A low-temperature electrolyte performance prediction method and system based on cross-scale collaborative optimization

Abstract

The application belongs to the technical field of intelligent molecular dynamics simulation, and relates to a low-temperature electrolyte performance prediction method and system based on cross-scale collaborative optimization, comprising: obtaining a low-temperature electrolyte structure model through molecular dynamics simulation; applying a Gaussian bias potential to lithium ion coordination in the low-temperature electrolyte structure model to obtain electrolyte dynamic data; obtaining a molecular graph vector and a non-covalent interaction energy vector of the electrolyte; fusing the electrolyte dynamic data, the molecular graph vector and the non-covalent interaction energy vector into a feature fusion vector; establishing a low-temperature electrolyte prediction model according to a graph neural network model, and training the low-temperature electrolyte prediction model through the feature fusion vector; and inputting a candidate electrolyte formula into the trained low-temperature electrolyte prediction model to generate key performance of the electrolyte. The application realizes high-precision prediction and screening of key performance of the electrolyte in a low-temperature range, and significantly reduces experimental cost and cycle of low-temperature electrolyte research and development.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization, belonging to the field of intelligent molecular dynamics simulation technology. Background Technology

[0002] With the increasing demand for low-temperature adaptability in new energy vehicles, polar scientific research equipment, and satellite energy storage systems, the performance of electrolytes in the -40°C to 0°C range has become a core factor restricting the low-temperature operating capability of batteries. For example, the conductivity of traditional lithium battery electrolytes drops by more than 70% at -30°C compared to room temperature, resulting in a battery capacity retention rate of less than 50%, severely impacting the equipment's range and reliability. Therefore, developing electrolyte formulations suitable for low-temperature scenarios and accurately predicting their key performance parameters has become an urgent need in the field of battery materials.

[0003] Traditional low-temperature electrolyte development relies on a "trial and error experiment + experience screening" model: extensive low-temperature experiments are conducted by manually adjusting the proportions of lithium salts (such as LiPF6), solvents (such as EC / DMC), and additives (such as FEC), taking months to complete. Furthermore, limitations in experimental conditions prevent the exploration of millions of formulation combinations. In terms of theoretical simulation, while early molecular dynamics (MD) simulations could obtain the microstructure of the electrolyte, the slowed molecular motion at low temperatures leads to a more than three-fold increase in the simulation equilibrium time; for example, an electrolyte at -40°C requires 72 hours to reach equilibrium, resulting in high computational costs and hindering high-throughput screening.

[0004] The core deficiency of existing technologies lies in the lack of a feature system and prediction model specific to low temperatures, specifically the following issues:

[0005] Lack of low-temperature characteristics: Low-temperature sensitive parameters such as ionic coordination number CN and hydrogen bond density HBN are not quantified, making it impossible to capture the dynamic changes in intermolecular interactions in the range of -40°C to 0°C.

[0006] Insufficient model adaptability: The general graph neural network does not focus on key low-temperature groups (such as fluorinated groups) and has insufficient learning ability for non-covalent interaction energies (electrostatic energy, dispersion energy) at low temperatures, resulting in inaccurate predictions of parameters such as freezing point and viscosity.

[0007] Low screening efficiency: The lack of automated low-temperature constraint embedding requires manual elimination of a large number of invalid formulas, and the screening cycle for thousands of formulas exceeds 10 days, which cannot meet the needs of rapid research and development. Summary of the Invention

[0008] To address the aforementioned issues, the present invention aims to provide a method and system for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization. This method achieves high-precision prediction and high-throughput screening of key electrolyte performance within the range of -40°C to 0°C, significantly reducing the experimental costs and time required for the development of low-temperature electrolytes.

[0009] To achieve the above objectives, the present invention proposes the following technical solution: a method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization, comprising the following steps: obtaining a low-temperature electrolyte structure model through molecular dynamics simulation; applying a Gaussian bias potential to the lithium-ion coordination in the low-temperature electrolyte structure model to obtain electrolyte dynamic data; acquiring the molecular graph vector and non-covalent interaction energy vector of the electrolyte; fusing the electrolyte dynamic data, molecular graph vector, and non-covalent interaction energy vector into a feature fusion vector; establishing a low-temperature electrolyte prediction model based on a graph neural network model, and training the low-temperature electrolyte prediction model with the feature fusion vector; inputting candidate electrolyte formulations into the trained low-temperature electrolyte prediction model to generate the key performance characteristics of the electrolyte.

[0010] Furthermore, the low-temperature electrolyte structure model includes the electrolyte molecular structure at several different temperature points, which are in the range of -40°C to 0°C; the electrolyte dynamic data includes ion mobility, solvation coordination number CN, and hydrogen bond density HBN; the solvation coordination number CN is obtained by calculating the number of solvent molecules in the first coordination shell of lithium ions through radial distribution function (RDF) integration; the hydrogen bond density HBN is obtained by calculating the hydrogen bond density per unit volume by statistically counting the number of hydrogen bonds with OH bond lengths.

[0011] Furthermore, the method for obtaining the molecular graph vector is as follows: the electrolyte molecules are decomposed into atomic nodes Li / C / O / F and bond edges; based on the atomic nodes Li / C / O / F and bond edges, a molecular graph vector is generated using the Weisfeiler-Lehman algorithm, and atomic charge and bond order information are encoded in the molecular graph vector; the non-covalent interaction energy includes electrostatic energy and dispersion energy at low temperature; the non-covalent interaction energy is obtained through density functional theory.

[0012] Furthermore, the feature fusion vector includes point-level features and bond-level features. The low-temperature electrolyte prediction model includes a convolutional layer, a pooling layer, and an output layer. The convolutional layer extracts local structural features using the GCNConv operator; it increases the weight of fluorinated groups while suppressing long-chain alkyl groups using a self-attention module; the pooling layer reduces the dimensionality of the local structural features, retains key nodes, and reduces computational complexity; the output layer outputs the key performance characteristics of the electrolyte, including conductivity, viscosity, and freezing point.

[0013] Furthermore, the weighted mean square error (MSE) of the loss function of the output layer is:

[0014] ,

[0015] The Loss expression represents the total loss function, which comprehensively measures the prediction errors of conductivity, viscosity, and freezing point. It is obtained through weighted summation. The mean squared error (MSE) of conductivity, viscosity and freezing point are expressed separately. 0.5, 0.3 and 0.2 are the weightings of conductivity, viscosity and freezing point, respectively, reflecting the differences in the importance of the indicators. That is, conductivity is the core indicator of low temperature battery and has the highest weight. Viscosity and freezing point have the next lower influence and their weights decrease.

[0016] in, N is the quantity of electrolyte formulation involved in the loss calculation. and The predicted and actual conductivity values ​​for the i-th sample are expressed separately. The formulas for calculating the mean square error of viscosity and freezing point are the same as those for conductivity.

[0017] Furthermore, the method for training the low-temperature electrolyte prediction model is as follows: based on the Uni-mol open-source model, using low-temperature simulation data, the low-temperature electrolyte prediction model is pre-trained to initialize the parameters of the low-temperature electrolyte prediction model; the feature fusion vector is used to train the low-temperature electrolyte prediction model; several electrolyte formulations with the largest weighted mean square error (MSE) are selected, and molecular dynamics simulations are performed according to the electrolyte formulations to expand the training set data; the expanded training set is input into the low-temperature electrolyte prediction model for training until the value of the loss function of the output layer is less than the threshold, and the trained low-temperature electrolyte prediction model is output.

[0018] Furthermore, the results of the low-temperature electrolyte prediction model are subjected to low-temperature constraint filtering, the conditions of which are:

[0019]

[0020] in, for conductivity, for viscosity; It is the freezing point;

[0021] Based on the results of low-temperature constraint filtering, the parameters of the molecular dynamics simulation are automatically calibrated.

[0022] Furthermore, the conductivity value of the electrolyte is obtained according to the low-temperature electrolyte prediction model, and a conductivity value versus time curve is established. The conductivity temperature coefficient is obtained according to the conductivity value versus time curve. The results of the low-temperature electrolyte prediction model are input into the Eyring equation to obtain the viscosity activation energy of the electrolyte.

[0023] Furthermore, molecular dynamics calculations are performed on the results of the low-temperature electrolyte prediction model to calculate the stability of the solvated sheath layer; the results of the low-temperature electrolyte prediction model are further screened based on the stability of the solvated sheath layer.

[0024] This invention also discloses a low-temperature electrolyte performance prediction system based on cross-scale collaborative optimization, comprising: a molecular dynamics simulation module for obtaining a low-temperature electrolyte structure model through molecular dynamics simulation; an electrolyte dynamic data acquisition module for applying a Gaussian bias potential to the lithium-ion coordination in the low-temperature electrolyte structure model to obtain electrolyte dynamic data; a molecular graph energy vector acquisition module for acquiring the molecular graph vector and non-covalent interaction energy vector of the electrolyte; a feature fusion module for fusing the electrolyte dynamic data, molecular graph vector, and non-covalent interaction energy vector into a feature fusion vector; a model training module for establishing a low-temperature electrolyte prediction model based on a graph neural network model and training the low-temperature electrolyte prediction model with the feature fusion vector; and a model output module for inputting candidate electrolyte formulations into the trained low-temperature electrolyte prediction model to generate the key performance characteristics of the electrolyte.

[0025] The technical solution of the present invention has at least the following technical effects or advantages:

[0026] 1. Existing technologies mainly rely on first-principles calculations or general molecular databases at room temperature for electrolyte formulation prediction, without including data specific to low temperatures (-40°C to 0°C). This invention, for the first time, quantifies dynamic parameters such as ion mobility, solvation coordination number (CN), and hydrogen bond density (HBN), solving the problem of insufficient description of intermolecular interactions at low temperatures in existing technologies. It integrates molecular dynamics simulations and graph neural networks to construct a direct mapping between low-temperature microstructure and macroscopic properties, with prediction errors controlled within 10%.

[0027] 2. Existing electrolyte formulation prediction technologies suffer from limited feature extraction, focusing only on static molecular features at room temperature, such as atomic charge, bond length, and Fukui function. They fail to quantify dynamic interactions at low temperatures, such as low-temperature non-covalent interaction energy and solvated sheath stability, making it difficult to comprehensively reflect the electrolyte's characteristics at low temperatures. This invention adds features reflecting the electrolyte's low-temperature performance, such as electrolyte dynamic data, molecular graph vectors, and non-covalent interaction energy vectors. Feature fusion is performed to extract multiple low-temperature sensitive features and integrate molecular graph vectors with non-covalent interaction energies, enabling a more accurate characterization of the electrolyte's low-temperature properties and providing richer and more effective information for training low-temperature electrolyte prediction models.

[0028] 3. Existing technologies generate virtual molecules based on room-temperature molecular nodes, without distinguishing the activity of molecular groups at low temperatures; in this invention, the charge distribution and coordination state of low-temperature sensitive elements (Li, F) are embedded in the virtual molecular nodes, and the edge features of the molecular graph vectors increase the non-covalent interaction energy parameters between molecules at low temperatures, fully considering the interaction between the focused fluorinated solvent (FEC) and the additive (LiDFP) at low temperatures.

[0029] 4. This invention incorporates low-temperature physical constraints, such as σ≥1.0 mS / cm and T_mp≤-50°C, automatically eliminating 97% of invalid formulations and reducing manual intervention costs; it improves the correlation between characteristics and low-temperature performance by 40%, controls the prediction error of key performance (such as conductivity and freezing point) to ≤10%, supports parallel evaluation of thousands of formulations, and shortens the R&D cycle from several weeks in traditional methods to several hours, significantly reducing trial and error costs.

[0030] 5. This invention realizes full automation and precision in the low-temperature electrolyte formulation process from data generation to model prediction to screening and optimization, providing an efficient technical tool for the research and development of low-temperature battery materials. Through parallel computing with GPU clusters, the single formulation prediction time is less than 10 ms, and the screening cycle of thousands of formulations is shortened from more than 10 days in traditional methods to minutes, improving efficiency by more than 100 times.

[0031] 6. The solution in this invention reduces the conductivity prediction error of the -40°C electrolyte by 40% and the viscosity prediction error by 35% compared to the traditional graph neural network model. The solution in this invention uses a weighted mean square error loss function to achieve joint optimization of multiple performance parameters, avoid single task bias, and reduce the overall prediction error by 20% compared to the single task model, thus breaking through the limitations of traditional methods in modeling multi-parameter coupling relationships.

[0032] 7. This invention dynamically expands the training set, reducing the cost of manual annotation; virtual screening replaces traditional experiments, while avoiding the large-scale use of toxic reagents. Attached Figure Description

[0033] Figure 1This is a flowchart of a low-temperature electrolyte performance prediction method based on cross-scale collaborative optimization in one embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the structure of a low-temperature electrolyte structure model in one embodiment of the present invention;

[0035] Figure 3 This is a flowchart of an automated screening method in one embodiment of the present invention. Detailed Implementation

[0036] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention is described in detail through specific embodiments. However, it should be understood that the specific embodiments are provided only for a better understanding of the present invention and should not be construed as limiting the present invention. In the description of the present invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0037] To address the problems of missing temperature-sensitive features, insufficient model prediction accuracy, and low screening efficiency in existing electrolyte performance testing methods, this invention proposes a low-temperature electrolyte performance prediction method and system based on cross-scale collaborative optimization. By integrating low-temperature electrolyte dynamic data and an improved attention mechanism, a low-temperature electrolyte prediction model is constructed that can accurately capture the evolution of ion transport and solvation structures at low temperatures. An automatic screening method with low-temperature physical constraints is also implemented, achieving high-precision prediction and high-throughput screening of key electrolyte properties (conductivity, viscosity, and freezing point) within the range of -40°C to 0°C. This invention overcomes the efficiency and accuracy bottlenecks in low-temperature electrolyte R&D through precise quantification of low-temperature sensitive features, deep collaboration of cross-scale models, and the engineering implementation of a fully automated screening system. Compared to traditional methods, the prediction error of key electrolyte properties is reduced by more than 60%, the R&D cycle is shortened by 90%, and experimental costs are reduced by 85%. It also provides a reusable low-temperature material design method example for the electrolyte industry, demonstrating significant scientific value and engineering application prospects. The following detailed description of the invention, with reference to the accompanying drawings, illustrates the solution in detail.

[0038] Example 1

[0039] This embodiment discloses a method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization, such as... Figure 1 As shown, it includes the following steps:

[0040] S1 obtained a low-temperature electrolyte structure model through molecular dynamics simulation.

[0041] Molecular dynamics simulations were performed using the molecular dynamics software GROMACS. GROMACS boasts high computational efficiency and parallel processing capabilities, supports multi-threaded GPU acceleration, significantly shortens low-temperature simulation time, and is suitable for high-throughput data generation. The force field selection for molecular dynamics simulations is compatible with OPLS-AA for both lithium salt and solvent molecules. The OPLS-AA force field can optimize lithium salt and solvent parameters, with a fluorinated group dispersion energy calculation error of <8%, and can accurately describe hydrogen bonding / coordination interactions.

[0042] 10,000 candidate formulations were generated by randomly combining lithium salts, solvents, and additives. The solvent ratio in each formulation fluctuated within ±10%, and the additive content ranged from 1% to 5%. A virtual molecular graph was constructed, with each formulation containing 5-10 molecular nodes. The edges of the molecular graph represented the non-covalent intermolecular interactions, thus obtaining a low-temperature electrolyte structure model containing lithium salt LiFSI (1.0M), solvent EC / DMC / FEC (volume ratio 3:5:2), and additive VC (3%). The low-temperature electrolyte structure model included the electrolyte molecular structure at several different temperature points, ranging from -40°C to 0°C. Specifically, in this embodiment, electrolyte molecular structures corresponding to five temperature points—-40°C, -30°C, -20°C, -10°C, and 0°C—were collected, with each temperature including 1000 temperature values. The periodic boundary condition in the molecular dynamics software GROMACS was: a box size of 12 nm × 12 nm × 12 nm.

[0043] S2 applies a Gaussian bias potential to the lithium-ion coordination in the low-temperature electrolyte structure model to obtain electrolyte dynamic data.

[0044] Using Metadynamics software, a Gaussian bias potential was applied to the lithium-ion coordination in the low-temperature electrolyte structure model to accelerate the sampling efficiency of solvation shell reconstruction at low temperatures and obtain electrolyte dynamic data. In this embodiment, the electrolyte dynamic data includes ion mobility, solvation coordination number (CN), and hydrogen bond density (HBN). For example, at -40°C, the lithium-ion coordination number is 3.2 ± 0.3, and the hydrogen bond density (HBN) decreases by 18% compared to the room-temperature hydrogen bond density (HBN).

[0045] The method for applying a Gaussian bias potential to lithium-ion coordination in a low-temperature electrolyte structure model is as follows: Define a collective variable CV, such as the coordination distance between lithium ions and solvent molecules; add a Metadynamics bias potential; and output the free energy surface and CV trajectory. Run the Metadynamics software; the output file is colvar.dat, which records the collective variable CV, the bias potential change, and the CV trajectory. The bias potential parameters are: Gaussian potential height σ = 0.15 kcal / mol, energy interval ΔE = 0.08 kcal / mol, and the target is Li.+ Coordination distance (2-4 Å).

[0046] Based on the output free energy surface, generate a free energy surface file (free_energy.dat) to confirm whether the energy barrier of the coordination distance has been effectively crossed.

[0047] The solvation coordination number (CN) is obtained by calculating the number of solvent molecules in the first coordination shell of lithium ions through radial distribution function (RDF) integration. The number of solvent molecules in the first coordination shell is directly related to the ion migration energy barrier. Specifically, statistics from the aforementioned colvar.dat file show that... + The average coordination number, such as CN = 3.2 ± 0.3 at -40°C.

[0048] Ion mobility: calculated using mean square displacement (MSD), formula:

[0049]

[0050] in, Phenotypic ion mobility reflects the ability of ions to move in an electrolyte. The mean square displacement is the statistical average of the squared displacements of the ion over time t, where t represents the simulation time. Express Boltzmann constant, T is the thermodynamic temperature. The Dirac function is used to filter ion-solvent pairs at a specific distance.

[0051] Ion mobility was obtained by analyzing the trajectory using the GROMACS tool gmx msd, combined with Einstein relations.

[0052] Hydrogen bond density (HBN) is calculated by counting the number of hydrogen bonds with OH bond lengths, reflecting the strength of intermolecular interactions at low temperatures. A higher HBN indicates a higher electrolyte viscosity. Specifically, the HBN is calculated by counting the number of hydrogen bonds with OH bond lengths <2.5 Å. (Unit: nm) -3 Existing technologies only focus on static structures (such as coordination number). In this embodiment, Metadynamics is used to accelerate sampling and obtain dynamic migration paths at low temperatures, thereby more accurately predicting ion transport performance.

[0053] S3 obtains the molecular diagram vector and non-covalent interaction energy vector of the electrolyte.

[0054] This method transforms the electrolyte molecular structure into molecular graph vectors that can be recognized by machine learning models and calculates the non-covalent interaction energies at low temperatures, overcoming the limitation of existing technologies where static features cannot characterize dynamic molecular interactions. Specifically, the molecular graph vectors are obtained by inputting the SMILES string of the electrolyte components, such as LiPF6: [Li+].[F-][P-](F)(F)(F)(F)F, EC: C1=COC(=O)C1, generating molecular objects; and then decomposing the electrolyte molecules into atomic nodes Li / C / O / F and formal charges (such as Li...). + The process involves identifying +1), hybridization states (e.g., sp² hybridization of C atoms), and bond edges. Atomic numbers are obtained, including both covalent and ionic bonds. Covalent bond types (single / double / ionic), bond lengths (Bond.GetLength), and aromaticity (Bond.GetIsAromatic) are recognized. Virtual edges are added to ion-molecule coordination interactions and marked as coordinate bonds. A 128-dimensional molecular graph vector is generated using the Weisfeiler-Lehman algorithm. Neighborhood information is aggregated three times, encoding atomic charges and bond order information. The output molecular graph vector is a floating-point array representing the structural features of the molecular graph.

[0055] Non-covalent interaction energy includes electrostatic energy at low temperatures. and dispersion energy Non-covalent interactions can be obtained through density functional DFT. When performing density functional calculations, the functional is B3LYP and the basis set is def2-SVP.

[0056] The method for calculating the decomposition of non-covalent interaction energy is as follows:

[0057] Calculate the total energy of isolated molecules such as fluoroethylene carbonate (FEC) and lithium difluorophosphate (LiDFP). and electrostatic energy component .

[0058] Constructing Li + -FEC coordination structure (Li-O spacing 2.2 Å), calculate the total energy of the complex. .

[0059] Electrostatic energy:

[0060] Dispersion energy:

[0061] Use a Python script to iteratively calculate all molecular pairs, such as Li +- Solvent, solvent-solvent, stores the average of the electrostatic and dispersion energies of all molecular pairs, which is used as the non-covalent interaction energy. For example, for every 5% increase in the FEC content of fluoroethylene carbonate, the dispersion energy increases by 3%. The non-covalent interaction energy can affect the solvation stability at low temperatures.

[0062] S4 fuses electrolyte dynamic data, molecular graph vectors, and non-covalent interaction energy vectors into a feature fusion vector.

[0063] The 30-dimensional electrolyte dynamic data, 128-dimensional molecular graph vector, and 10-dimensional non-covalent interaction energy vector are concatenated into a 168-dimensional feature fusion vector. This feature fusion vector is then standardized and input into the low-temperature electrolyte prediction model.

[0064] The standardization method is as follows: Z-score standardization is performed using Scikit-learn's StandardScaler. The output is a standardized vector with a mean of 0 and a variance of 1, which is suitable for the input requirements of the low-temperature electrolyte prediction model.

[0065] Through the above steps, the automated generation of molecular structure and multidimensional features is achieved. Combined with low-temperature specific calculation methods and graph theory algorithms, physically meaningful and multidimensional training data are provided for the low-temperature electrolyte prediction model, significantly improving the accuracy of low-temperature electrolyte performance prediction. Traditional feature engineering neglects intermolecular non-covalent interactions; this scheme improves the accuracy of solvation stability prediction by quantifying electrostatic and dispersion energies using DFT.

[0066] S5 establishes a low-temperature electrolyte prediction model based on a graph neural network model, and trains the low-temperature electrolyte prediction model through feature fusion vectors.

[0067] like Figure 2 As shown, a 168-dimensional feature fusion vector is received. This feature fusion vector includes point-level features and bond-level features. The point-level features include solvation coordination number CN, hydrogen bond density HBN, and atomic charge. The bond-level features include bond length and electrostatic energy. and dispersion energy .

[0068] The low-temperature electrolyte prediction model includes an input layer, a convolutional layer, a pooling layer, and an output layer. The convolutional layer extracts local structural features using the GCNConv operator.

[0069] The input layer receives a 168-dimensional feature fusion vector, including electrolyte dynamics data, molecular graph vectors, and non-covalent interaction energies.

[0070] Convolutional Layers: PyTorch Geometric operators are used to extract local graph structural features. The first layer is expanded to a 256-dimensional hidden layer, and the second layer is compressed to 128 dimensions to reduce computational complexity. The hidden layer dimension is 256, and the activation function is ReLU. A self-attention module is used to increase the weight of fluorinated groups (such as -CF2- in FEC) while suppressing long-chain alkyl groups, focusing on key low-temperature interactions. In this embodiment, the weight of fluorinated groups is increased to 25%, while the weight of long-chain alkyl groups is suppressed to 15%. A temperature weight factor is introduced into the self-attention module, assigning higher weights (weight coefficient 1.5 times) to samples from -40°C to -20°C, enhancing learning in extreme low-temperature scenarios. The validation set MSE is reduced by 40% compared to the basic approach.

[0071] Pooling layers are used to reduce the dimensionality of local structural features. Top-K pooling (K=0.5) retains key nodes and reduces computational complexity.

[0072] The output layer is used to output key properties of the electrolyte, including conductivity σ at -40°C, viscosity η at -30°C, and freezing point T. mp For example, the predicted formula of a certain electrolyte is: σ=0.62 mS / cm (-40°C), η=28 mPa·s (-30°C), T_mp=-48°C.

[0073] The weighted mean square error (MSE) of the loss function of the output layer is:

[0074] ,

[0075] The Loss expression represents the total loss function, which comprehensively measures the prediction errors of conductivity, viscosity, and freezing point. It is obtained through weighted summation. The mean squared error (MSE) of conductivity, viscosity and freezing point are expressed separately. 0.5, 0.3 and 0.2 are the weightings of conductivity, viscosity and freezing point, respectively, reflecting the differences in the importance of the indicators. That is, conductivity is the core indicator of low temperature battery and has the highest weight. Viscosity and freezing point have the next lower influence and their weights decrease.

[0076] in, N is the quantity of electrolyte formulation involved in the loss calculation. and The predicted and actual conductivity values ​​for the i-th sample are presented separately. The formulas for calculating the mean square error of viscosity and freezing point are the same as those for conductivity. Conductivity is a key indicator, with a weighting of 50%.

[0077] The method for training the low-temperature electrolyte prediction model is as follows: Based on the Uni-mol open-source model, 10,000 sets of low-temperature simulation data are used to pre-train the low-temperature electrolyte prediction model, initializing the parameters of the low-temperature electrolyte prediction model to improve the low-temperature feature learning efficiency by 30%. In this embodiment, the parameters for training the low-temperature electrolyte prediction model are: learning rate 1e... -4 Adam optimizer, batch size 64, training 100 epochs.

[0078] The low-temperature electrolyte prediction model is trained using feature fusion vectors. Several electrolyte formulations with the largest weighted mean square error (MSE) are selected; in this embodiment, formulations with the top 10% prediction variance (e.g., prediction variance > 0.05 mS / cm) are chosen. Molecular dynamics simulations are performed based on these electrolyte formulations (e.g., LiPF6 + VC 2%) to expand the training set by adding 2000 new data sets. The expanded training set is then input into the low-temperature electrolyte prediction model for training until the loss function of the output layer is less than a threshold, at which point the trained low-temperature electrolyte prediction model is output. This embodiment introduces an expanded training set to reduce reliance on labeled data. For example, if the predicted conductivity variance of a certain electrolyte formulation is 0.08 mS / cm, triggering molecular dynamics simulations reduces the prediction error from 15% to 7%.

[0079] In this embodiment, FP16 mixed-precision computing is used to train the low-temperature electrolyte prediction model, which reduces memory usage while maintaining accuracy, supports larger batch training (batch_size increased from 64 to 128), and improves training speed by 50%.

[0080] S6 inputs the candidate electrolyte formulations into the trained low-temperature electrolyte prediction model to generate the key performance characteristics of the electrolyte.

[0081] like Figure 3 As shown, the results of the low-temperature electrolyte prediction model are subjected to low-temperature constraint filtering. The conditions for low-temperature constraint filtering are:

[0082]

[0083] in, for conductivity, for viscosity; The freezing point is used; the PyTorch DDP framework is adopted, and parallel evaluation is performed through a GPU cluster. The screening time for thousands of formulations is less than 30 minutes, which is 800% more efficient than a single card. For example, the formulation prediction result of an electrolyte is σ=0.48 mS / cm, which does not meet the screening conditions of low-temperature constraint filtration, so the prediction result is directly filtered.

[0084] The parameters of the molecular dynamics simulation are automatically calibrated based on the results filtered by low-temperature constraints. For example, the LJ well depth ε of the FEC molecule is adjusted from 0.35 kcal / mol to 0.38 kcal / mol, forming a closed loop of simulation-experiment-correction data. After low-temperature screening, the error is reduced by 30%, making the low-temperature electrolyte prediction model more adaptable to extreme low-temperature scenarios. For instance, the measured conductivity of a certain formulation's electrolyte at -40°C is σ = 0.55 mS / cm. By reverse-calibrating the Lennard-Jones potential parameter ε of the FEC molecule, changing ε from 0.35 kcal / mol to 0.38 kcal / mol, the prediction error of the low-temperature electrolyte prediction model is reduced from 12% to 6%.

[0085] The conductivity of the electrolyte is obtained based on a low-temperature electrolyte prediction model. A conductivity-time curve is established, and the conductivity temperature coefficient Δσ / ΔT is obtained through linear fitting based on this curve. For example, in a certain formulation, Δσ / ΔT = 0.02 mS / (cm·°C). The results of the low-temperature electrolyte prediction model are then input into the Eyring equation to obtain the viscosity activation energy. ,For example =15 kJ / mol, to quantify the sensitivity of key electrolyte properties to temperature changes at low temperatures, and enhance the ability of low-temperature electrolyte prediction models to capture nonlinear relationships. For example, the temperature dependence of electrolyte conductivity σ in the range of -40°C to 0°C was reduced by 20% in modeling error.

[0086] From 10,000 formulations, 300 were selected (e.g., LiFSI 1.0M + VC 3%). The screening process took 20 minutes. Molecular dynamics calculations (20 ns) were performed on the results of the low-temperature electrolyte prediction model to calculate the solvation sheath stability (SCS, standard deviation <0.5). For example, the SCS of a certain electrolyte formulation was 0.45 (meeting the standard). Based on the solvation sheath stability, the results of the low-temperature electrolyte prediction model were further screened, and finally 20 high-stability electrolyte formulations were selected. This second screening improved the experimental verification pass rate by more than 80%.

[0087] Example 2

[0088] Based on the same inventive concept, this embodiment discloses a low-temperature electrolyte performance prediction system based on cross-scale collaborative optimization, comprising:

[0089] The molecular dynamics simulation module is used to obtain the structure model of the low-temperature electrolyte through molecular dynamics simulation.

[0090] The electrolyte dynamic data acquisition module is used to apply a Gaussian bias potential to the lithium ion coordination in the low-temperature electrolyte structure model to obtain electrolyte dynamic data.

[0091] The molecular diagram energy vector acquisition module is used to acquire the molecular diagram vector and non-covalent interaction energy vector of the electrolyte.

[0092] The feature fusion module is used to fuse electrolyte dynamic data, molecular diagram vectors, and non-covalent interaction energy vectors into a feature fusion vector;

[0093] The model training module is used to establish a low-temperature electrolyte prediction model based on the graph neural network model, and to train the low-temperature electrolyte prediction model through feature fusion vectors.

[0094] The model output module is used to input candidate electrolyte formulations into the trained low-temperature electrolyte prediction model to generate the key performance characteristics of the electrolyte.

[0095] 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.

[0096] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] 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.

[0098] 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.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific embodiments of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention. The above content is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims

1. A method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization, characterized in that, Includes the following steps: A structural model of the low-temperature electrolyte was obtained through molecular dynamics simulations. A Gaussian bias potential is applied to the lithium ion coordination in the low-temperature electrolyte structure model to obtain electrolyte dynamic data; Obtain the molecular diagram vector and non-covalent interaction energy vector of the electrolyte; The electrolyte dynamic data, molecular graph vector, and non-covalent interaction energy vector are fused into a feature fusion vector. A low-temperature electrolyte prediction model is established based on the graph neural network model, and the feature fusion vector is used to train the low-temperature electrolyte prediction model. The candidate electrolyte formulations are input into the trained low-temperature electrolyte prediction model to generate the key performance characteristics of the electrolyte.

2. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in claim 1, characterized in that, The low-temperature electrolyte structure model includes the electrolyte molecular structure at several different temperature points, which are in the range of -40°C to 0°C. The electrolyte dynamic data includes ion mobility, solvation coordination number CN, and hydrogen bond density HBN. The solvation coordination number CN is obtained by calculating the number of solvent molecules in the first coordination shell of lithium ions through radial distribution function (RDF) integration. The hydrogen bond density HBN is obtained by calculating the hydrogen bond density per unit volume by statistically counting the number of hydrogen bonds with OH bond lengths.

3. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in claim 1, characterized in that, The method for obtaining the molecular graph vector is as follows: the electrolyte molecules are decomposed into atomic nodes Li / C / O / F and bond edges. Based on the atomic nodes Li / C / O / F and bond edges, a molecular graph vector is generated using the Weisfeiler-Lehman algorithm, and atomic charge and bond order information are encoded in the molecular graph vector. The non-covalent interaction energy includes electrostatic energy and dispersion energy at low temperature. The non-covalent interaction energy is obtained through density functional theory.

4. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in claim 1, characterized in that, The feature fusion vector includes point-level features and bond-level features. The low-temperature electrolyte prediction model includes a convolutional layer, a pooling layer, and an output layer. The convolutional layer extracts local structural features using the GCNConv operator; it increases the weight of fluorinated groups while suppressing long-chain alkyl groups using a self-attention module; the pooling layer reduces the dimensionality of the local structural features, retains key nodes, and reduces computational complexity; the output layer outputs the key performance characteristics of the electrolyte, including conductivity, viscosity, and freezing point.

5. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in claim 4, characterized in that, The weighted mean square error (MSE) of the loss function of the output layer is: ， The Loss expression represents the total loss function, which comprehensively measures the prediction errors of conductivity, viscosity, and freezing point. It is obtained through weighted summation. The mean squared error (MSE) of conductivity, viscosity and freezing point are expressed separately. 0.5, 0.3 and 0.2 are the weightings of conductivity, viscosity and freezing point, respectively, reflecting the differences in the importance of the indicators. That is, conductivity is the core indicator of low temperature battery and has the highest weight. Viscosity and freezing point have the next least influence and their weights decrease in that order. in, N is the quantity of electrolyte formulation involved in the loss calculation. and The predicted and actual conductivity values ​​for the i-th sample are expressed separately. The formulas for calculating the mean square error of viscosity and freezing point are the same as those for conductivity.

6. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in claim 5, characterized in that, The method for training the low-temperature electrolyte prediction model is as follows: based on the Uni-mol open-source model, using low-temperature simulation data, the low-temperature electrolyte prediction model is pre-trained, and the parameters of the low-temperature electrolyte prediction model are initialized; the feature fusion vector is used to train the low-temperature electrolyte prediction model; several electrolyte formulations with the largest weighted mean square error (MSE) are selected, and molecular dynamics simulations are performed based on the electrolyte formulations to expand the training set data; The expanded training set is input into the low-temperature electrolyte prediction model for training until the value of the loss function of the output layer is less than the threshold, and then the trained low-temperature electrolyte prediction model is output.

7. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in any one of claims 1-6, characterized in that, The results of the low-temperature electrolyte prediction model are subjected to low-temperature constraint filtering, and the conditions for the low-temperature constraint filtering are as follows: ， in, for conductivity, for viscosity; It is the freezing point; Based on the results of low-temperature constraint filtering, the parameters of the molecular dynamics simulation are automatically calibrated.

8. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in any one of claims 1-6, characterized in that, The conductivity value of the electrolyte is obtained based on the low-temperature electrolyte prediction model, and a conductivity value versus time curve is established. The conductivity temperature coefficient is obtained based on the conductivity value versus time curve. The results of the low-temperature electrolyte prediction model are input into the Eyring equation to obtain the viscosity activation energy of the electrolyte.

9. The method for predicting the performance of low-temperature electrolytes based on cross-scale collaborative optimization as described in any one of claims 1-6, characterized in that, Molecular dynamics calculations were performed on the results of the low-temperature electrolyte prediction model to calculate the stability of the solvated sheath layer; The results of the low-temperature electrolyte prediction model are further screened based on the stability of the solvated sheath.

10. A low-temperature electrolyte performance prediction system based on cross-scale collaborative optimization, characterized in that, include: The molecular dynamics simulation module is used to obtain the structure model of the low-temperature electrolyte through molecular dynamics simulation. The electrolyte dynamic data acquisition module is used to apply a Gaussian bias potential to the lithium ion coordination in the low-temperature electrolyte structure model to obtain electrolyte dynamic data. The molecular diagram energy vector acquisition module is used to acquire the molecular diagram vector and non-covalent interaction energy vector of the electrolyte. The feature fusion module is used to fuse the electrolyte dynamic data, molecular diagram vector, and non-covalent interaction energy vector into a feature fusion vector; The model training module is used to establish a low-temperature electrolyte prediction model based on the graph neural network model, and to train the low-temperature electrolyte prediction model with the feature fusion vector. The model output module is used to input candidate electrolyte formulations into the trained low-temperature electrolyte prediction model to generate the key performance characteristics of the electrolyte.

Images