Artificial Intelligence-Based Optimization System and Method for Liquid Additives in Lithium Metal Batteries

By optimizing the system with artificial intelligence, the high-throughput generation and real-time dynamic control of liquid additives are achieved, which solves the problems of long R&D cycle and insufficient performance of lithium metal batteries, improves the cycle life and safety of batteries, and is applicable to a variety of high energy density batteries.

CN123120127APending Publication Date: 2026-11-17ZHEJIANG FENGLI TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610987058.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively address the challenges of long development cycles, vast chemical spaces, complex deep performance coupling, inefficient cross-scale data utilization, and lack of dynamic interface control capabilities in lithium metal batteries, leading to insufficient cycle life and safety performance.

Method used

By employing an AI-based multi-scale prediction model, a deep active learning inverse generation framework, and a real-time closed-loop system, combined with voltage curve image encoding and deep reinforcement learning, we can achieve ultra-high throughput generation of liquid additive formulations, accurate performance prediction, and synthesis path planning, and dynamically control the formulation and charging strategies during battery operation.

Benefits of technology

It significantly shortens the R&D cycle, improves the cycle life and ionic conductivity of lithium metal batteries, enhances safety performance, and the system has cross-system scalability, making it suitable for the R&D of various high-energy-density battery electrolyte systems.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

This invention provides an artificial intelligence-based optimization system and method for liquid additives in lithium metal batteries. The system includes a data management module that integrates and standardizes literature, patent data, and multi-scale physical simulation data; a multi-scale prediction model module that uses graph neural networks to encode molecules and combines substitution-invariant attention networks to predict properties such as ionic conductivity and lifetime, while embedding physical information constraints; a reverse generation and screening module that generates high-throughput formulations through a conditional diffusion model and selects samples using Bayesian active learning; a synthesis path planning module that recommends mixing process parameters; an experimental execution and testing module that prepares and assembles batteries at high throughput, using in-situ characterization and intelligent identification of the state using 3D RGB images of voltage curves; and a closed-loop optimization module that incrementally updates the model based on feedback data and invokes reinforcement learning to implement online dynamic control. This invention achieves efficient optimization and cross-system transfer of battery liquid additives, significantly improving cycle life and safety.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to lithium battery technology, and specifically to an artificial intelligence-based closed-loop optimization system and method for lithium metal battery electrolytes. Background Technology

[0002] Lithium metal batteries have become a core solution for next-generation high-energy-density energy storage systems due to their extremely high theoretical specific capacity (3860 mAh / g). However, problems such as lithium dendrite growth, side reactions at the interface between liquid additives and electrodes, and poor stability of the solid electrolyte interphase (SEI) film severely limit their cycle life and safety performance. In hybrid solid-liquid lithium metal batteries, liquid additives, as the "blood" regulating interface characteristics, directly determine the quality of the SEI film and lithium deposition behavior through their solvation structure, ionic conductivity, and redox stability, and are key to solving the aforementioned bottlenecks.

[0003] Traditional liquid additive development mainly relies on a "trial and error" approach, which has the following core challenges: The chemical space is extremely vast: liquid additives involve multi-component complexes such as solvents, lithium salts, and functional additives, and the basic molecular library contains hundreds of thousands of candidate combinations. Traditional high-throughput experiments cannot fully cover the effective chemical space. Deep coupling of multidimensional performance: There are strong nonlinear and non-equilibrium characteristics between "composition-structure-solventization behavior-SEI characteristics-battery performance", making it difficult to manually uncover the deep patterns, resulting in a research and development cycle that is usually as long as 18 to 24 months. Inefficient use of cross-scale data: Experimental, multi-scale simulation and literature data lack standardized integration, machine learning models are prone to overfitting, and it is difficult to achieve closed-loop optimization of "prediction-experiment-feedback-iteration"; Lack of dynamic interface control capability: Existing R&D systems are mostly based on static or offline prediction, which cannot be combined with the dynamic evolution of the interface during battery charging and discharging, making it difficult to achieve synergistic optimization of electrolyte formulation and real-time charging strategy.

[0004] While existing technologies employ machine learning-based electrolyte / additive screening methods, they still suffer from several shortcomings: First, they fail to fully integrate prior knowledge from physics such as quantum chemistry and molecular dynamics, resulting in prediction accuracy highly limited by the size of the known dataset. Second, they lack a complete closed-loop integration encompassing "molecular generation - synthetic pathway prediction - performance verification - automatic model updating." Third, they fail to achieve dynamic adaptation between liquid additive formulations and the real-time operating status of the battery. Therefore, current screening mechanisms struggle to support the stringent requirements of hybrid solid-liquid lithium metal batteries, which simultaneously demand high energy density and long-cycle stability. Summary of the Invention

[0005] In view of the shortcomings of the prior art, this invention provides an optimization system and method for liquid additives for lithium metal batteries based on artificial intelligence. By using a multi-scale prediction model embedded with physical information, a deep active learning reverse generation framework, and a real-time closed-loop system of experiment-model, it achieves ultra-high throughput reverse generation of liquid additive formulations, accurate performance prediction, and automatic planning of synthesis paths. At the same time, by introducing a state recognition mechanism based on voltage curve image encoding and a deep reinforcement learning mechanism, it realizes the coordinated control of formulation and charging strategy during the dynamic operation of the battery, which significantly shortens the research and development cycle and significantly improves the cycle life, ionic conductivity, and safety performance of lithium metal batteries.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An AI-based optimization system for liquid additives in lithium metal batteries, comprising: The data management module is used to collect and standardize multi-source data, constructing a standardized database containing mapping relationships between composition, structure, solvation, and performance. This multi-source data includes: ① molecular structure and performance data from literature / patents (≥240,000 single molecules, ≥10,000 mixed formulations); ② multi-scale simulation calculation data, such as quantum chemical calculations and molecular dynamics simulation data (≥100,000 sets); ③ experimental test data, such as conductivity, cycle life, and SEI composition; ④ real-time battery operation data, such as voltage curves, temperature, and capacity decay rates. After cleaning and normalization, this multi-source data is used to construct a standardized database containing mapping relationships between composition, structure, solvation, and performance, and the data confidence levels are labeled.

[0007] Multi-scale prediction model module: used to build prediction models with embedded physical information, integrate physical constraints at different scales, and predict key performance indicators and microscopic properties of liquid additives. Deep active learning inverse generation module: Used to construct an inverse generation framework based on diffusion model with battery performance target as constraint, generate candidate liquid additive formulations, and combine active learning strategy to screen priority experimental samples from the candidate liquid additive formulations. Synthesis Path Planning Module: Based on the reaction prediction model, it performs reverse synthesis path planning for priority experimental samples and outputs the corresponding synthesis process parameters; Experiment execution and real-time detection module: used to prepare liquid additives and assemble batteries according to synthesis process parameters, conduct electrochemical performance tests on batteries to obtain experimental data, and obtain microscopic characteristic data and macroscopic operating status through in-situ characterization and voltage curve analysis; the experiment execution and real-time detection module uses voltage curve three-dimensional imaging technology to encode the charge and discharge data within a set number of cycles into RGB state images, and classifies the battery operating status into stable state, transition state or decay state by recognizing RGB state images.

[0008] Closed-loop optimization module: used to feed back the experimental data and microscopic characteristic data obtained by the experimental execution and real-time detection module to the data management module, so as to dynamically update the multi-scale prediction model module and dynamically adjust the liquid additive formula or battery charging strategy for the identified abnormal battery operating states.

[0009] As a preferred technical solution, the multi-scale prediction model module includes a molecular characterization sub-model, a mixture prediction sub-model, and an interpretability analysis sub-model: Molecular characterization sub-model: A graph neural network (GNN) is used to encode the discrete graph topological structure of solvent, lithium salt, and additive molecules in the electrolyte system, and embeds physical feature vectors containing molecular flexibility and functional group activity. Specifically, the molecule is represented as a graph G=(V,E), with atoms as nodes V and chemical bonds as edges E. A message-passing mechanism is used to aggregate neighborhood atomic features, and physical prior feature vectors are embedded in the final fully connected molecular layer. These physical prior feature vectors include the number of flexible rotational bonds, van der Waals volume, dipole moment, and the highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) energy levels.

[0010] The mixture prediction sub-model employs a permutation-invariant network embedded with physical information. Since the liquid additive formulation consists of disordered blends of multiple components, the network eliminates the influence of input order on the prediction results by weighted summation of the GNN feature vectors of each independent molecule based on an attention mechanism. Its input consists of the structural topological features of each component molecule and their corresponding mass / molar ratio. The output simultaneously predicts the macroscopic and microscopic properties of the mixed liquid additive. Macroscopic properties include ionic conductivity, oxidation potential, and thermal decomposition temperature; microscopic properties include the structure of the first solvation shell and the inorganic / organic component ratio of the SEI film.

[0011] Physical information embedding constraint mechanism: The model forcibly embeds multi-scale physical equation constraints into the loss function. Specifically, the total loss function, Loss, is a data-driven loss. d Loss with physical constraint penalty p Weighted composition: Loss = Loss d +λ1×Loss p ; Among them, the physical constraint penalty term Loss p The fundamental laws of electrochemistry and quantum chemistry must be enforced, including: ① the Arrhenius equation constraint on the change of ionic conductivity with temperature; ② the energy conservation boundary constraint between the electrochemical stability window and the HOMO / LUMO energy levels of each component molecule.

[0012] Interpretability analysis sub-model: The first derivative of the output index with respect to the input molecular descriptor is calculated through the gradient decoupling mechanism, and combined with the symbolic regression algorithm, the complex neural network black box is explicitly mapped into a closed algebraic equation, and the dominant molecular descriptors affecting the overall performance of liquid additives, such as fluorine content, dielectric constant, and ion-solvent binding energy, are identified.

[0013] The deep active learning inverse generation module includes a candidate space construction unit, an inverse generation unit, and an active learning filtering unit: Candidate space construction unit: Integrating external PubChem and eMolecules standard chemical databases, a candidate space containing at least 1 million active molecular structures is constructed. The solvents are defined as at least one of ethers, carbonates, ketones, and sulfones, or mixtures thereof; the lithium salts are at least one of LiPF6, LiFSI, LiTFSI, and LiClO4, or mixtures thereof; the additives are at least one of LiNO3, fluoroethylene carbonate (FEC), thiocarbonates, and small molecule additives containing nitrogen or sulfur functional groups.

[0014] Inverse generation unit based on conditional diffusion model: The model uses the set battery performance targets, such as cycle life of 1000 cycles and ionic conductivity ≥10mS / cm, as the conditional embedding vector C.

[0015] To address the complex characteristics of formulations that combine "discrete molecular structure" and "continuous proportions," a dual-channel forward diffusion and reverse denoising logic is established: ① The molecular diagram topology channel adopts discrete diffusion: In forward diffusion, random mask noise is gradually added to the atomic categories C, O, F, S, P, N and chemical bond types such as single bond, double bond, and aromatic bond through discrete Markov transfer matrix until the structure becomes a disordered uniform distribution diagram.

[0016] ② The proportioning channel uses continuous Gaussian diffusion: the proportioning vector X of solvent, lithium salt, and functional additives is subjected to Markov Gaussian diffusion, and Gaussian noise is gradually introduced according to the following formula: ; Where, β t These are timing noise scheduling parameters.

[0017] ③ Conditional Inverse Denoising Generation: During inverse generation, the input is a completely random, unordered graph structure and a randomized allocation vector. The denoising network (a temporal graph neural network combined with a multilayer perceptron) is guided by the conditional vector C to progressively predict and remove noise estimates є. θ (X) t Through multi-step reverse iteration, the SMILES string or adjacency matrix of the novel molecule and the corresponding component molar ratio are reconstructed and generated.

[0018] Active learning screening unit: Using Gaussian process regression combined with a Bayesian model averaging algorithm, the information gain and performance potential of the candidate liquid additive formulations are calculated. Samples whose information gain and performance potential meet preset thresholds are selected as priority experimental samples. The mean performance μ(x) and prediction uncertainty (variance σ(x)) of all inversely generated candidate formulations under the current multi-scale prediction model are calculated. An acquisition function α is constructed using the confidence upper bound algorithm. UCB (x) = μ(x) + к×σ(x); where к is the weighting coefficient for balancing optimization and exploration. The 20 to 50 formula samples with the highest acquisition function scores are selected as priority experimental samples to achieve a dual selection process of "maximizing information gain" and "maximizing performance potential".

[0019] The synthesis path planning module is based on a reaction prediction model, specifically employing a dual-drive architecture based on graph attention network reaction center localization and Transformer sequence inverse translation: Reaction center prediction: Input the graph topology of the novel additive molecule output by the reverse generation module, use graph attention network to calculate the electron cloud density and chemical bond breaking probability between each atomic pair in the molecule, locate the bond breaking site (reaction center) of the reverse synthesis, match it with the standard chemical reaction template library, and disassemble it into a first-order reaction intermediate.

[0020] Sequence Reverse Translation and Process Output: The intermediate is converted into a sequence of SMILES text strings and input into a Transformer encoder-decoder network pre-trained on a massive standard chemical reaction dataset (such as USPTO). The encoder captures long-range atomic dependencies, and the decoder autoregressively "translates" the target molecule string into a SMILES sequence of commercially available basic reaction raw materials. Simultaneously, a multi-task classification and regression network head is connected to the decoder. Based on the matched chemical reaction type, such as nucleophilic substitution, electrophilic addition, and esterification, the network automatically calculates and synchronously outputs the corresponding synthetic process parameters, including: the optimal reaction temperature range, specific catalyst type, reaction solvent medium, and the reaction stirring time required to achieve the preset yield.

[0021] The experimental execution and real-time detection module integrates automated high-throughput liquid additive preparation equipment, electrochemical testing station and in-situ characterization equipment, and automatically completes the liquid additive preparation, soft pack / button battery assembly and charge-discharge cycle testing of priority experimental samples.

[0022] In-situ characterization equipment includes at least one of in-situ Raman spectroscopy and in-situ transmission electron microscopy, used to capture the dynamic evolution of lithium ion deposition / stripping morphology on electrode surface in real time during battery charging and discharging, and to quantitatively detect the spatial distribution and growth kinetics of chemical components of microscopic inorganic substances such as LiF and Li2O and organic substances such as alkyl lithium carbonate in the solid electrolyte interface film.

[0023] During battery charge-discharge cycles, the data acquisition system captures real-time "voltage-capacity-cycle count" time-series data for each cycle. Every preset number of cycles, preferably every 3 cycles, the corresponding charge-discharge voltage curves are time-series stacked and mapped in three-dimensional space along the time axis / cycle count axis. After normalization, this is converted into a three-dimensional matrix image containing two-dimensional pixel coordinates and RGB color channel features. The geometric texture and color gradient changes in the image directly correspond to the battery's internal polarization voltage growth rate, capacity decay slope, and lithium dendrite micro-short circuit characteristics. A convolutional neural network classifier then classifies the current battery operating state in real-time into a stable state, transitional state, or decay state.

[0024] The closed-loop optimization module feeds back the actual electrochemical test data, such as measured conductivity, capacity retention, and in-situ characterization of microstructure data, obtained by the experimental execution and real-time detection modules, weighted by confidence level, to the data management module. A mini-batch gradient descent algorithm is used to incrementally and supervisedly fine-tune the parameters of the top fully connected layer of the multi-scale prediction model module, achieving closed-loop adaptive iterative updates of the prediction model.

[0025] Online Real-Time Regulation Based on Deep Reinforcement Learning: For abnormal battery operating states identified as "decaying states," the system constructs a control loop based on a deep Q-network. The environmental states are the current battery voltage polarization, temperature, and capacity decay rate, with the bounded reward function being maintaining battery capacity retention and suppressing side reactions. The DQN network solves the optimal compensation strategy for the current state online in real-time, including: dynamically increasing the mass percentage of LiNO3 additive in the current battery system via microfluidic valves, or sending commands at the control terminal to reduce the charging current rate and simultaneously lower the cutoff voltage to prevent battery degradation and achieve real-time dynamic closed-loop optimization.

[0026] An optimization method for liquid additives in lithium metal batteries based on artificial intelligence, implemented based on the aforementioned optimization system, the method includes the following steps: S1. Data Standardization and Model Initialization: Collect multi-source data through the data management module and perform cleaning, normalization and confidence labeling to build a standardized database; Based on the initial database, use the physical information loss function that introduces first-principles energy level boundaries and Arrhenius equation constraints to train the multi-scale prediction model module and set the model baseline error. S2. Performance Targets and Constraints: Set the target performance targets and engineering application constraints for the liquid additive, and limit the initial concentration and molar ratio range of each component of the liquid additive, such as solvent, lithium salt, and functional additives. S3. Reverse generation and active learning screening: The target performance objective is input into the conditional diffusion model as a conditional embedding vector. A high-throughput candidate recipe array (≥1000 groups) is generated through dual-channel discrete / continuous denoising iterative reconstruction. The active learning screening unit is used to calculate the UCB acquisition function score of each candidate recipe based on Gaussian process regression, and 20 to 50 groups of high information gain and high potential priority experimental samples are selected from them. S4. Synthesis Path Planning and Experiment Execution: The Transformer inverse translation model of the synthesis path planning module is used to decompose the precursor raw materials of the priority experimental samples and output synthesis process parameters including temperature, catalyst, and stirring time. The experiment execution and real-time detection module automatically completes the synthesis of liquid additives, battery assembly, electrochemical charge-discharge testing, and in-situ Raman / TEM microscopic characterization. Simultaneously, the real-time operating state (stable state / transition state / degradation state) of the battery is obtained through voltage curve three-dimensional imaging technology. S5. Real-time closed-loop optimization: The acquired macroscopic performance data and microscopic characteristic data are fed back to the standardized database in a weighted manner according to confidence level using the closed-loop optimization module, and the multi-scale prediction model is updated incrementally. At the same time, if the battery is detected to be in a decay state, the DQN deep reinforcement learning network is called to execute the online closed-loop control strategy, and the proportion of specific liquid additives (such as LiNO3) is adjusted by microfluidics, or the charging rate and cutoff voltage are dynamically adjusted. S6. Optimal formulation output: Iterate through steps S3 to S5, usually for 3-6 rounds, until the optimal liquid additive formulation that fully meets all the target performance objectives is selected. Finally, output the corresponding molecular topology, molar ratio, synthesis route process parameters and battery performance verification report.

[0027] The beneficial effects of this invention are: 1. This invention replaces the traditional blind trial-and-error method with "diffusion model reverse generation + active learning Bayesian screening". Through deep correlation between macroscopic performance and microscopic chemical characteristics, and nonlinear selection of high-throughput candidate formulation arrays and priority experimental samples, the R&D cycle of liquid additives is significantly shortened from the original 18-24 months to 4-6 months, and the number of experimental trials is reduced by more than 70%.

[0028] 2. This invention successfully integrates the microscopic physical constraints of first principles, the mesoscopic solvation structural features, and macroscopic electrochemical testing by constructing a joint network with embedded physical information. This overcomes the excessive reliance of traditional AI on data volume and still has extremely high prediction accuracy with small sample datasets. By using symbolic regression, the interpretability of the mechanism is realized, which can accurately guide the directional design of formulations with high ionic conductivity and strong SEI film stability.

[0029] 3. This invention breaks through the limitation of existing AI systems that can only perform "offline static prediction". This invention innovatively adopts voltage curve RGB state image encoding technology, which can sensitively capture the transition state and decay state of the battery during cycling. During operation, it uses reinforcement learning to dynamically compensate the formula (such as adjusting the LiNO3 ratio) or co-optimize the charging strategy to block dendrite growth, thereby improving the cycle life of lithium metal batteries by ≥200%.

[0030] 4. The synthesis path planning module of the system of this invention can directly output the retrosynthesis process parameters in an end-to-end manner, seamlessly connecting to the automated production line; the system architecture has strong cross-system scalability, and is not only suitable for hybrid solid-liquid lithium metal batteries, all-solid-state or negative electrode-free lithium metal batteries, but can also be horizontally extended to the research and development of multiple next-generation high-energy-density battery electrolyte systems such as sodium-ion and potassium-ion. Detailed Implementation

[0031] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0032] Example 1: This example is an optimized example of LiNO3 additive generated by AI reverse engineering.

[0033] The purpose of this embodiment is to achieve precise reverse design of efficient LiNO3 interface stabilizer formulations in liquid electrolyte systems and to efficiently optimize high ionic conductivity and long cycle life.

[0034] The data management module performs standardized integration of multi-source data: The data management module extracts the topological structures of 324,000 organic small molecule monomers and the electrochemical characteristics of 12,500 multi-component mixed liquid formulations through a standard text mining engine, including initial coulombic efficiency, oxidation potential, and electrochemical window.

[0035] Multiscale physics simulation data: Calculations were performed based on first-principles calculations. Using VASP software with a generalized gradient approximation (GGA-PBE) functional, the geometry of LiNO3, LiFSI, and ether solvent molecules was optimized. The solvation binding energy and the lowest unoccupied molecular orbital (LUMO) energy level were calculated, generating a dataset of 121,000 quantum chemical features. For molecular dynamics (MD) simulations, GROMACS software (using an OPLS-AA force field) was used to construct a mixed solvent-lithium salt local coordination system containing 2000 molecules. After a 50 ns relaxation, the three-dimensional free diffusion coefficient and radial distribution function (RDF) feature data of lithium ions under specific ratios were output.

[0036] The measured ionic conductivity was normalized to the continuous interval [0,1] using a maximum-minimum mapping. The cycle life was linearly normalized and labeled with a baseline of 1000 cycles. The confidence weight coefficients of the experimental data were also set.

[0037] The network architecture of the multi-scale prediction model module includes a molecular characterization sub-model and a mixture prediction sub-model. The molecular characterization sub-model uses a graph isomorphic network to transform 1,2-dimethoxyethane, 1,3-dioxopentane, and LiNO3 molecules into molecular graph topological embedding vectors with 256 feature channels, capturing the degrees of freedom of rotational single bonds and the local charge density of nitro functional groups. The mixture prediction sub-model constructs a permutation-invariant network based on a self-attention mechanism. The network input consists of the feature vector matrix of each independent molecule and the corresponding molar ratio scalar. A multi-head attention layer aggregates the disordered interaction features between multiple components. The output is connected to a multi-task regression head, simultaneously outputting the predicted ionic conductivity, electrochemical oxidation decomposition potential, and interfacial lithium deposition overpotential.

[0038] The model's loss function incorporates a physical boundary penalty term. Specifically, for predicting the change of ionic conductivity with temperature, the Arrhenius empirical equation is introduced as a constraint: ;

[0039] If the conductivity predicted by the neural network at each temperature (25℃, 50℃, 85℃) does not satisfy the monotonicity of the nonlinear function and the activation energy boundary, then a loss of 10 is given. 3 The high penalty weights are used to constrain the model's predictive extrapolation ability under small sample sizes.

[0040] Mechanistic decoupling analysis: By decoupling the deep weights of the network through symbolic regression, an explicit formula showing the positive correlation between the molar concentration of LiNO3 and the arrangement density of the first solvation shell of lithium salt anions was extracted, confirming that LiNO3 can preferentially participate in the reduction reaction and reconstruct the SEI film composition.

[0041] The deep active learning reverse generation and sample selection part includes the following: Formulation design boundary constraints: The chemical element space for reverse generation is limited to C, H, O, N, F, S, and Li. The main solvent system is limited to DME and DOL, the main salt is LiPF6 or LiFSI, and the core targeted optimized functional additives are LiNO3 and fluoroethylene carbonate. The molar ratio of DME and DOL is limited to 2:1 to 4:1, and the mass percentage of LiNO3 is limited to 0.5% to 5.0%. Dual-channel conditional diffusion model generation: The targeted performance condition vector C = [cycle lifetime ≥ 1000 cycles, ionic conductivity ≥ 10 mS / cm] is set as the conditional control input. After a 1000-step reverse Markov denoising process, the model outputs the molar percentage of components in the continuous ratio channel and generates a flux candidate formulation array consisting of 1200 non-repeating formulations in the discrete graph topology channel through reverse reconstruction. The performance prediction mean μ(x) and uncertainty variance σ(x) of these 1200 formulations are calculated using Gaussian process regression. Use the upper confidence bound (UCB) to obtain the function α UCB The information gain of each candidate sample is evaluated using μ(x) = μ(x) + 2.5 × σ(x). Finally, the top 20 formula samples with the highest acquisition function scores are automatically selected as the priority experimental sample queue.

[0042] Synthesis path planning and high-throughput experiment execution are as follows: Reaction Path Planning: The synthesis path planning module calls a Transformer-based retrosynthetic prediction model, inputting novel additive-derived molecules unknown in the priority experimental samples. The model automatically translates and outputs a two-step nucleophilic substitution synthesis route starting from commercially available precursors, and simultaneously outputs process parameters: optimal reaction temperature 25℃±2℃, mechanical stirring speed 300 rpm, stirring time under inert gas protection 30 min, and recommended solvent anhydrous acetonitrile.

[0043] Automated preparation and battery assembly: A high-throughput liquid addition robot automatically controls a microfluidic pump according to the generated process parameters to automatically mix and inject the selected No. 7 formulation (DME:DOL = 3:1 molar ratio, containing 1.2 MiFSI and 2.5 wt.% LiNO3) into the CR2032 coin cell lithium metal battery system.

[0044] In-situ characterization and three-dimensional curve imaging: In-situ characterization: Real-time monitoring of 874 cm⁻¹ using in-situ Raman spectroscopy during charging and discharging. -1 No. 3 -The intensity decay of the vibration peak was used to quantitatively obtain its reduction and consumption kinetics; at the same time, the electrode surface was dynamically observed by in-situ transmission electron microscopy (TEM) to confirm that the lithium deposition behavior is a planar dense morphology rather than a dendrite tip morphology. 3D Status Imaging: The data acquisition module captures the voltage-capacity (VQ) curve for the entire battery lifecycle every 3 cycles. This curve is then time-stacked and transformed into a 224×224×3 RGB 3D matrix status image, where the horizontal axis represents voltage, the vertical axis represents capacity, and the three RGB color channels correspond to the voltage polarization rate, coulombic efficiency decay rate, and ohmic resistance change, respectively. A convolutional neural network classifier reads this RGB image and outputs the current dynamic health rating of the battery in real time.

[0045] Online closed-loop control and updates: When the battery cycled to the 450th cycle, the RGB image showed obvious color distortion and abrupt changes in geometric texture. The system identified in real time that the battery had transitioned from a deteriorating transition state to a "degraded state," with micro-polarization and local micro-short circuits occurring internally. The system immediately activated the deep reinforcement learning (DQN) control loop, and the online reverse output optimization compensation action was as follows: through a microfluidic micro-injection linkage device, LiNO3 functional additives were dynamically added to the battery system to increase its final mass percentage to 3.5% wt.%, while simultaneously sending control commands to the Blue Electric test station to lower the charging cutoff voltage from 4.3V to 4.2V and reduce the charging current rate to 0.5C.

[0046] Implementation effect verification: In terms of cycle life, after undergoing active closed-loop microfluidic replenishment and strategy regulation, the internal micro-short circuits disappeared, and the battery successfully operated for more than 2000 cycles under 1C high-current charge-discharge, with a capacity retention rate of up to 95.2%, which is 500% higher than the 400 cycles of the unoptimized system. Regarding ionic conductivity, the liquid additive system showed a measured ionic conductivity of 8 mS / cm at 25℃ and reached 13.3 mS / cm at 85℃. In terms of interface characteristics, in-situ testing showed that the negative electrode surface formed a layer rich in Li3N and LiN. x O y The high-density bilayer SEI film of the component significantly reduces the interfacial electronic impedance, and dendrite growth is completely blocked.

[0047] Example 2: Multi-objective optimized design of FEC+LiFSI hybrid additives.

[0048] The purpose of this embodiment is to synergistically optimize the liquid formulation of high-energy-density, high-voltage soft-pack lithium metal batteries, and solve the multi-objective, strong nonlinear, and non-equilibrium problem between high-rate charge-discharge performance and liquid thermal / electrochemical stability under high temperature and high voltage.

[0049] Data Management and Multi-Scale Prediction: This involved collecting and standardizing datasets from literature and patents containing fluoroethylene carbonate and lithium difluorosulfonylimide complex systems. The datasets included 245,000 chemical molecular topology diagrams and measured thermal decomposition temperatures (T0.05). d A total of 2100 sets of data were collected, including oxidation potential and high-rate capacity retention. Multi-objective feature joint modeling: A permutation-invariant attention network was used, with the GIN encoding vector of the molecules and the mass fractions of the electrolyte components simultaneously as joint inputs. In the physical information embedding layer, Pareto multi-objective cooperative boundary constraints were forcibly introduced: Minimize [f1(oxidation overpotential), -f2(conductivity), -f3(thermal decomposition temperature)], to enable the model to apply to the complex liquid system at high voltages (≥4.5V vs. Li / Li). + Joint adaptive accurate prediction of multidimensional complex nonlinear space with high current (charging and discharging above 3C).

[0050] Reverse generation and active learning space reduction: High-throughput candidate space expansion, limiting the solvents in candidate formulations to ethylene carbonate, diethyl carbonate, and FEC, and the lithium salts to LiFSI and LiPF6. The core parameter ranges for optimization are limited to: the molar ratio of FEC to EC between 0.2 and 0.8, and the molar concentration of LiFSI in the liquid system between 0.5M and 3.0M.

[0051] Formulation flux generation and Bayesian screening: Capacity retention ≥85% under high-rate 3C charging and no exothermic decomposition at thermal stability temperature ≥85℃ were used as target constraint vectors. The diffusion model generated 2000 non-repeating candidate formulations of locally coordinated composite liquid additives through 800 steps of inverse denoising iteration. The active learning screening unit used the Bayesian model averaging (BMA) algorithm to maximize the information entropy gain of the candidate space, selecting and outputting a priority experimental formulation sample queue of 50 formulations with Pareto performance potential.

[0052] Synthetic Path Planning and Automated Experimental Validation: ① Path and Parameter Recommendation: The synthetic path planning module, based on 50 priority experimental formulations and combined with a pre-trained reaction prediction network, recommends process parameters online: In a fully automated high-throughput gas inert protection chamber, the mixing environment temperature is set at 15℃±0.5℃, and the moisture and oxygen content is limited to less than 0.1 ppm, with mechanical stirring time of 45 min. ② High-Throughput Soft-Pack Battery Testing: The high-throughput fully automated liquid addition and assembly line automatically prepares liquids according to process parameters and assembles high-voltage, low-electrolyte lithium metal soft-pack batteries with an Ah specification. ③ In-situ Evaluation of Multi-Objective Electrochemical Performance: The soft-pack batteries are directly connected to a multi-channel electrochemical workstation for 0.5C to 5C step-rate charge-discharge cycle testing, and in-situ adiabatic accelerated calorimetry (ARC) testing is performed in an 85℃ constant temperature chamber, recording the self-exothermic onset temperature. Simultaneously, an in-situ Raman spectrometer focuses on the surface of the negative electrode foil to monitor the solvation and desolvation kinetics of F-rich molecules during cycling in real time.

[0053] Closed-loop iteration and real-time control: After the first round of experiments, feedback results showed that some high-concentration LiFSI formulations, while exhibiting excellent rate performance, showed trace gas production at 85℃, indicating insufficient thermal stability. The closed-loop optimization module reduced the weight of these failed test samples according to their corresponding low confidence levels and fed them back into the standardized database. The multi-scale prediction model completed a round of supervised incremental updates through mini-batch gradient descent, readjusting the correlation coefficient between the FEC molar ratio and the free energy of the LiFSI solvation shell. In the second round of iteration, the system recommended the optimized formulation No. 22: precisely positioning the FEC:EC molar ratio at 0.52:0.48, adjusting the LiFSI concentration to 2 M, and introducing 0.5 wt.% thiocarbonate functional additive.

[0054] Implementation Results Verification: Cycling and Rate Performance. The final electrolyte formula, selected through three rounds of closed-loop iterative optimization, resulted in a high-voltage soft-pack lithium metal battery that stably cycled 2150 cycles under a 4.5V high cutoff voltage and 1C cycle conditions, maintaining a capacity retention of ≥90%. Even at an extremely high charging rate of 5C, the battery still stably released 90.1% of its initial capacity, achieving a Pareto optimal balance between high lifespan and high rate capability.

[0055] Thermal stability: The battery was stored continuously at 85℃ for 200 hours without any gas generation, swelling, decomposition, or heat release.

[0056] SEI film microstructure: In-situ TEM and XPS analysis showed that an ultrathin monolayer SEI film with a thickness of only 4 nm, rich in LiF and with an inorganic phase volume fraction of ≥60%, was formed in-situ on the negative electrode surface. This film effectively isolated the electrolyte solvent from further continuous side reactions.

[0057] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An optimization system for liquid additives in lithium metal batteries based on artificial intelligence, characterized in that, Data management module: used to collect multi-source data and perform standardized integration to build a standardized database containing the mapping relationship between composition, structure, solvation and performance; wherein, the multi-source data includes literature / patent data, multi-scale simulation calculation data, experimental test data and real-time battery operation data; Multi-scale prediction model module: used to build prediction models with embedded physical information, integrate physical constraints at different scales, and predict key performance indicators and microscopic properties of liquid additives. Deep active learning inverse generation module: Used to construct an inverse generation framework based on diffusion model with battery performance target as constraint, generate candidate liquid additive formulations, and combine active learning strategy to screen priority experimental samples from the candidate liquid additive formulations. Synthesis Path Planning Module: Based on the reaction prediction model, it performs reverse synthesis path planning for priority experimental samples and outputs the corresponding synthesis process parameters; Experiment execution and real-time detection module: used to prepare liquid additives and assemble batteries according to synthesis process parameters, conduct electrochemical performance tests on batteries to obtain experimental data, and obtain microscopic characteristic data and macroscopic operating status through in-situ characterization and voltage curve analysis; Closed-loop optimization module: used to feed back the experimental data and microscopic characteristic data obtained by the experimental execution and real-time detection module to the data management module, so as to dynamically update the multi-scale prediction model module and dynamically adjust the liquid additive formula or battery charging strategy for the identified abnormal battery operating states.

2. The optimization system for lithium metal battery liquid additives based on artificial intelligence according to claim 1, characterized in that, The multi-scale prediction model module includes: Molecular characterization sub-model: A graph neural network is used to encode solvent, lithium salt and additive molecules and embed physical feature vectors containing molecular flexibility and functional group activity. Mixture prediction sub-model: Based on an attention mechanism-based permutation invariant network, it handles the disorder of multi-component formulations and predicts the comprehensive electrochemical performance of liquid additives for hybrid solid-liquid lithium metal batteries. Interpretable analysis sub-model: Identify the core molecular descriptors that affect the performance of liquid additives through gradient decoupling and symbolic regression algorithms.

3. The optimization system for lithium metal battery liquid additives based on artificial intelligence according to claim 1, characterized in that, The deep active learning inverse generation module includes: Candidate space building unit: used to integrate external chemical databases to build candidate spaces containing at least 1 million molecules and to limit the range of chemical structures of solvents, lithium salts and additives; Reverse generation unit: Based on the diffusion model, it generates the molecular structure and formulation of liquid additives that meet multiple objective constraints, and outputs the molecular structure topology, component molar ratio and additive mass percentage content. Active learning screening unit: Using Gaussian process regression combined with Bayesian model averaging algorithm, the information gain and performance potential of the candidate liquid additive formulations are calculated, and samples whose information gain and performance potential meet the preset threshold are selected as priority experimental samples.

4. The optimization system for lithium metal battery liquid additives based on artificial intelligence according to claim 1, characterized in that, The experimental execution and real-time detection module uses voltage curve three-dimensional imaging technology to encode the charge and discharge data within a set number of cycles into RGB state images. By recognizing the RGB state images, the battery operating state is classified into a stable state, a transitional state, or a decay state.

5. The optimization system for lithium metal battery liquid additives based on artificial intelligence according to claim 1, characterized in that, The closed-loop optimization module weights the experimental test data and the in-situ characterization results according to the data confidence level, and feeds the weighted data back to the data management module, and updates the multi-scale prediction model module through the mini-batch gradient descent algorithm. When the battery operating state is identified as a decay state, the proportion of functional additives in the liquid additive is dynamically adjusted by combining a deep reinforcement learning network, or the charging rate and cutoff voltage are dynamically updated.

6. The optimization system for lithium metal battery liquid additives based on artificial intelligence according to claim 1, characterized in that, Within the chemical range defined by the candidate spatial building blocks: The solvent includes at least one of ether solvents, carbonate solvents, and ketone solvents, or a mixture thereof; The lithium salt includes at least one of LiPF6, LiFSI, LiTFSI, and LiClO4, or a mixture thereof; The additives include at least one of LiNO3, fluoroethylene carbonate, thiocarbonates, and small molecule additives containing nitrogen or sulfur functional groups.

7. The optimization system for lithium metal battery liquid additives based on artificial intelligence according to claim 1, characterized in that, The experimental execution and real-time detection module includes an in-situ characterization and analysis unit; the in-situ characterization and analysis unit includes at least one of an in-situ Raman spectrometer and an in-situ transmission electron microscope, used to detect changes in the microscopic chemical composition of lithium deposition morphology and solid electrolyte interface film in real time.

8. An optimization method for liquid additives in lithium metal batteries based on artificial intelligence, characterized in that, The optimization system for lithium metal battery liquid additives based on artificial intelligence, as described in any one of claims 1 to 7, comprises the following steps: S1. Data Standardization and Model Initialization: Collect multi-source data through the data management module and perform cleaning, normalization and confidence labeling to build a standardized database; train a multi-scale prediction model embedding physical information based on the initial database and set the model baseline error; S2. Performance Targets and Constraints: Set the target performance targets and engineering application constraints for the liquid additive, and limit the concentration range of each component of the liquid additive. S3. Reverse generation and active learning screening: Based on the reverse generation framework, candidate liquid additive formulations that meet the target constraints are generated, and the information gain and performance prediction score of the candidate liquid additive formulations are calculated through the active learning screening unit to select priority experimental samples. S4. Synthesis Path Planning and Experiment Execution: The synthesis path planning module is used to plan the reverse synthesis path for priority experimental samples and output the synthesis process parameters. The experiment execution and real-time detection module automatically completes the synthesis of liquid additives and battery assembly, and performs electrochemical testing and in-situ characterization to obtain performance data, microscopic characteristic data and real-time battery operating status. S5. Real-time closed-loop optimization: The acquired performance data and micro-characteristic data are weighted according to confidence level and fed back to the standardized database using the closed-loop optimization module to update the multi-scale prediction model; at the same time, the liquid additive formula or charging strategy is dynamically adjusted in combination with the real-time operating status of the battery. S6. Optimal formulation output: Iterate through steps S3 to S5 until the optimal liquid additive formulation that meets all target performance objectives is selected, and output the corresponding molecular structure, molar ratio, synthesis process and battery performance verification report.

9. The optimization method for liquid additives for lithium metal batteries based on artificial intelligence according to claim 8, characterized in that, In step S3, at least 1,000 candidate formulations are generated based on the reverse generation framework, and a priority experimental sample consisting of 20 to 50 samples is selected from them by an active learning screening unit.

10. The optimization method for liquid additives in lithium metal batteries based on artificial intelligence according to claim 8, characterized in that, In step S5, the dynamic adjustment specifically includes: When the real-time operating state of the battery is identified as a decay state, the closed-loop optimization module enables a deep reinforcement learning algorithm to maintain the battery capacity retention rate as the reward function and outputs a compensation strategy for the current decay state online in reverse. The compensation strategy includes increasing the mass percentage of LiNO3 in the formulation, or reducing the charging current rate and simultaneously lowering the cutoff voltage at the charging control terminal.