction is a bottom-up process that iteratively samples a set of edges and shrinks them to a non-terminal node. For each shrinking step, a production rule is constructed and added to the grammar to obtain a new graph with fewer nodes and edges. Specifically for a molecule, at iteration t, the current graph H M , t = (V, E H ) Sample m edges, whose set is represented as ,in represents the i-th hyperedge sampled in the t-th round. The nodes connected by these edges The set is represented as . Then extract all connected components about these edges ,in and Respectively represent the nodes and edges in the i-th connected component sampled in the t-th round, and convert each connected component into a generation rule. The generation rule defines the nodes that are connected to the nodes in the connected component but do not belong to the connected component as anchor nodes ; The edges between anchor nodes and internal structures are represented as , the formulas for both are:

[0221] ;

[0222] ;

[0223] Among them, s represents a node belonging to a connected component, and v represents a node that does not belong to a connected component but is connected to the node s of the connected component.

[0224] Next, construct the generation rule p i : LHS→RHS, the left LHS and the right RHS are as follows, H represents the graph composed of edge E and node V, V L 、E L 、V R and E R Respectively represent the node on the left, the edge on the left, the node on the right, and the edge on the right. The left side contains non-terminal nodes N:

[0225] ;

[0226] ;

[0227] Each connected component is then replaced by a non-terminal node N and the graph is updated to H M , t+1 , and continue iterating until the entire graph is reduced to a single non-terminal node and replaced by the initial node X.

[0228] Step A3.3: Optimization and learning of graph grammar rules;

[0229] The optimization goal of graph grammar is to maximize the evaluation index of the input (for example, diversity and synthesis possibility). However, the optimization goal is usually non-differentiable and the optimization path cannot be directly calculated by gradient descent. Therefore, Monte Carlo sampling (MCsampling) and reinforcement learning policy gradient algorithm are used to optimize the objective function.

[0230] Step A3.4: Evaluate the metrics;

[0231] In order to evaluate the performance of the molecular generation model, this embodiment introduces evaluation indicators: effectiveness, uniqueness, novelty, diversity, retrosynthetic score (RS) and membership. This embodiment can evaluate the quality and diversity of generated molecules, or be used to evaluate the similarity and extrapolation ability of generated molecules with the distribution of training data.

[0232] Validity is the proportion of chemically valid molecules among the generated molecules. Uniqueness is the proportion of unique molecules among the generated molecules. Novelty is the proportion of molecules not in the training dataset among the generated molecules. Diversity is the average pairwise molecular distance between the generated molecules. Retrosynthetic score (RS) is the success rate of synthesizing the current molecule from commercially available raw materials. Membership is the proportion of generated molecules that belong to the training data.

[0233] Step A4: Use the discriminant model to guide the generation of the model.

[0234] By creating graph grammar generation rules from training data, although the molecules generated by the generative model are similar in chemical structure to the training data for polymer electrolytes, it cannot be ensured that these generated molecules have good properties such as conductivity. Therefore, the present invention proposes a method of combining a discriminant model with a generative model, so that the discriminant model can effectively guide the construction of the generative model, thereby ensuring that the generated molecules have excellent properties. The specific method includes the following steps A4.1 and A4.2:

[0235] Step A4.1: Construct a prediction function based on the discriminant model;

[0236] The input of this function is the molecular SMILES formula generated by the generative model, and the output is one or more properties predicted by the discriminative model based on the molecular SMILES formula, such as σ, t Li+ 、T g and T d This function specifically includes SMILES-style feature extraction and performance prediction, where the feature extraction part is the same as the steps when building the discriminant model to ensure consistency with the input of the trained discriminant model.

[0237] Step A4.2: Update the objective function of the graph grammar rules;

[0238] In the optimization of graph grammar rules, the objective function is proposed as , where M i and λ i Represent the value and weight of the i-th evaluation indicator respectively. The evaluation indicators are mentioned above, and they can evaluate the quality and diversity of the generated molecules, or be used to evaluate the similarity and extrapolation ability of the generated molecules with the distribution of training data. In order to combine the generative model with the discriminant model, the present invention integrates the performance value predicted by the discriminant model into the objective function of the graph grammar rule, so that the optimization of the graph grammar rule iterates in the direction of generating molecules with better performance.

[0239] The following are the instructions for normalization:

[0240] The value range of the above evaluation index is between 0 and 1. In order to make the performance score and the evaluation index score consistent when integrated into the same objective function, the predicted performance value needs to be normalized to obtain a performance score, and its value range is also between [0, 1]. Since the output of the discriminant model does not have a fixed range, the minimum and maximum normalization method cannot be directly used. The present invention adopts the threshold clipping mapping method, which defines a valid range based on the existing data set, thereby mapping the extreme values ​​of the discriminant model output to between 0 and 1.

[0241] First, define the maximum and minimum values, and record the 5th and 95th percentiles of the existing data set as P5 and P95 respectively, and use them as references for the minimum and maximum values ​​to avoid the influence of extreme outliers in the data on the setting. Then, perform threshold truncation and normalization on the output x of the discriminant model, and the formula is as follows:

[0242] ;

[0243] Threshold truncation means that if x is lower than the set minimum value, it is normalized to 0, and if it is higher than the maximum value, it is normalized to 1; and x within the maximum and minimum value range performs a simple linear normalization to map them to [0, 1].

[0244] In addition, not all indicators or performance are better the lower they are. For example, in practical applications, it is hoped that a polymer electrolyte with a lower glass transition temperature can be obtained. In this case, the above formula needs to be modified, namely:

[0245] ;

[0246] The description of the objective function weight is as follows:

[0247] After normalizing the performance value predicted by the discriminant model to obtain the performance score, the graph grammar rule objective function is updated to , where M i represents the evaluation index (such as diversity, RS, etc.) or the normalized performance score (such as conductivity score, lithium ion migration number score, etc.), λ i is the corresponding weight. Then the weight λ needs to be set and optimized.

[0248] The initial weight setting can be set according to the importance of different indicators or performance scores. If the actual demand is to focus on generating molecules with better performance, then the weight of the performance score should be set higher, so that the performance will have a greater impact on the objective function, and the optimized graph grammar rules will tend to generate molecules with better performance.

[0249] If the actual requirement is multi-objective optimization, the weights should be able to balance various indicators so that the generated molecules have the best overall performance. The balanced weight method can be used, that is, a set of weights that make multiple objective function values ​​higher can be found through grid search or Bayesian optimization.

[0250] The generative model guided by the discriminant model in this embodiment can generate high-quality molecules that meet specific electrolyte requirements and provide support for the development of polymer electrolyte materials.

[0251] The following is an example of Example B, Example C and Example D to illustrate the scheme of designing lithium battery polymer electrolyte using the above-mentioned coupled discriminant and generative models.

[0252] Embodiment B: High conductivity molecules are generated by combining a conductivity discriminant model with a generative model, specifically comprising the following steps B1-B3.

[0253] Step B1: Construct a conductivity discriminant model.

[0254] Specifically, the process of constructing the conductivity discriminant model includes the following steps B1.1-B1.5:

[0255] Step B1.1: Construct a conductivity data set of polymer electrolytes;

[0256] The conductivity data set of polymer electrolytes can include multiple data (e.g., 9117 data). Each data includes 5 features, namely, polymer monomer SMILES formula, polymer weight-average molecular weight, lithium salt SMILES formula, lithium salt mass molar concentration, conductivity temperature, and a label: conductivity. Features and labels are converted to units, missing values ​​are filled, and outliers are processed.

[0257] See also Figure 4 , Figure 4A probability density distribution diagram of the conductivity after logarithmization provided in the embodiment of the present application. Log Conductivity (log S cm -1 ) represents the logarithmic value of conductivity in Siemens per centimeter, and the ordinate represents probability.

[0258] Data processing and visualization were implemented using Python 3.11.7. The pandas library was used for data processing, and the matplotlib library was used for visualization.

[0259] Step B1.2: Feature extraction;

[0260] Mordred descriptors are used to extract structural features for polymer and lithium salt SMILES formulas. This includes about 28 specific features related to polarization, hydrogen bonds, rotation bonds, mass, acidity and alkalinity, and the number of atoms. Together with the original features polymer weight-average molecular weight, lithium salt mass molar concentration, and conductivity temperature, the total number of features in the next step of feature selection is 59.

[0261] All the features extracted by the above electrolyte discriminant model are as follows:

[0262] 'polymer apol' (polymer atomic polarizability), 'polymer bpol' (polymer bond polarizability), 'polymer nHBAcc' (polymer hydrogen bond acceptor number), 'polymer nHBDon' (polymer hydrogen bond donor number), 'polymer nRot' (polymer rotation bond number), 'polymer RotRatio' (polymer rotation bond ratio), 'polymer Vabc' (polymer van der Waals volume), 'polymer MW' (polymer exact molecular weight), 'polymer AMW' (polymer average exact molecular weight), 'polymer nAcid' (polymer acidic group number), 'polymer nBase' (polymer basic group number), 'polymer nAtom' (polymer atomic number), 'polymer nHeavyAtom' (polymer heavy atom number), 'polymernSpiro' (polymer spiro atom number), 'polymer nBridgehead' (polymer bridgehead atom number), 'polymer polymer nHetero', 'polymer nH', 'polymer nB', 'polymer nC', 'polymer nN', 'polymer nO', 'polymer nS', 'polymer nP', 'polymer nF', 'polymer nCl', 'polymer nBr', 'polymer nI', 'polymer nX', 'polymer Kier1', 'polymer Kier2', 'salt apol', 'salt apol' bpol' (bond polarizability of lithium salt molecule), 'salt nHBAcc' (number of hydrogen bond acceptors of lithium salt molecule), 'salt nHBDon' (number of hydrogen bond donors of lithium salt molecule), 'salt nRot' (number of rotation bonds of lithium salt molecule),'salt MW' (precise molecular weight of lithium salt molecules), 'saltAMW' (precise average molecular weight of lithium salt molecules), 'salt nAcid' (number of acidic groups in lithium salt molecules), 'saltnBase' (number of basic groups in lithium salt molecules), 'salt nAtom' (number of atoms in lithium salt molecules), 'saltnHeavyAtom' (number of heavy atoms in lithium salt molecules), 'salt nSpiro' (number of spiro atoms in lithium salt molecules), 'saltnBridgehead' (number of bridgehead atoms in lithium salt molecules), 'salt nHetero' (number of bridgehead atoms in lithium salt molecules), 'salt nH' (number of H atoms in lithium salt molecules), 'salt nB' (number of B atoms in lithium salt molecules), 'salt nC' (number of C atoms in lithium salt molecules), 'salt nN' (number of N atoms in lithium salt molecules), 'salt nO' (number of O atoms in lithium salt molecules), 'saltnS' (the number of S atoms in lithium salt molecules), 'salt nP' (the number of P atoms in lithium salt molecules), 'salt nF' (the number of F atoms in lithium salt molecules), 'salt nCl' (the number of Cl atoms in lithium salt molecules), 'salt nBr' (the number of Br atoms in lithium salt molecules), 'saltnI' (the number of I atoms in lithium salt molecules), 'salt nX' (the number of halogen atoms in lithium salt molecules), polymer molecular weight, lithium salt mass molar concentration and conductivity temperature. ,

[0263] Step B1.3: Data splitting;

[0264] The data set is split into a training set and a test set, where the training set is 80% of the total data (e.g., 7293 records); the test set is 20% of the total data (e.g., 1824 records).

[0265] Step B1.4: Feature screening;

[0266] After data standardization and normalization, feature screening is performed. The first step is variance filtering. The threshold is set to 0.005. 22 features have variances below the threshold and are deleted, leaving 37 features. The next step is correlation filtering. The threshold is set to 0.9. 10 features have correlations above this threshold and are deleted, leaving 27 features. Then, RFECV is used to select features. The random forest model is selected as the model, and the K-fold cross validation is set to 5 folds. After screening, the optimal number of features is finally determined to be 12. The correlation matrix of these features is as follows: Figure 5 As shown, Figure 5 A schematic diagram of a feature correlation matrix after conductivity data feature screening provided in an embodiment of the present application.

[0267] Figure 5 where mw represents molecular weight, salt nN represents the number of nitrogen atoms in the salt, salt nC represents the number of carbon atoms in the salt, salt nAcid represents the number of acid groups in the salt, salt AMW represents the average molecular weight of the salt, salt nHBAcc represents the number of hydrogen bond acceptors in the salt, salt bpol represents the polarity of the salt, salt apol represents the nonpolarity of the salt, salt nS represents the number of sulfur atoms in the salt, polymer nN represents the number of nitrogen atoms in the polymer, salt nP represents the number of phosphorus atoms in the salt, polymer AMW represents the average molecular weight of the polymer, polymer RotRatio represents the rotation ratio of the polymer, polymer nRot represents the number of rotatable bonds in the polymer, polymer nHBDon represents the number of hydrogen bond donors in the polymer, temperature represents temperature, molality represents molar concentration, conductivity represents conductivity, Upper Triangle represents upper triangle, SpearmanCorrelation represents Spearman correlation coefficient, Lower Triangle represents lower triangle, and Pearson Correlation represents Pearson correlation coefficient.

[0268] Step B1.4: Model training and evaluation;

[0269] After feature screening, the data is modeled by machine learning. Three algorithms, random forest, gradient boosting tree and XGBoost, are used, and the training set is trained through the scikit-learn library. At the same time, Optuna is used to find the best hyperparameter combination for modeling in a given hyperparameter space. For different algorithms, the optimized hyperparameters mainly include the number of evaluators, the depth of the model, the learning rate, etc. The optimized indicator is the R of the model on the validation set. 2 , the number of optimized experiments is 100. After the model training of the three algorithms is completed, the prediction performance of the model is evaluated using the test set. The evaluation indicators used include RMSE, MAE and R 2 , the evaluation results are shown in Table 1:

[0270] Table 1 Evaluation index calculation results

[0271]

[0272] RMSE Train Represents the root mean square error of the training set, MAE Train represents the mean absolute error of the training set, R 2 Train Represents the coefficient of determination of the training set, RMSE Test Represents the root mean square error of the test set, MAETest represents the mean absolute error of the test set, R 2 Test represents the coefficient of determination of the test set, XGBoost represents distributed gradient boosting library, RF represents random forest, and GBDT represents gradient boosted tree.

[0273] Comprehensive comparison shows that the prediction effect of the XGBoost model is relatively good, and the algorithm model training speed is the fastest and the time consumption is the shortest. Therefore, the model constructed by this algorithm is selected as the conductivity discriminant model.

[0274] Step B2: Build a molecular generative model based on graph grammar and combine it with the discriminant model;

[0275] Specifically, the process of implementing this step includes the following steps B2.1-B2.2:

[0276] Step B2.1: Constructing a polymer monomer molecular structure data set;

[0277] Since the purpose of the generative model in this embodiment is to generate molecules that can be used in lithium metal battery polymer electrolytes, the training data is currently used molecules of this type, so that the generative model will learn the structural characteristics of such molecules when learning the graph grammar generation rules and generate new molecules of this type. The training set used in this embodiment includes a total of 183 used polymer monomer molecular formulas.

[0278] Step 2.2: Set the objective function for graph grammar rule optimization.

[0279] The objective function for optimizing the graph grammar rules of the generative model is: , where M i and λ i Represent the value and weight of the i-th evaluation index respectively. The evaluation index and weight to be optimized need to be selected before optimization. The evaluation indexes selected in this embodiment are diversity, RS, and conductivity score, and the corresponding weights are λ1=1, λ2=1, and λ3=2. The set weights are only set according to the importance of the evaluation index of the molecule to be generated in this embodiment, and are not optimized.

[0280] The conductivity score is obtained by the conductivity prediction function. The prediction function includes passing the molecules generated by the generative model through the conductivity discriminant model to obtain the predicted conductivity and normalize it to obtain the conductivity score. The normalization method used is the threshold cutoff normalization method based on the existing conductivity data. The P5 and P95 of the logarithmic conductivity of the existing data are -8.648 and -2.928 respectively, and the unit is log S cm -1 . Therefore, the threshold truncation normalization formula is:

[0281] ;

[0282] Step B3: Use the discriminant model to guide the construction of the graph grammar generation model;

[0283] First, we use the RDKit package to extract a graph of molecules in the training data, including a set of nodes and a set of edges. Then we use the feature extraction function f(e) to extract features from each edge e. The feature extraction function used in this example is a pre-trained graph neural network implemented by the pytorch library, and the feature dimension extracted is 300. The edge weight function is defined as a composite function F(e;θ)=F that depends on the parameter θ. θ (f(e)), which uses a neural network model with two fully connected layers, the input layer dimension is 300, the hidden layer dimension is set to 128, the Dropout rate is set to 0.5, the output layer dimension is set to 2, and the softmax activation function is used to generate a probability distribution of X being 0 and 1, and X follows the selection probability based on edge e Bernoulli distribution. Edge weight function F θ It is a parameterized function that depends on the parameter θ. It further maps the output of the edge feature extraction function f(e) to the final result, that is, setting the corresponding weight for each edge.

[0284] Then the graph grammar rules are optimized, that is, the parameter θ of the edge weight function F(e;θ). Each graph grammar rule needs to be evaluated to obtain the value of its objective function. The specific method is to use the graph grammar to generate a certain number of molecules, and then evaluate these molecules with the selected evaluation index. After obtaining the value of the evaluation index, the objective function can be obtained. The number of generated molecules for evaluation is set to 100. After obtaining the objective function, the neural network of the edge weight function is iteratively optimized, the maximum number of iterations is set to 50, and the learning rate is set to 0.01, which is implemented through the pytorch library.

[0285] After iterative training, an optimized graph grammar rule generation model can be finally obtained.

[0286] Step B4: Generate samples using the generative model and predict conductivity using the discriminant model;

[0287] The optimal graph grammar rule generation model was used to randomly generate 2K molecular samples, and the discriminant model was used to predict their conductivity. Examples of the generated high conductivity molecules are as follows: Figure 6 shown. Figure 6 A schematic diagram of a high conductivity molecule generated by a trained graph grammar generation model provided in an embodiment of the present application.

[0288] Figure 6 A represents a specific atom or chemical group, log represents logarithm, Conductivity represents conductivity, S cm-1 represents Siemens per centimeter, N represents nitrogen atoms, O represents oxygen atoms, and F represents fluorine atoms. The logarithm of conductivity can be: -3.569 log S cm -1 、-3.587 log S cm -1 、-3.662 log S cm -1 、-4.057 log S cm -1 、-3.623 log S cm -1 、-4.176 log S cm -1 、-4.135 log S cm -1 , -4.119 log S cm -1 、-3.587log S cm -1 .

[0289] Example C: Molecules with excellent thermal properties are generated by combining multiple discriminant models with a generative model, specifically including the following steps C1-C4.

[0290] Step C1: constructing a glass transition temperature discriminant model and a thermal decomposition temperature discriminant model;

[0291] Specifically, the process of constructing the glass transition temperature discriminant model and the thermal decomposition temperature discriminant model includes the following steps C1.1-C1.5:

[0292] Step C1.1: Glass transition temperature and thermal decomposition temperature data set;

[0293] The glass transition temperature dataset has 8075 data points, and the thermal decomposition temperature dataset has 1742 data points. The initial features of these two datasets are only the polymer monomer SMILES formula and a label: glass transition temperature or thermal decomposition temperature. The thermal decomposition temperature is screened and only the temperature corresponding to a 5% reduction in mass when thermally decomposed is retained, that is, T d5 .

[0294] See also Figure 7 , Figure 7 A probability density distribution diagram of glass transition temperature data provided in an embodiment of the present application, Figure 7 The horizontal axis represents the glass transition temperature (°C), and the vertical axis represents the probability density.

[0295] See also Figure 8 , Figure 8 A probability density distribution diagram of thermal decomposition temperature data provided in an embodiment of the present application, Figure 8The abscissa Thermal decomposition temperature (°C) represents the thermal decomposition temperature, and the ordinate represents the probability density.

[0296] Step C1.2: feature extraction;

[0297] Mordred descriptors were used to extract structural features for polymer SMILES formula, including 31 specific features related to polarization properties, hydrogen bonds, rotation bonds, mass, acidity and basicity, and atomic number.

[0298] Step C1.3: data splitting;

[0299] The data splitting part is the same as that in Example B, where the glass transition temperature and thermal decomposition temperature data sets are split into a training set and a test set, wherein the training set is 80% of the total data and the test set is 20% of the total data.

[0300] Step C1.4: Feature screening;

[0301] The feature screening part is roughly the same as that of Example B, except that RFECV is not used in this example. After the glass transition temperature and thermal decomposition temperature data are screened, the number of features is 18 and 20, respectively. The correlation matrix of the features is as follows: Fig. 9 and Fig.10 shown.

[0302] Fig. 9A schematic diagram of a correlation matrix of characteristics of glass transition temperature data provided in an embodiment of the present application, polymer apol represents polymer non-polarity, polymer bpol represents polymer polarity, polymer nHBAcc represents the number of polymer hydrogen bond acceptors, polymer nHBDon represents the number of polymer hydrogen bond donors, polymer nRot represents the number of rotatable bonds of the polymer, polymer AMW represents the average molecular weight of the polymer, polymer nBase represents the number of polymer bases, polymer nBridgehead represents the number of polymer bridgehead atoms, polymer nHetero represents the number of polymer heteroatoms, polymer nH represents the number of polymer hydrogen atoms, polymer nN represents the number of polymer nitrogen atoms, polymer nO represents the number of polymer oxygen atoms, polymer nS represents the number of polymer sulfur atoms, polymer nP represents the number of polymer phosphorus atoms, polymer nF represents the number of polymer fluorine atoms, polymer nCl: represents the number of polymer chlorine atoms, polymer nBr represents the number of polymer bromine atoms, polymer Kier2 represents the polymer Kier index 2, Tg represents glass transition temperature, Upper Triangle represents the upper triangle, Spearman Correlation represents the Spearman correlation coefficient, Lower Triangle represents the lower triangle, and Pearson Correlation represents the Pearson correlation coefficient.

[0303] Fig.10A schematic diagram of a correlation matrix of characteristics of thermal decomposition temperature data provided in an embodiment of the present application, polymer apol represents polymer non-polarity, polymer bpol represents polymer polarity, polymer nHBAcc represents the number of polymer hydrogen bond acceptors, polymer nHBDon represents the number of polymer hydrogen bond donors, polymer nRot represents the number of polymer rotatable bonds, polymer MW represents polymer molecular weight, polymer AMW represents polymer average molecular weight, polymer nBase represents polymer base number, polymer nHeavyAtom represents polymer heavy atom number, polymer nBridgehead represents polymer bridgehead atom number, polymer nHetero represents polymer heteroatom number, polymer nC represents polymer carbon atom number, polymer nN represents polymer nitrogen atom number, polymer nO represents polymer oxygen atom number, polymer nS represents polymer sulfur atom number, polymer nP represents polymer phosphorus atom number, polymer nF represents polymer fluorine atom number, polymer nCl represents polymer chlorine atom number, polymer Kier1 represents polymer Kier index 1, polymer Kier2 represents polymer Kier index 2, T d5 Indicates thermal decomposition temperature.

[0304] Step C1.5: Model training and evaluation;

[0305] The model training and evaluation part is the same as that of Example B. The evaluation results of the glass transition temperature discriminant model are shown in Table 2 below:

[0306] Table 2 Evaluation results of glass transition temperature discriminant model

[0307]

[0308] Table 3 Evaluation results of glass transition temperature discriminant model

[0309]

[0310] Comprehensive comparison shows that the prediction effect of the XGBoost model is relatively good for both glass transition temperature and thermal decomposition temperature, and the algorithm model has the fastest training speed and the shortest training time. Therefore, the model constructed by this algorithm is selected as the discriminant model for glass transition temperature and thermal decomposition temperature.

[0311] Step C2: construct a molecular generative model based on graph grammar and combine it with the discriminant model;

[0312] Specifically, the process of constructing a molecular generative model and combining it with a discriminant model includes the following steps C2.1-C2.2:

[0313] Step C2.1: constructing a polymer monomer molecular structure data set;

[0314] This step is the same as in Example B.

[0315] Step C2.2: Setting the objective function for graph grammar rule optimization;

[0316] This step is substantially the same as in Example B. The evaluation indicators selected in this example include diversity, RS, glass transition temperature score and thermal decomposition temperature score, and the corresponding weights are λ1=1, λ2=1, λ3=2, and λ4=2.

[0317] The normalized formula for the glass transition temperature threshold cutoff is:

[0318] ;

[0319] The thermal decomposition temperature threshold cutoff normalization formula is:

[0320] ;

[0321] Step C3: Use the discriminant model to guide the construction of the graph grammar generation model:

[0322] This part is the same as Example B. Finally, an optimized graph grammar rule generation model can be obtained, which combines the glass transition temperature and thermal decomposition temperature discriminant model and can be used to generate molecules with excellent thermal properties.

[0323] Step C4: Generate samples from the generative model and predict performance of the discriminant model:

[0324] The optimal graph grammar rule generation model was used to randomly generate 2K molecular samples, and the discriminant model was used to predict their glass transition temperature and thermal decomposition temperature. Examples of molecules with excellent thermal properties generated include Fig.11 As shown, Fig.11 A schematic diagram of a molecule with excellent thermal performance provided in an embodiment of the present application, A represents a specific atom or chemical group, O represents an oxygen atom, N represents a nitrogen atom, P represents a phosphorus atom, F represents a fluorine atom, T d Indicates thermal decomposition temperature (such as 429.589°C, 400.651°C, 424.797°C, 397.963°C), T g Indicates the glass transition temperature (e.g., -47.960°C, -57.374°C, -48.916°C, -60.827°C).

[0325] Embodiment D: Molecules with excellent comprehensive performance are generated by combining multiple discriminant models with a generative model, specifically comprising the following steps D1-D4.

[0326] Step D1: Construct discriminant models for each performance:

[0327] Among them, the discriminant model of conductivity, glass transition temperature and thermal decomposition temperature is the same as that of Example B and Example C. The construction of the discriminant model of lithium ion migration number is introduced in steps D1.1-D1.5 below:

[0328] Step D1.1: Construct a lithium ion migration data set;

[0329] The lithium ion migration number dataset contains 3959 data points. The initial features include polymer monomer SMILES formula, polymer molecular weight and lithium salt. Fig.12 , Fig.12 This is a probability density distribution diagram of lithium ion migration number data provided in the embodiment of the present application, with the horizontal axis Li + The transference number represents the lithium ion migration number, and the ordinate represents the probability density.

[0330] Step D1.2: feature extraction;

[0331] In this step, Mordred descriptors are used to extract structural features for polymer SMILES formula, including 31 specific features related to polarization properties, hydrogen bonds, rotation bonds, mass, acidity and alkalinity, and number of atoms.

[0332] Step D1.3: Data splitting;

[0333] The data splitting part is the same as that in Example 1, and the lithium ion migration number data set is split into a training set and a test set, wherein the training set is 80% of the total data and the test set is 20% of the total data.

[0334] Step D1.4: Feature screening;

[0335] The feature screening part is roughly the same as that in Example 1. In this example, RFECV is not used. After the lithium ion migration number data is screened, the number of features is 12, and the correlation matrix of the features is as follows: Fig.13 As shown, Fig.13A schematic diagram of a correlation matrix of characteristics of lithium ion transfer number data provided in an embodiment of the present application, in which molality represents molar concentration, polymer_molecular_weight represents polymer molecular weight, nHBAcc represents the number of hydrogen bond acceptors, nHBDon represents the number of hydrogen bond donors, nRot represents the number of rotatable bonds, RotRatio represents the rotation ratio, AMW represents the average molecular weight, nAcid represents the number of acid groups, nBase represents the number of bases, nN represents the number of nitrogen atoms, nS represents the number of sulfur atoms, nF represents the number of fluorine atoms, and transference_number represents the transfer number.

[0336] Step D1.5: Model training and evaluation;

[0337] The model training and evaluation part is the same as that in Example 1. The evaluation results of the lithium ion migration number discriminant model are shown in Table 4 below:

[0338] Table 4 Evaluation results of the lithium ion migration number discriminant model

[0339]

[0340] A comprehensive comparison shows that the prediction effect of the XGBoost model is relatively good for the lithium ion migration number, and the algorithm model has the fastest training speed and the shortest training time. Therefore, the model constructed by this algorithm is selected as the discriminant model of the lithium ion migration number.

[0341] Step D2: Build a molecular generative model based on graph grammar and combine it with the discriminative model:

[0342] The specific step D2 may specifically include step D2.1 and step D2.2.

[0343] Step D2.1: polymer monomer molecular structure data set;

[0344] This step is the same as in Example B.

[0345] Step D2.2: Set the objective function for graph grammar rule optimization

[0346] This step is substantially the same as in Example B. The evaluation indicators selected in this example include diversity, RS, conductivity score, glass transition temperature score, thermal decomposition temperature score, and lithium ion migration number score, and the corresponding weights are λ1=1, λ2=1, λ3=2, λ4=2, λ5=2, and λ6=2.

[0347] The threshold cutoff normalization formulas of the three performance prediction functions of conductivity score, glass transition temperature score, and thermal decomposition temperature score are the same as those of Example B and Example C. The threshold cutoff normalization formula of the lithium ion migration number is:

[0348] ;

[0349] Step D3: Use the discriminant model to guide the construction of the graph grammar generation model:

[0350] This step is the same as Example B, and ultimately an optimized graph grammar rule generation model can be obtained. The generation model combines the conductivity, glass transition temperature, thermal decomposition temperature and lithium ion migration number discriminant model, and can be used to generate molecules with excellent comprehensive performance.

[0351] Step D4: Generate samples from the generative model and predict performance from the discriminant model:

[0352] Use the best graph grammar rule generation model to randomly generate 2K molecular samples, and use the discriminant model to predict their glass transition temperature and thermal decomposition temperature. Fig.14 , Fig.14 A schematic diagram of a molecule with excellent comprehensive performance provided in the embodiment of the present application. In the figure, A represents a specific atom or chemical group, O represents an oxygen atom, H represents a hydrogen atom, N represents a nitrogen atom, and the logarithm of the conductivity is -7.242 log S cm -1 、-7.110 log S cm -1 、-7.633 log S cm -1 、-7.517 log S cm -1 , lithium ion migration number Li + The transference numbers are 0.204, 0.195, 0.201 and 0.201, and the thermal decomposition temperature T d are 421.673 °C, 403.030 °C, 423.244 °C and 427.649 °C, and the glass transition temperature T g are 197.984°C, 109.603°C, 117.496°C and 190.446°C.

[0353] See also Fig.15 , Fig.15 A flow chart of a method for designing a lithium battery polymer electrolyte by coupling a discriminant and a generative model provided in an embodiment of the present application includes the following steps: (1) collecting and organizing data sets; (2) establishing a discriminant model; (3) establishing a generative model; and (4) using the discriminant model to guide the generative model.

[0354] This embodiment uses a machine learning model to directly predict the ionic conductivity of solid electrolytes. By learning and analyzing a large number of data sets, the structure-performance relationship of polymer electrolytes is constructed to predict ionic conductivity. This embodiment uses a machine learning method to construct a discriminant model to predict the core performance indicators of lithium metal battery polymer electrolytes. This embodiment uses graph grammar generation rules to ensure the performance of generated molecules for polymer electrolyte molecules.

[0355] In order to solve the problem of few types of polymer solid electrolytes for lithium metal batteries, this embodiment uses a generative model based on graph grammar that can be trained with very little data; this embodiment constructs a structured data set of solid polymer electrolytes for lithium metal batteries; this embodiment establishes a discriminant model of the core performance indicators of polymers used in lithium metal batteries; this embodiment couples the multi-objective discriminant model of polymer solid electrolytes with the generative model to reverse design polymer solid electrolytes.

[0356] A system for generating a molecular structure formula of a polymer electrolyte provided in an embodiment of the present application may include:

[0357] A data set acquisition module, used to acquire performance data sets and molecular structure data sets of polymer electrolytes;

[0358] A first model building module is used to build a discriminant model using the performance data set; wherein the discriminant model is used to calculate the performance index of the molecular structure formula, and the performance index includes any one or a combination of any several of the following: conductivity, lithium ion migration number, glass transition temperature and thermal decomposition temperature;

[0359] A second model building module is used to build a generative model using the molecular structure formula data set; wherein the generative model is used to generate a molecular structure formula of a polymer electrolyte;

[0360] A model training module, used for training the generative model based on the discriminant model;

[0361] The molecular generation module is used to generate a plurality of new molecular structural formulas of polymer electrolytes using the trained generative model if a molecular structure design task is received, and to screen the molecular structural formulas of the new polymer electrolytes using the discriminant model to obtain molecular structural formulas that meet the molecular structure design task.

[0362] This embodiment constructs a discriminant model through a performance data set of a polymer electrolyte. The discriminant model can learn the relationship between different molecular structures and performance. The discriminant model can be used to evaluate whether the newly generated molecular structure meets specific performance requirements. This embodiment constructs a generative model through a molecular structural formula data set of a polymer electrolyte. The generative model can automatically generate a new molecular structural formula of a polymer electrolyte based on a known molecular structure. This embodiment trains the generative model based on the discriminant model so that the generative model can generate a molecular structural formula whose performance indicators meet the requirements. After receiving a molecular structure design task, this embodiment can use the generative model to generate a plurality of new molecular structural formulas of polymer electrolytes, and screen them through the discriminant model to obtain a molecular structural formula that meets the molecular structure design task. The above process uses a discriminant model to guide the generative model to generate a molecular structural formula without manual participation. Therefore, this embodiment can automatically generate the molecular structural formula of the polymer electrolyte, thereby improving the efficiency and effectiveness of generating the molecular structural formula.

[0363] Furthermore, it also includes:

[0364] The data set processing module is used to perform data cleaning and data visualization operations on the performance data set and the molecular structure formula data set after obtaining the performance data set and the molecular structure formula data set of the polymer electrolyte.

[0365] Furthermore, the process of the first model building module using the performance data set to build a discriminant model includes: extracting data features from the performance data set, and standardizing and normalizing the data features; filtering the data features to obtain target data features; and performing machine learning modeling based on the target data features and corresponding labels in the performance data set to obtain the discriminant model.

[0366] Further, the performance data set includes any one or a combination of any one of a conductivity data set, a lithium ion migration number data set, a glass transition temperature data set and a thermal decomposition temperature data set; the conductivity data set includes a polymer structure, a polymer molecular weight, a lithium salt structure, a lithium salt concentration, a temperature and a conductivity label; the lithium ion migration number data set includes a polymer structure, a lithium salt structure, a lithium salt concentration, a temperature, an additive type, an additive content and a lithium ion migration number label; the glass transition temperature data set includes a polymer structure and a glass transition temperature label; the thermal decomposition temperature data set includes a polymer structure and a thermal decomposition temperature label;

[0367] Correspondingly, the first model building module performs machine learning modeling based on the target data features and corresponding labels in the performance data set to obtain the discriminant model, including:

[0368] Performing machine learning modeling based on target data features and conductivity labels in the conductivity data set to obtain a conductivity discriminant model;

[0369] And / or, performing machine learning modeling based on the target data features and lithium ion migration number labels in the lithium ion migration number data set to obtain a lithium ion migration number discriminant model;

[0370] And / or, performing machine learning modeling based on target data features and glass transition temperature labels in the glass transition temperature data set to obtain a glass transition temperature discriminant model;

[0371] And / or, machine learning modeling is performed based on the target data features and thermal decomposition temperature labels in the thermal decomposition temperature data set to obtain a thermal decomposition temperature discriminant model.

[0372] Furthermore, the process of the second model building module using the molecular structure formula data set to build a generative model includes: converting the molecular structure in the molecular structure formula data set into a graph structure; wherein the nodes in the graph structure represent atoms and the edges represent chemical bonds; generating a corresponding graph grammar according to the graph structure, and building the generative model based on the graph grammar.

[0373] Furthermore, the process of training the generative model based on the discriminant model by the model training module includes: constructing an objective function of the generative model; wherein the value of the objective function is a weighted calculation result of multiple evaluation indicators, and the evaluation indicators include scores corresponding to performance indicators calculated by the discriminant model; and iteratively training the generative model based on the objective function.

[0374] Furthermore, the process of iteratively training the generative model based on the objective function by the model training module includes: controlling the generative model to generate the i-th round output result; calculating the value of the objective function corresponding to the i-th round output result; judging whether the value of the objective function calculated most recently is within a preset interval; if so, judging that the generative model training is completed; if not, adjusting the parameters of the generative model, adding 1 to the value of i and entering the operation of controlling the generative model to generate the i-th round output result;

[0375] Among them, the evaluation indicators also include effectiveness index, uniqueness index, novelty index, diversity index, retrosynthesis score index and membership index; the effectiveness index is used to describe the proportion of valid molecular structures in the output results of the generative model; the uniqueness index is used to describe the proportion of non-repeated molecular structures in the output results of the generative model; the novelty index is used to describe the proportion of molecular structures in the output results of the generative model that are not in the molecular structure formula data set; the diversity index is used to describe the structural similarity between the output results of the generative model; the retrosynthesis score index is used to describe the synthesis success rate of the output results of the generative model; the membership index is used to describe the proportion of molecular structures in the output results of the generative model that belong to the molecular structure formula data set.

[0376] Since the embodiments of the system part correspond to the embodiments of the method part, please refer to the description of the embodiments of the method part for the embodiments of the system part, which will not be repeated here.

[0377] The present application also provides a storage medium on which a computer program is stored, and when the computer program is executed, the steps provided in the above embodiment can be implemented. The storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0378] The present application also provides an electronic device, which may include a memory and a processor, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps provided in the above embodiment may be implemented. Of course, the electronic device may also include various network interfaces, power supplies and other components.

[0379] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of this application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of this application.

[0380] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A method for generating a molecular structural formula of a polymer electrolyte, characterized in that: include: Obtain performance data sets and molecular structure data sets of polymer electrolytes; Constructing a discriminant model using the performance data set; wherein the discriminant model is used to calculate performance indicators of the molecular structure formula, the performance indicators including any one or a combination of any several of conductivity, lithium ion migration number, glass transition temperature and thermal decomposition temperature; Constructing a generative model using the molecular structural formula data set; wherein the generative model is used to generate a molecular structural formula of a polymer electrolyte; Training the generative model based on the discriminative model; If a molecular structure design task is received, the trained generative model is used to generate a plurality of new molecular structural formulas of polymer electrolytes, and the discriminant model is used to screen the molecular structural formulas of the new polymer electrolytes to obtain molecular structural formulas that meet the molecular structure design task.

2. The method for generating the molecular structural formula of a polymer electrolyte according to claim 1, characterized in that: After obtaining the performance data set and molecular structure data set of the polymer electrolyte, it also includes: Data cleaning and data visualization operations are performed on the performance data set and the molecular structure formula data set.

3. The method for generating the molecular structural formula of a polymer electrolyte according to claim 1, characterized in that: Constructing a discriminant model using the performance data set includes: Extracting data features from the performance data set, and standardizing and normalizing the data features; Filtering the data features to obtain target data features; Machine learning modeling is performed based on the target data features and corresponding labels in the performance data set to obtain the discriminant model.

4. The method for generating a molecular structural formula of a polymer electrolyte according to claim 3, characterized in that: The performance data set includes any one or a combination of any one of a conductivity data set, a lithium ion migration number data set, a glass transition temperature data set and a thermal decomposition temperature data set; the conductivity data set includes a polymer structure, a polymer molecular weight, a lithium salt structure, a lithium salt concentration, a temperature and a conductivity label; the lithium ion migration number data set includes a polymer structure, a lithium salt structure, a lithium salt concentration, a temperature, an additive type, an additive content and a lithium ion migration number label; the glass transition temperature data set includes a polymer structure and a glass transition temperature label; the thermal decomposition temperature data set includes a polymer structure and a thermal decomposition temperature label; Accordingly, machine learning modeling is performed based on the target data features and corresponding labels in the performance data set to obtain the discriminant model, including: Performing machine learning modeling based on target data features and conductivity labels in the conductivity data set to obtain a conductivity discriminant model; And / or, performing machine learning modeling based on the target data features and lithium ion migration number labels in the lithium ion migration number data set to obtain a lithium ion migration number discriminant model; and / or, performing machine learning modeling based on target data features and glass transition temperature labels in the glass transition temperature data set to obtain a glass transition temperature discriminant model; And / or, machine learning modeling is performed based on the target data features and thermal decomposition temperature labels in the thermal decomposition temperature data set to obtain a thermal decomposition temperature discriminant model.

5. The method for generating a molecular structural formula of a polymer electrolyte according to claim 1, characterized in that: Constructing a generative model using the molecular structure formula dataset includes: Converting the molecular structures in the molecular structure formula data set into a graph structure; wherein the nodes in the graph structure represent atoms and the edges represent chemical bonds; A corresponding graph grammar is generated according to the graph structure, and the generative model is constructed based on the graph grammar.

6. The method for generating a molecular structural formula of a polymer electrolyte according to claim 5, characterized in that: Training the generative model based on the discriminant model includes: Constructing an objective function of the generative model; wherein the value of the objective function is a weighted calculation result of multiple evaluation indicators, and the evaluation indicators include scores corresponding to performance indicators calculated by the discriminant model; The generative model is iteratively trained based on the objective function.

7. The method for generating a molecular structural formula of a polymer electrolyte according to claim 6, characterized in that: Iteratively training the generative model based on the objective function includes: Controlling the generative model to generate an i-th round of output results; Calculate the value of the objective function corresponding to the output result of the i-th round; Determine whether the value of the objective function calculated most recently is within a preset range; If yes, it is determined that the generative model training is complete; If not, adjust the parameters of the generative model, add 1 to the value of i and enter the operation of controlling the generative model to generate the i-th round output result; Among them, the evaluation indicators also include effectiveness index, uniqueness index, novelty index, diversity index, retrosynthesis score index and membership index; the effectiveness index is used to describe the proportion of valid molecular structures in the output results of the generative model; the uniqueness index is used to describe the proportion of non-repeated molecular structures in the output results of the generative model; the novelty index is used to describe the proportion of molecular structures in the output results of the generative model that are not in the molecular structure formula data set; the diversity index is used to describe the structural similarity between the output results of the generative model; the retrosynthesis score index is used to describe the synthesis success rate of the output results of the generative model; the membership index is used to describe the proportion of molecular structures in the output results of the generative model that belong to the molecular structure formula data set.

8. A system for generating a molecular structure formula of a polymer electrolyte, characterized in that: include: A data set acquisition module, used to acquire performance data sets and molecular structure data sets of polymer electrolytes; A first model building module is used to build a discriminant model using the performance data set; wherein the discriminant model is used to calculate the performance index of the molecular structure formula, and the performance index includes any one or a combination of any several of the following: conductivity, lithium ion migration number, glass transition temperature and thermal decomposition temperature; A second model building module is used to build a generative model using the molecular structure formula data set; wherein the generative model is used to generate a molecular structure formula of a polymer electrolyte; A model training module, used for training the generative model based on the discriminant model; The molecular generation module is used to generate a plurality of new molecular structural formulas of polymer electrolytes using the trained generative model if a molecular structure design task is received, and to screen the molecular structural formulas of the new polymer electrolytes using the discriminant model to obtain molecular structural formulas that meet the molecular structure design task.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method for generating the molecular structure formula of a polymer electrolyte according to any one of claims 1 to 7 when calling the computer program in the memory.

10. A storage medium, characterized in that: The storage medium stores computer executable instructions, and when the computer executable instructions are loaded and executed by the processor, the steps of the method for generating the molecular structure formula of the polymer electrolyte according to any one of claims 1 to 7 are implemented.

Images