Lithium battery electrolyte property prediction method based on deep learning

Abstract

The invention relates to the technical field of molecular property prediction, in particular to a lithium battery electrolyte property prediction method based on deep learning, comprising: S1, acquiring initial component information and external condition parameters of a target electrolyte, the external condition parameters comprising temperature, working voltage window and charge-discharge rate; s2, on the basis of the input parameter set, key structure representations corresponding to a plurality of key evolution stages are generated through a pre-trained multi-stage dynamic evolution network; s3, inputting the key structure characterization into a corresponding property evaluation network to obtain contribution degree evaluation of each evolution stage to the target property; and S4, based on each contribution degree evaluation and the corresponding evolution node weight, fusing and determining a final property prediction value of the target electrolyte. According to the method, virtual simulation is carried out on the whole process evolution of the electrolyte from the initial state to aging through the multi-stage dynamic evolution network, and the adaptability of the model to complex working conditions is improved.

Description

Technical Field

[0001] This invention relates to the field of molecular property prediction technology, and more specifically to a method for predicting the properties of electrolytes for lithium batteries based on deep learning. Background Technology

[0002] The performance, safety, and lifespan of lithium-ion batteries largely depend on the physicochemical properties of their electrolytes. Traditionally, electrolyte development has relied heavily on time-consuming and laborious trial-and-error experiments, as well as computational guidance based on limited experience or simplified theoretical models.

[0003] In recent years, machine learning methods have provided new approaches for the rapid prediction of electrolyte properties. These methods typically transform the molecular structure of the electrolyte into a fixed descriptor vector, and then directly map it to the target property through a complex neural network model. However, this approach has inherent limitations. The properties of the electrolyte are not determined by the initial molecular structure, but by macroscopic properties exhibited by factors such as solvation and desolvation of lithium ions, formation of a solid electrolyte interfacial film at the electrode interface, component degradation, and structural aging. A single structure is insufficient to reveal the differentiated contributions of different evolutionary stages to a particular comprehensive property, which is not conducive to guiding targeted optimization of electrolyte formulations. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a deep learning-based method for predicting the properties of electrolytes used in lithium batteries.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for predicting the properties of electrolytes for lithium batteries based on deep learning, comprising the following steps: S1. Obtain the initial component information and external condition parameters of the target electrolyte, including temperature, operating voltage window and charge / discharge rate; S2. Based on the input parameter set, generate key structural representations corresponding to several key evolutionary stages through a pre-trained multi-stage dynamic evolution network. S3. Input the key structural representations into the corresponding property evaluation networks to obtain the contribution assessment of each evolution stage to the target property. S4. Based on the assessment of each contribution and its corresponding evolution node weights, the final property prediction values ​​of the target electrolyte are determined by fusion.

[0006] In a preferred embodiment, in step S1, the initial component information of the target electrolyte is obtained through a data interface. The initial component information includes the type and ratio of the solvent system, the type and concentration of the lithium salt, and the chemical composition and mass percentage of all additives, and is stored in a structured form. Subsequently, external condition parameters are acquired, including temperature, operating voltage window, and charge / discharge rate. The temperature parameter is acquired in real time by a temperature sensor deployed in the electrolyte testing environment. The operating voltage window is determined based on the battery type and the thermodynamic stability window of the positive and negative electrode materials used in the target electrolyte. The charge / discharge rate is the current rate of the battery's rated capacity.

[0007] In a preferred embodiment, in step S2, the structured initial component information is converted into a dense vector through an embedding layer, while the external condition parameters are normalized. The dense vector is then concatenated with the normalized external condition parameters, and the concatenated vector is mapped to an initial state vector representing the initial state of the electrolyte through a fully connected coding network. Subsequently, the initial state vector and external condition parameters including the operating voltage window and charge / discharge rate are received, and the iterative simulation process is started from the initial state vector based on the preset total number of cycles. In each iteration step, the evolutionary computing unit receives the state vector output from the previous iteration step, the operating voltage and charge / discharge rate drive signal values ​​corresponding to the current step, and the dynamic coefficient determined by the temperature parameter. The evolutionary computation unit calculates the changes in the structure and properties of the electrolyte system based on the driving signal value and dynamic coefficient through pre-trained parameter mapping relationships, and applies the changes to the input previous state vector to generate the updated state vector for the current step. After a new state vector is generated in each iteration step, the built-in evolution stage determination module is immediately started. The evolution stage determination module runs three independent sub-determiners in parallel. After a new state vector is generated in each iteration step, the embedded evolution stage determination module is called synchronously. The evolution stage determination module integrates three parallel independent sub-determiners, corresponding to the three stages of solvation structure formation, interface film formation, and structure aging, respectively. The core operation of each sub-determiner is to first perform structure decoding and physicochemical quantity calculation on the current state vector, thereby extracting specific feature parameters. The specific steps of the structure decoding and calculation process are as follows: The state vector encoding restores the structural information of the system at the atomic and molecular scale through the decoding layer, including but not limited to the spatial coordinates of lithium ions, the spatial relative positions of particles, the type and state identifiers of particles, chemical bond connections, and statistical information. Based on the basic information obtained by decoding, macroscopic observable measurements are calculated in real time, including solvation structure-related features, interface film-related features, and aging signal-related features. The solvation structure-related feature quantities establish a spatial neighborhood range for each lithium ion based on the spatial position coordinates of the lithium ions obtained by decoding, and identify the spatial coordinates of all anions and coordinating atoms in the solvent molecules within this range, forming a list of coordinating neighbors for each lithium ion. For each lithium ion, the total number of anions and solvent coordinating atoms in its coordination neighbor list is counted and recorded as the instantaneous coordination number of the lithium ion. The average lithium-ion coordination number of the system is obtained by arithmetically averaging the instantaneous coordination numbers of all lithium ions. For each lithium ion, determine whether there is an anion in its coordination neighbor list, and whether the distance between the anion and the lithium ion is less than a preset contact coordination distance threshold. If so, determine that the lithium ion and the anion form a contact ion pair. The percentage of all lithium ions identified as forming contact ion pairs in the statistical system is calculated to determine the contact ion pair ratio. The interface film related features are based on the electrode surface. Along the direction perpendicular to the electrode surface, the space outside the electrode surface is divided into thin layers of thickness. For each thin layer, the number density of particles belonging to the solid electrolyte interface film components in the thin layer is statistically analyzed based on the particle position and type information obtained by decoding. Starting from the electrode surface, each spatial thin layer is scanned sequentially in a vertical outward direction. When the particle number density of the membrane components in multiple consecutive layers is higher than the preset membrane density threshold, these thin layers are identified as solid electrolyte interface membrane layers. The vertical distance between the outermost thin layer identified as a membrane layer and the electrode surface is calculated as the solid electrolyte interface membrane thickness. Within the spatial region identified as a membrane layer, the total number of particles belonging to the inorganic phase and the total number of particles belonging to the organic phase are counted respectively. Based on the statistical number of particles in each phase and the effective volume of a single particle, the volume fraction of the inorganic phase and the organic phase in the total volume of the membrane layer are calculated respectively. Based on the difference between the total volume of the membrane layer and the total volume of each phase, the volume fraction of the porous phase is calculated. Obtain the preset intrinsic ionic conductivity of each component, including the intrinsic conductivity of the inorganic phase, the intrinsic conductivity of the organic phase, and the intrinsic conductivity of the pore-filled phase; Based on the effective medium theory, the volume fraction and intrinsic conductivity of each component are substituted into the following relationship to obtain the estimated ionic conductivity of the interface film, which can be specifically expressed as: ; in, Indicates the volume fraction of inorganic phase. Indicates the intrinsic conductivity of the inorganic phase. This indicates an estimate of ionic conductivity. This indicates the volume fraction of the organic phase. Indicates the intrinsic conductivity of the organic phase. Indicates the pore volume fraction. Indicates the intrinsic conductivity of the porous phase; The aging signal related feature quantity is based on the particle type and state identifier obtained by decoding. The total number of all lithium ions marked as freely migratable in the current system is counted. The total number of migratable lithium ions currently counted is compared with the baseline total number of migratable lithium ions established in the early stage of the simulation evolution. The absolute reduction is calculated to obtain the active lithium loss. Based on the particle type information obtained from decoding, the total number of solvent molecules of a specific type in the current system is counted. The total number of key solvent molecules is divided by the current total volume of the system to obtain the current concentration of the key solvent. Furthermore, the percentage decrease of this concentration relative to the initial concentration is calculated to obtain the solvent consumption percentage. In multiple consecutive iterations, the cumulative amount of active lithium loss and the change in the key solvent concentration are recorded. By calculating the changes in these characteristic quantities per unit simulation time, the active lithium loss rate and the solvent consumption rate are obtained as auxiliary aging evaluation indicators. After completing the real-time calculation of the feature values, the three sub-determiners perform independent threshold determinations. Based on the pre-acquired target average lithium-ion coordination number and target contact ion ratio, as well as the calculated average lithium-ion coordination number and contact ion ratio, the absolute change in the average lithium-ion coordination number is calculated. The specific calculation formula is as follows: ; in, Indicates absolute change. This indicates the current average lithium-ion coordination number. This represents the initial average lithium-ion coordination number, used to calculate the absolute change in the contact ion ratio. ,in This indicates the current contact ion ratio. The formation degree component of the average lithium-ion coordination number is calculated using the following formula: ; in, Indicates the degree of formation component, This represents the target average lithium-ion coordination number, indicating the proportion of the current change to the target change. The degree of formation component of the contact ion ratio is calculated using the following formula: ; in, Indicates the degree of formation component, This indicates the target contact ion ratio. The initial contact ion ratio is represented by the formation degree component. Based on the formation degree component of the average lithium-ion coordination number and the contact ion ratio, the overall structure formation degree is calculated. The specific calculation formula is as follows: ; in, Indicates the degree of formation of the overall structure. , This represents the preset weighting coefficients, reflecting the relative importance of the two features to the formation of the overall structure; Set a time window containing the most recent k iteration steps, and obtain the average lithium-ion coordination number sequence calculated in these steps. and contact ion comparison sequence The linear regression method was used to fit the two sequences relative to the step number, and the linear regression slope of the average lithium ion coordination number sequence was calculated. The absolute value of the slope is the average rate of change of the average lithium ion coordination number within the nearest window. The linear regression slope of the contact ion pair ratio sequence was also calculated. The absolute value of the slope is the average rate of change of the contact ion pair ratio within the nearest window. The linear regression slope is compared with a preset rate of change stability threshold, which defines a maximum rate of change that is allowed to approach a stable state. If the absolute value of the linear regression slope is less than the preset rate of change stability threshold, and the linear regression slope is less than the preset rate of change stability threshold, then it is determined that these two feature parameters have approached stability. If the system determines that the following conditions are met simultaneously, if the rate of change is lower than the first preset stability threshold and the degree of structure formation is greater than or equal to the second preset formation threshold, then the electrolyte system is deemed to have completed the solvation structure formation stage in the current iteration step, and a solvation structure formation completion identifier is added to the state vector corresponding to this step. Based on the solid electrolyte interface film thickness sequence obtained by real-time calculation, the instantaneous growth rate of the film thickness in the current step is calculated, and based on the estimated value sequence of the membrane ionic conductivity, its change trajectory over time is monitored. If the system determines that the instantaneous growth rate has decreased from the historical peak and has been continuously lower than the third preset rapid growth threshold in a consecutive preset number of iteration steps, and at the same time the estimated value of the membrane ionic conductivity has decreased and remained within the range of the fourth preset densification conductivity threshold in a consecutive preset number of iteration steps, then the electrolyte system is considered to have completed the interface film formation stage in the current iteration step, and an electrolyte interface film formation completion label is added to the state vector corresponding to this step. Based on the calculated cumulative loss of active lithium and the real-time concentration of key solvents, the relative loss percentages of these values ​​relative to the stable baseline value established at the beginning of the cycle are calculated to obtain the cumulative loss of active lithium. If the cumulative loss of active lithium reaches the fifth preset lithium loss failure threshold, the electrolyte system is identified as entering the structural aging stage in the current iteration step, and an aging start marker is added to the state vector corresponding to this step. The system repeatedly executes the above iterative simulation, real-time decoding and stage determination steps to generate a complete state vector evolution sequence in sequence until the preset total number of cycles is reached. Finally, from the entire state vector sequence, all state vectors carrying stage identifier tags are selected. The state vectors contain a set of state vectors with key stage identifiers, which are defined and output as the key structural representation of the entire simulation evolution process.

[0008] In a preferred embodiment, in step S3, the evolutionary stage to which the key structural characterization belongs and the initial component information contained therein are input into the corresponding pre-trained property evaluation sub-network. Based on the key structural representations obtained from simulated evolution and the labels of their respective evolutionary stages, a contribution evaluation is performed. The key structural representations carrying labels indicating the completion of solvation structure formation and containing information on the corresponding initial components are input into the pre-trained first property evaluation subnetwork. The first property evaluation subnetwork decouples the input key structural representation into independent initial component feature vectors and solvation structure feature vectors. Based on the mapping relationship established by the pre-trained network parameters, it analyzes the influence weight of each group of molecular features in the initial component feature vector on a specific structural feature dimension in the solvation structure feature vector. Then, by simulating component feature perturbation and forward predicting macroscopic property changes, it quantifies the contribution of the solvation structure features formed by the current initial component information to the target macroscopic property, and outputs the first contribution evaluation value. The key structural characterization carrying the completed label of the solid electrolyte interface membrane formation and containing the corresponding initial component information is input into the pre-trained second property evaluation subnetwork. The second property evaluation subnetwork decouples the input key structural representation into independent initial component feature vectors and interface membrane structure feature vectors. Based on the causal analysis model embedded in the pre-trained network parameters, it calculates the causal effect by taking the features of each component in the initial component feature vector as the cause and the key membrane features and interface stability evaluation in the interface membrane structure feature vector as the effect. By quantifying the expected change in the interface stability evaluation score when the component features are fixed, it evaluates the causal contribution strength of the interface membrane features formed by the current initial component information to the interface stability, and outputs the second contribution evaluation value. The key structural representation corresponding to the structural aging stage and containing the corresponding initial component information is input into the third sub-network. The initial component information in the input key structural representation and the long-term evolved structural state vector are separated by feature separation. Based on the pre-trained network parameters, the correlation between the performance degradation signal in the separated long-term evolved structural state vector and the separated initial component information is analyzed. Based on the aforementioned correlation, the proportion of responsibility each component in the initial formulation should bear for the overall performance degradation is assessed and quantified as the third contribution assessment value. The first contribution assessment, the second contribution assessment, and the third contribution assessment are then summarized to form a structured set of contribution assessments for each stage.

[0009] In a preferred embodiment, in step S4, contribution evaluation values ​​from each property evaluation sub-network and their corresponding evolution node weights are obtained. Based on the evolution node weights, each contribution evaluation value is fused into a weighted contribution value through a weighted aggregation function. Based on all weighted contribution values, an initial property prediction value is generated through an integrated prediction function. The initial property prediction value is calibrated based on preset global calibration parameters to generate the final property prediction value. The global calibration parameters are a set of parameters predetermined through regression analysis based on the statistical relationship between simulated predicted values ​​and actual experimental measured values ​​of historical electrolyte systems, used to correct systematic prediction biases. The final property prediction value and its corresponding input parameters and calibration process are recorded. When the number of accumulated prediction records reaches a preset threshold, the global calibration parameters are refitted and updated based on the newly accumulated experimental data to achieve continuous optimization of the prediction model.

[0010] The beneficial effects of this invention are: This invention uses a multi-stage dynamic evolution network to virtually simulate the entire process evolution of the electrolyte from its initial state to aging, which improves the model's adaptability to complex working conditions. By using external condition parameters as inputs to the evolution network, it can predict the performance differences of the same formulation under different temperatures and rate of return, thus realizing performance analysis in multiple scenarios. Attached Figure Description

[0011] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0014] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0015] like Figure 1 This embodiment provides a deep learning-based method for predicting the properties of electrolytes used in lithium batteries, comprising the following steps: S1. Obtain the initial component information and external condition parameters of the target electrolyte, including temperature, operating voltage window and charge / discharge rate; Furthermore, the initial component information of the target electrolyte is obtained through a data interface. The initial component information includes the type and ratio of the solvent system, the type and concentration of the lithium salt, and the chemical composition and mass percentage of all additives, and is stored in a structured form. Subsequently, external condition parameters are acquired, including temperature, operating voltage window, and charge / discharge rate. The temperature parameter is acquired in real time by a temperature sensor deployed in the electrolyte testing environment. The operating voltage window is determined based on the battery type (e.g., lithium-ion battery, lithium metal battery) and the thermodynamic stability window of its positive and negative electrode materials used in the target electrolyte. The charge / discharge rate is the current rate of the battery's rated capacity.

[0016] S2. Based on the input parameter set, generate key structural representations corresponding to several key evolutionary stages through a pre-trained multi-stage dynamic evolution network. Furthermore, the structured initial component information is converted into a dense vector through an embedding layer, while the external condition parameters are normalized. The dense vector is then concatenated with the normalized external condition parameters, and the concatenated vector is mapped to an initial state vector representing the initial state of the electrolyte through a fully connected encoding network. Subsequently, the initial state vector and external condition parameters including the operating voltage window and charge / discharge rate are received, and the iterative simulation process is started from the initial state vector based on the preset total number of cycles. In each iteration step, the evolutionary computing unit receives the state vector output from the previous iteration step, the operating voltage and charge / discharge rate drive signal values ​​corresponding to the current step, and the dynamic coefficient determined by the temperature parameter. The evolutionary computation unit calculates the changes in the structure and properties of the electrolyte system based on the driving signal value and dynamic coefficient through pre-trained parameter mapping relationships, and applies the changes to the input previous state vector to generate the updated state vector for the current step. After a new state vector is generated in each iteration step, the built-in evolution stage determination module is immediately started. The evolution stage determination module runs three independent sub-determiners in parallel. After a new state vector is generated in each iteration step, the embedded evolution stage determination module is called synchronously. The evolution stage determination module integrates three parallel independent sub-determiners, corresponding to the three stages of solvation structure formation, interface film formation, and structure aging, respectively. The core operation of each sub-determiner is to first perform structure decoding and physicochemical quantity calculation on the current state vector, thereby extracting specific feature parameters. The specific steps of the decoding and calculation process are as follows: The state vector encoding restores the structural information of the system at the atomic and molecular scale through the decoding layer, including but not limited to the spatial coordinates of lithium ions, the spatial relative positions of particles, the type and state identifiers of particles, chemical bond connections, and statistical information, such as the particle number distribution and energy state distribution of various chemical species. Based on the basic information obtained by decoding, macroscopic observable measurements are calculated in real time, including solvation structure-related features, interface film-related features, and aging signal-related features. The solvation structure-related feature quantities establish a spatial neighborhood range for each lithium ion based on the spatial position coordinates of the lithium ions obtained by decoding, and identify the spatial coordinates of all anions and coordinating atoms in the solvent molecules within this range, forming a list of coordinating neighbors for each lithium ion. For each lithium ion, the total number of anions and solvent coordinating atoms in its coordination neighbor list is counted and recorded as the instantaneous coordination number of the lithium ion. The average lithium-ion coordination number of the system is obtained by arithmetically averaging the instantaneous coordination numbers of all lithium ions. For each lithium ion, determine whether there is an anion in its coordination neighbor list, and whether the distance between the anion and the lithium ion is less than a preset contact coordination distance threshold. If so, determine that the lithium ion and the anion form a contact ion pair. The percentage of all lithium ions identified as forming contact ion pairs in the statistical system is calculated to determine the contact ion pair ratio. The interface film related features are based on the electrode surface. Along the direction perpendicular to the electrode surface, the space outside the electrode surface is divided into thin spatial layers of thickness. For each thin spatial layer, based on the particle position and type information obtained by decoding, the number density of particles belonging to the solid electrolyte interface film components (including inorganic decomposition product particles and organic decomposition product particles) in the thin layer is counted. Starting from the electrode surface, each spatial thin layer is scanned sequentially in a vertical outward direction. When the particle number density of the membrane components in multiple consecutive layers is higher than the preset membrane density threshold, these thin layers are identified as solid electrolyte interface membrane layers. The vertical distance between the outermost thin layer identified as a membrane layer and the electrode surface is calculated as the solid electrolyte interface membrane thickness. Within the spatial region identified as a membrane layer, the total number of particles belonging to the inorganic phase and the total number of particles belonging to the organic phase are counted respectively. Based on the statistical number of particles in each phase and the effective volume of a single particle, the volume fraction of the inorganic phase and the organic phase in the total volume of the membrane layer are calculated respectively. Based on the difference between the total volume of the membrane layer and the total volume of each phase, the volume fraction of the porous phase is calculated. Obtain the preset intrinsic ionic conductivity of each component, including the intrinsic conductivity of the inorganic phase, the intrinsic conductivity of the organic phase, and the intrinsic conductivity of the pore-filled phase; Based on the effective medium theory, the volume fraction and intrinsic conductivity of each component are substituted into the following relationship to obtain the estimated ionic conductivity of the interface film, which can be specifically expressed as: ; in, Indicates the volume fraction of inorganic phase. Indicates the intrinsic conductivity of the inorganic phase. This indicates an estimate of ionic conductivity. This indicates the volume fraction of the organic phase. Indicates the intrinsic conductivity of the organic phase. Indicates the pore volume fraction. Indicates the intrinsic conductivity of the porous phase; The aging signal related feature quantity is based on the particle type and state identifier obtained by decoding. The total number of all lithium ions marked as freely migratable in the current system is counted. The total number of migratable lithium ions currently counted is compared with the baseline total number of migratable lithium ions established in the early stage of the simulation evolution. The absolute reduction is calculated to obtain the active lithium loss. Based on the particle type information obtained from decoding, the total number of solvent molecules of a specific type in the current system is counted. The total number of key solvent molecules is divided by the current total volume of the system to obtain the current concentration of the key solvent. Furthermore, the percentage decrease of this concentration relative to the initial concentration is calculated to obtain the solvent consumption percentage. In multiple consecutive iterations, the cumulative amount of active lithium loss and the change in the key solvent concentration are recorded. By calculating the changes in these characteristic quantities per unit simulation time, the active lithium loss rate and the solvent consumption rate are obtained as auxiliary aging evaluation indicators. After completing the real-time calculation of the feature values, the three sub-determiners perform independent threshold determinations. Based on the pre-acquired target average lithium-ion coordination number and target contact ion ratio, as well as the calculated average lithium-ion coordination number and contact ion ratio, the absolute change in the average lithium-ion coordination number is calculated. The specific calculation formula is as follows: ; in, Indicates absolute change. This indicates the current average lithium-ion coordination number. This represents the initial average lithium-ion coordination number, used to calculate the absolute change in the contact ion ratio. ,in This indicates the current contact ion ratio. The formation degree component of the average lithium-ion coordination number is calculated using the following formula: ; in, Indicates the degree of formation component, This represents the target average lithium-ion coordination number, indicating the proportion of the current change to the target change. The degree of formation component of the contact ion ratio is calculated using the following formula: ; in, Indicates the degree of formation component, This indicates the target contact ion ratio. The initial contact ion ratio is represented by the formation degree component. Based on the formation degree component of the average lithium-ion coordination number and the contact ion ratio, the overall structure formation degree is calculated. The specific calculation formula is as follows: ; in, Indicates the degree of formation of the overall structure. , This represents the preset weighting coefficients, reflecting the relative importance of the two features to the formation of the overall structure; Set a time window containing the most recent k iteration steps, and obtain the average lithium-ion coordination number sequence calculated in these steps. and contact ion comparison sequence The linear regression method was used to fit the two sequences relative to the step number, and the linear regression slope of the average lithium ion coordination number sequence was calculated. The absolute value of the slope is the average rate of change of the average lithium ion coordination number within the nearest window. The linear regression slope of the contact ion pair ratio sequence was also calculated. The absolute value of the slope is the average rate of change of the contact ion pair ratio within the nearest window. The linear regression slope is compared with a preset rate of change stability threshold, which defines a maximum rate of change that is allowed to approach a stable state. If the absolute value of the linear regression slope is less than the preset rate of change stability threshold, and the linear regression slope is less than the preset rate of change stability threshold, then it is determined that these two feature parameters have approached stability. If the system determines that the following conditions are met simultaneously, if the rate of change is lower than the first preset stability threshold and the degree of structure formation is greater than or equal to the second preset formation threshold, then the electrolyte system is deemed to have completed the solvation structure formation stage in the current iteration step, and a solvation structure formation completion identifier is added to the state vector corresponding to this step. Based on the solid electrolyte interface film thickness sequence obtained by real-time calculation, the instantaneous growth rate of the film thickness in the current step is calculated, and based on the estimated value sequence of the membrane ionic conductivity, its change trajectory over time is monitored. If the system determines that the instantaneous growth rate has decreased from the historical peak and has been continuously lower than the third preset rapid growth threshold in a consecutive preset number of iteration steps, and at the same time the estimated value of the membrane ionic conductivity has decreased and remained within the range of the fourth preset densification conductivity threshold in a consecutive preset number of iteration steps, then the electrolyte system is considered to have completed the interface film formation stage in the current iteration step, and an electrolyte interface film formation completion label is added to the state vector corresponding to this step. Based on the calculated cumulative loss of active lithium and the real-time concentration of key solvents, the relative loss percentages of these values ​​relative to the stable baseline value established at the beginning of the cycle are calculated to obtain the cumulative loss of active lithium. If the cumulative loss of active lithium reaches the fifth preset lithium loss failure threshold, the electrolyte system is identified as entering the structural aging stage in the current iteration step, and an aging start marker is added to the state vector corresponding to this step. It should be noted that the first preset stability threshold refers to the critical parameter used to determine whether the system structure tends to dynamic equilibrium during the solvation structure formation stage. Specifically, it refers to the upper limit of the change rate of key characteristic parameters such as the average lithium-ion coordination number and the contact ion ratio. The second preset formation threshold refers to the completion standard used to determine whether the structure has been fully formed during the solvation structure formation stage. Specifically, it refers to the lower limit of the comprehensive structure formation degree that must be reached or exceeded by the calculated value of the average lithium-ion coordination number and the contact ion ratio. The third preset rapid growth threshold refers to the critical growth rate value used to define the rapid growth state of the film layer during the interface film formation stage. When the instantaneous growth rate of the film thickness exceeds this threshold, it indicates that the interface film is in a rapid deposition state. The first stage, when the rate drops from its peak and remains below this threshold, indicates that the rapid growth phase of the film has ended and the growth has slowed down. The fourth preset densification conductivity threshold refers to the standard range of ionic conductivity used to determine whether the film has formed a dense structure during the interface film formation stage. When the estimated film ionic conductivity decreases and stabilizes within this range, it indicates that the film has changed from a loose and porous state to a dense state, which can effectively block the continuous decomposition of the electrolyte while allowing selective ion conduction. The fifth preset lithium loss failure threshold refers to the critical value used to determine the cumulative amount of active lithium loss when the battery performance begins to decline irreversibly during the structural aging stage. When the amount of active lithium loss reaches this threshold, it means that the battery capacity decay has reached the preset failure standard.

[0017] The system repeatedly executes the above iterative simulation, real-time decoding and stage determination steps to generate a complete state vector evolution sequence in sequence until the preset total number of cycles is reached. Finally, it automatically selects all state vectors carrying stage identifier tags from the entire state vector sequence. The state vectors contain a set of state vectors with key stage identifiers, which are defined and output as the key structural representation of the entire simulation evolution process.

[0018] S3. Input the key structural representations into the corresponding property evaluation networks to obtain the contribution assessment of each evolution stage to the target property. Furthermore, based on the evolutionary stage to which the key structural characterization belongs and the initial component information contained therein, it is input into the corresponding pre-trained property evaluation sub-network; Based on the key structural representations obtained from simulated evolution and the labels of their respective evolutionary stages, a contribution evaluation is performed. The key structural representations carrying labels indicating the completion of solvation structure formation and containing information on the corresponding initial components are input into the pre-trained first property evaluation subnetwork. The first property evaluation subnetwork decouples the input key structural representation into independent initial component feature vectors and solvation structure feature vectors. Based on the mapping relationship established by the pre-trained network parameters, it analyzes the influence weight of each group of molecular features in the initial component feature vector on a specific structural feature dimension in the solvation structure feature vector. Then, by simulating component feature perturbation and forward predicting macroscopic property changes, it quantifies the contribution of the solvation structure features formed by the current initial component information to the target macroscopic property, and outputs the first contribution evaluation value. The key structural characterization carrying the completed label of the solid electrolyte interface membrane formation and containing the corresponding initial component information is input into the pre-trained second property evaluation subnetwork. The second property evaluation subnetwork decouples the input key structural representation into independent initial component feature vectors and interface membrane structure feature vectors. Based on the causal analysis model embedded in the pre-trained network parameters, it calculates the causal effect by taking the features of each component in the initial component feature vector as the cause and the key membrane features and interface stability evaluation in the interface membrane structure feature vector as the effect. By quantifying the expected change in the interface stability evaluation score when the component features are fixed, it evaluates the causal contribution strength of the interface membrane features formed by the current initial component information to the interface stability, and outputs the second contribution evaluation value. The key structural representation corresponding to the structural aging stage and containing the corresponding initial component information is input into the third sub-network. The initial component information in the input key structural representation and the long-term evolved structural state vector are separated by feature separation. Based on the pre-trained network parameters, the correlation between the performance degradation signal in the separated long-term evolved structural state vector and the separated initial component information is analyzed. It should be noted that the first property evaluation subnetwork is for contribution evaluation based on solvation structure characteristics. It is used to analyze the contribution of initial chemical components to the macroscopic transport properties of the electrolyte by affecting the bulk solvation structure. The second property evaluation subnetwork is for contribution evaluation based on interface film characteristics. It is used to analyze the causal contribution of initial chemical components to the stability of the electrode interface by affecting the formation and characteristics of the solid electrolyte interface film.

[0019] Based on the aforementioned correlation, the proportion of responsibility each component in the initial formulation should bear for the overall performance degradation is assessed and quantified as the third contribution assessment value. The first contribution assessment, the second contribution assessment, and the third contribution assessment are then summarized to form a structured set of contribution assessments for each stage.

[0020] S4. Based on the assessment of each contribution and its corresponding evolution node weights, the final property prediction values ​​of the target electrolyte are determined by fusion. Furthermore, the contribution evaluation values ​​from each property evaluation sub-network and their corresponding evolution node weights are obtained. Based on the evolution node weights, the contribution evaluation values ​​are fused into a weighted contribution value through a weighted aggregation function. Based on all weighted contribution values, an initial property prediction value is generated through an integrated prediction function. The initial property prediction value is calibrated based on preset global calibration parameters to generate the final property prediction value. The global calibration parameters are a set of parameters predetermined through regression analysis based on the statistical relationship between simulated predicted values ​​and actual experimental measured values ​​of historical electrolyte systems, used to correct systematic prediction biases. The final property prediction value and its corresponding input parameters and calibration process are recorded. When the number of accumulated prediction records reaches a preset threshold, the global calibration parameters are refitted and updated based on the newly accumulated experimental data to achieve continuous optimization of the prediction model.

[0021] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0022] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0023] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0024] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0025] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0026] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0027] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the properties of electrolytes for lithium batteries based on deep learning, characterized by, The method comprises the following steps: S1, obtaining initial component information of a target electrolyte and external condition parameters, the external condition parameters including temperature, working voltage window and charge / discharge rate; S2, based on the input parameter set, generating key structure representations corresponding to several key evolution stages through a pre-trained multi-stage dynamic evolution network; S3, inputting the key structure representations into corresponding property evaluation networks respectively to obtain contribution degree evaluations of each evolution stage to the target property; S4, based on each contribution degree evaluation and its corresponding evolution node weight, fusing to determine the final property prediction value of the target electrolyte.

2. The deep learning-based electrolyte property prediction method for lithium batteries according to claim 1, characterized by, In S1, the initial component information of the target electrolyte is obtained through a data interface, including the type and ratio of the solvent system, the type and concentration of lithium salt, and the chemical components and mass percentage of all additives, and is stored in a structured form; Then, the external condition parameters are obtained, including temperature, working voltage window and charge / discharge rate, the temperature parameter is collected in real time by a temperature sensor deployed in the electrolyte test environment, the working voltage window is determined based on the battery type to which the target electrolyte is applied and the thermodynamic stability window of the positive and negative electrode materials thereof, and the charge / discharge rate is the current rate of the rated capacity of the battery. 3.The deep learning-based electrolyte property prediction method for lithium batteries according to claim 1, characterized by, In S2, the structured initial component information is converted into a dense vector through an embedding layer, and the external condition parameters are normalized, the dense vector and the normalized external condition parameters are spliced, and the spliced vector is mapped into an initial state vector representing the initial state of the electrolyte through a fully connected encoding network; Then, the initial state vector and the external condition parameters including the working voltage window and the charge / discharge rate are received, and based on the preset total cycle number, the iteration simulation process is started from the initial state vector; In each iteration step, the evolution calculation unit receives the state vector output by the previous iteration step, the working voltage and charge / discharge rate driving signal values corresponding to the current step, and the dynamic coefficient determined by the temperature parameter; The evolution calculation unit calculates the structure and property change amount of the electrolyte system according to the driving signal values and the dynamic coefficient through a pre-trained parameter mapping relationship, and applies the change amount to the input previous state vector to generate an updated state vector of the current step.

4. The deep learning-based electrolyte property prediction method for lithium batteries according to claim 3, characterized by, After generating a new state vector at each iteration step, an embedded evolution stage determination module is started immediately, and the evolution stage determination module runs three independent sub-determinators in parallel, and synchronously calls the embedded evolution stage determination module after generating a new state vector at each iteration step; The evolution stage determination module is integrated with three independent sub-determinators running in parallel, corresponding to the solvent structure formation, interface film formation and structure aging stages respectively, and the core operation of each sub-determinator is to first perform structure decoding and physical and chemical quantity calculation on the current state vector to extract specific feature parameters.

5. The deep learning-based electrolyte property prediction method for lithium batteries according to claim 4, characterized by, The specific steps of the structure decoding and calculation process are as follows: The state vector coding restores the structural information of the system at the atomic and molecular scales through the decoding layer, including but not limited to lithium ion spatial position coordinates, spatial relative positions of particles, particle types and state identifiers, chemical bond connection relationships, and statistical information. Based on the basic information obtained by decoding, macroscopic observable quantities are calculated in real time, including solvation structure-related characteristic quantities, interface film-related characteristic quantities, and aging signal-related characteristic quantities.

6. The deep learning-based electrolyte property prediction method for lithium batteries according to claim 5, characterized by, The solvation structure-related characteristic quantity establishes the spatial neighborhood range for each lithium ion based on the decoded lithium ion spatial position coordinates, and identifies the spatial coordinates of all coordinating atoms in the anions and solvent molecules within the range to form a coordination neighbor list for each lithium ion. For each lithium ion, the total number of anions and solvent coordinating atoms in the coordination neighbor list is counted, denoted as the instantaneous coordination number of the lithium ion. The instantaneous coordination numbers of all lithium ions are arithmetically averaged to obtain the average lithium ion coordination number representing the solvation structure of the system. For each lithium ion, it is determined whether there is an anion in its coordination neighbor list and the distance between the anion and the lithium ion is less than a preset contact coordination distance threshold. If so, it is determined that the lithium ion and the anion form a contact ion pair. The number of lithium ions in the system that are determined to form a contact ion pair is counted, and the percentage of the total number of lithium ions in the system is calculated to obtain the contact ion pair proportion. The interface film-related characteristic quantity takes the electrode surface as the reference and divides the space outside the electrode surface into spatial layers with a thickness perpendicular to the electrode surface. For each spatial thin layer, the number density of particles belonging to the solid electrolyte interface film component within the thin layer is counted based on the decoded particle position and type information. Starting from the electrode surface, each spatial thin layer is scanned in the vertical outward direction. When the particle number density of a plurality of layers is higher than a preset film density threshold, these thin layers are determined to be solid electrolyte interface film layers. The vertical distance between the outer boundary of the thin layer determined to be the film layer and the electrode surface is calculated as the solid electrolyte interface film thickness. In the spatial region determined to be the film layer, the total number of particles belonging to the inorganic phase and the total number of particles belonging to the organic phase are counted respectively. Based on the statistical number of each phase particles and the effective volume of each single particle, the volume fraction of the inorganic phase and the organic phase in the total volume of the film layer is calculated respectively. Based on the difference between the total volume of the film layer and the total volume of each phase, the volume fraction of the pore phase is calculated. The preset intrinsic ion conductivities of each component are obtained, including the intrinsic conductivities of the inorganic phase, the organic phase, and the pore filling phase. Based on the effective medium theory, the volume fraction and intrinsic conductivity of each component are substituted into the following relationship to obtain the estimated ion conductivity of the interface film, which can be specifically represented as: ； wherein, represents the inorganic phase volume fraction, represents the inorganic phase intrinsic conductivity, represents the estimated ionic conductivity, represents the organic phase volume fraction, represents the organic phase intrinsic conductivity, represents the pore volume fraction, represents the pore-filling phase intrinsic conductivity; The aging signal related characteristic quantity is calculated according to the particle type and state identification obtained by decoding, the total number of lithium ions marked as free migration in the current system is counted, the total number of migratory lithium ions obtained by the current counting is compared with the benchmark total number of migratory lithium ions established at the initial stage of simulation evolution, the absolute reduction amount is calculated, and the active lithium loss amount is obtained.

7. The deep learning-based electrolyte property prediction method for lithium batteries according to claim 6, characterized by, The solvation structure related characteristic quantity is established according to the spatial position coordinates of lithium ions obtained by decoding, the spatial neighborhood range of each lithium ion is established, and the spatial coordinates of all anions and solvent molecules in the range are identified to form a coordination neighbor list of each lithium ion; For each lithium ion, the total number of anions and solvent coordination atoms in the coordination neighbor list is counted, and the instantaneous coordination number of the lithium ion is recorded; The instantaneous coordination numbers of all lithium ions are arithmetically averaged to obtain the average lithium ion coordination number representing the solvation structure of the system; For each lithium ion, it is judged whether there is an anion in the coordination neighbor list, and the distance between the anion and the lithium ion is less than the preset contact coordination distance threshold value, if there is, it is determined that the lithium ion and the anion form a contact ion pair; The number of lithium ions in the system that are determined to form a contact ion pair is counted, and the percentage of the total number of lithium ions in the system is calculated to obtain the contact ion pair proportion; The interface film related characteristic quantity takes the electrode surface as the reference, divides the space outside the electrode surface into spatial layers with a thickness in the direction perpendicular to the electrode surface, and for each spatial thin layer, counts the number density of particles belonging to the solid electrolyte interface film component in the thin layer according to the particle position and type information obtained by decoding; Starting from the electrode surface, each spatial thin layer is scanned in the vertical outward direction, and when the particle number density of the film component of a plurality of layers is higher than the preset film density threshold value, these thin layers are determined as the solid electrolyte interface film layer, and the vertical distance between the outer boundary of the thin layer determined as the film layer and the electrode surface is calculated as the thickness of the solid electrolyte interface film. In the spatial region determined as the film layer, the total number of particles belonging to the inorganic phase and the total number of particles belonging to the organic phase are counted respectively, the volume fractions of the inorganic phase and the organic phase in the total volume of the film layer are calculated respectively according to the statistical number of each phase particles and the effective volume of a single particle, and the volume fraction of the pore phase is calculated according to the difference between the total volume of the film layer and the total volume of each phase. The preset intrinsic ion conductivities of each component are obtained, including the intrinsic conductivity of the inorganic phase, the intrinsic conductivity of the organic phase, and the intrinsic conductivity of the pore filling phase. Based on the effective medium theory, the volume fraction and the intrinsic conductivity of each component are substituted into the following relationship to obtain the estimated ion conductivity of the interface film, which can be specifically represented as: ； wherein, represents the inorganic phase volume fraction, represents the inorganic phase intrinsic conductivity, represents the estimated ionic conductivity, represents the organic phase volume fraction, represents the organic phase intrinsic conductivity, represents the pore volume fraction, represents the pore-filling phase intrinsic conductivity; The aging signal related characteristic quantity is calculated according to the particle type and state identification obtained by decoding, the total number of lithium ions marked as free migration in the current system is counted, the total number of migratory lithium ions obtained by the current counting is compared with the benchmark total number of migratory lithium ions established at the initial stage of simulation evolution, the absolute reduction amount is calculated, and the active lithium loss amount is obtained. 8.The deep learning-based electrolyte property prediction method for lithium batteries according to claim 1, characterized by, The S3, based on the key structure characterization belonging to the evolution stage and the initial component information contained therein, inputs it into the corresponding pre-trained property evaluation subnetwork; Based on the key structure characterization and its evolution stage label obtained by analog evolution, the contribution degree evaluation is performed, and the key structure characterization carrying the solvation structure formation completion label and containing the corresponding initial component information is input into the pre-trained first property evaluation subnetwork; The first property evaluation subnetwork decouples the input key structure characterization into independent initial component feature vectors and solvation structure feature vectors, analyzes the influence weight of each component molecular feature in the initial component feature vector on a specific structure feature dimension in the solvation structure feature vector based on the mapping relationship established by the pre-trained network parameters, and then quantifies the contribution of the solvation structure feature formed by the current initial component information to the target macroscopic property by simulating component feature disturbance and forward predicting macroscopic property changes. The output is the first contribution degree evaluation value; The key structure characterization carrying the solid electrolyte interface film formation completion label and containing the corresponding initial component information is input into the pre-trained second property evaluation subnetwork; The second property evaluation subnetwork decouples the input key structure characterization into independent initial component feature vectors and interface film structure feature vectors, and based on the causal analysis model embedded in the pre-trained network parameters, takes each component feature in the initial component feature vector as the cause and takes the key film feature and interface stability evaluation in the interface film structure feature vector as the effect. The cause and effect calculation is performed, the causal contribution intensity of the interface film feature formed by the current initial component information is evaluated by quantifying the expected change of the interface stability evaluation score of the fixed component feature, and the second contribution degree evaluation value is output. The key structure characterization corresponding to the structure aging stage and containing the corresponding initial component information is input into the third subnetwork, the initial component information in the input key structure characterization is separated from the long-term evolved structure state vector, and based on the pre-trained network parameters, the correlation between the performance degradation signal in the separated long-term evolved structure state vector and the separated initial component information is analyzed. Based on the correlation, the responsibility proportion of each component in the initial formula for the overall performance degradation is evaluated and quantified as the third contribution degree evaluation value. The first contribution degree evaluation, the second contribution degree evaluation and the third contribution degree evaluation are summarized to form a structured contribution degree evaluation set of each stage. 9.The deep learning-based electrolyte property prediction method for lithium batteries according to claim 1, characterized by, In the S4, the contribution degree evaluation values and their corresponding evolution node weights from each property evaluation subnetwork are obtained, the evolution node weights are used to fuse each contribution degree evaluation value into a weighted contribution value through a weighted aggregation function, the initial property prediction value is generated based on all weighted contribution values through an integrated prediction function, and the final property prediction value is generated by calibrating the initial property prediction value based on the preset global calibration parameter; The global calibration parameter is determined in advance by regression analysis of a parameter set based on a statistical relationship between simulated prediction values and actual experimental measurement values of a historical electrolyte system, and is used to correct systematic prediction deviation; The final property prediction value, its corresponding input parameter and calibration process are recorded, and when the number of accumulated prediction records reaches a preset threshold, the global calibration parameter is re-fitted and updated based on newly accumulated experimental data.

Images