Simulation simulation system for research and development of new lithium battery material

Abstract

The invention discloses an analogue simulation system for lithium battery new material research and development, which relates to the technical field of lithium battery new material research and development and comprises a theoretical simulation calculation module, a machine learning model module, a high-throughput preparation module, a high-throughput characterization module, a closed-loop feedback module and a special database module for positive electrode material genes. The theoretical simulation calculation module executes first principle calculation and molecular dynamics simulation to obtain atomic-scale characteristics and mesoscopic dynamic parameters of the lithium battery positive electrode material. According to the analogue simulation system for research and development of the new lithium battery material, a'theoretical simulation-high-throughput preparation-high-throughput characterization-model updating 'whole-process intelligent iteration system is constructed: in a theoretical simulation stage, a machine learning model calculates preparation parameters based on multi-scale data preliminary screening of first principle calculation and molecular dynamics simulation; characterization data is fed back to a database, model incremental training is triggered, and the research and development iteration period is efficiently shortened.

Description

Technical Field

[0001] This invention relates to the field of lithium battery new material research and development technology, and in particular to a simulation system for lithium battery new material research and development. Background Technology

[0002] In the current field of lithium battery new material research and development, the traditional approach faces multiple core challenges. On the one hand, the requirements for material performance in fields such as new energy and advanced manufacturing are stringent, and the cross-scale correlation mechanism is unclear. First-principles calculations can only resolve atomic-level properties, and molecular dynamics simulations focus on mesoscopic dynamics. The connection between these two and macroscopic battery performance relies on experience, leading to a break in the "atomic-mesoscopic-macroscopic" scale during research and development. The experimental trial-and-error cycle is long, and the optimization of high-throughput preparation parameters requires repeated iterations. The parameter combination trial and error of the cathode material co-precipitation process alone requires a large number of experiments and takes several weeks. The computational resources are huge, and pure physical simulations are difficult to efficiently support the design of multi-system materials.

[0003] On the other hand, although there are technological attempts to use machine learning for materials research and development, most of them are purely data-driven "black box" models that do not incorporate the physical characteristics of molecular dynamics and first-principles calculations. The models have poor interpretability and weak cross-system generalization ability, and significant migration errors between different cathode materials or catalytic systems. Moreover, existing research and development has not formed a closed loop of "theoretical simulation - high-throughput preparation - high-throughput characterization - model update". The massive amount of computational data, preparation parameters and characterization results are difficult to integrate and utilize efficiently, and cannot provide accurate data support for subsequent research and development, which restricts the efficiency of the transformation of new lithium battery materials from laboratory research and development to industrial application.

[0004] Therefore, it is necessary to propose a simulation system for the research and development of new lithium battery materials to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a simulation system for the research and development of new lithium battery materials, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a simulation system for the research and development of new lithium battery materials, comprising a theoretical simulation calculation module, a machine learning model module, a high-throughput preparation module, a high-throughput characterization module, a closed-loop feedback module, and a dedicated database module for cathode material genes;

[0007] The theoretical simulation calculation module performs first-principles calculations and molecular dynamics simulations to obtain the atomic-level characteristics and mesoscopic dynamic parameters of lithium battery cathode materials.

[0008] The machine learning model module constructs and trains a model based on the aforementioned parameters to predict material properties and optimize preparation parameters;

[0009] The high-throughput preparation module prepares cathode materials according to the optimized parameters output by the model;

[0010] The high-throughput characterization module collects material structure and electrochemical performance data;

[0011] The closed-loop feedback module feeds back the characterization data to the theoretical simulation calculation module and the machine learning model module, updating the data and model parameters to form a closed loop for the entire R&D process.

[0012] Preferably, in the theoretical simulation calculation module, the DFT uses VASP software based on GGA-PBE functionals to calculate the crystal structure, electronic structure, and Li of the cathode material. + Migration barrier;

[0013] MD uses LAMMPS software to select ReaxFF, PCFF or CVFF force fields depending on the material type, and outputs the ion diffusion coefficient and SEI film growth rate.

[0014] Preferably, the machine learning model module includes supervised learning, GNN, DQN and unsupervised learning sub-models;

[0015] The supervised learning sub-model uses the random forest algorithm, taking DFT electronic structure features and MD mesoscopic parameters as input, and outputs Li + Migration barrier;

[0016] The GNN sub-model is based on the DGL library, takes an atomic graph as input, and outputs polysulfide conversion efficiency or cycle lifetime.

[0017] Dynamically optimize the preparation parameters of the DQN sub-model;

[0018] The unsupervised learning sub-model uses k-means clustering to identify the causes of cyclic decay.

[0019] Preferably, the dedicated database module for cathode material genes is encrypted with AES-256 and stores DFT / MD raw files, machine learning model weight files, preparation process parameters and characterization data. It supports multi-dimensional retrieval and has a response time of less than 1 second.

[0020] Preferably, the high-throughput preparation module adopts the co-precipitation method, equipped with a multi-channel reactor and a high-temperature sintering furnace. The parameter control range is: source solution concentration 0.8-1.5 mol / L, reaction temperature 40-60℃, pH value 9-11, stirring speed 500-800 rpm, sintering temperature 800-1000℃, holding time 2-6 h, and product particle size variation coefficient ≤8%.

[0021] Preferably, the high-throughput characterization module includes a structural and electrochemical characterization unit;

[0022] The structural characterization units employed XPS, XRD, and X-ray fluorescence. XRD data was analyzed using a CNN model, and the lattice constant and grain size were inverted within one hour, with errors of [missing information]. <5nm; Electrochemical characterization units are tested for charge-discharge and cycle performance in a specific mode.

[0023] Preferably, the machine learning model module further includes a multi-task learning sub-model, which takes XPS element content and MD diffusion coefficient as input, and synchronously outputs capacity retention rate and 5C ratio capacity retention rate after 1000 cycles, with a deviation ≤3%.

[0024] Preferably, the GNN sub-model adopts transfer learning, transferring the GNN model of the polyethylene cracking system to the design of Fe-V dual-atom catalysts for lithium-sulfur batteries. Only 200 target system atom diagrams need to be fine-tuned, the convergence speed is increased by 50%, and the data requirement is reduced by 40%.

[0025] Preferably, the feedback process of the closed-loop feedback module is as follows: when the deviation between the characterization data and the model prediction value is >5%, the entire sample process data is added to the database;

[0026] For every 100 new data sets added to the database, incremental training of the model is triggered, improving prediction accuracy by 3%-5%.

[0027] Preferably, the simulation system is adapted for the research and development of NCM, LFP, LCO cathode materials and Fe-V diatomic catalysts for lithium-sulfur batteries.

[0028] The technical effects and advantages of this invention are as follows:

[0029] 1. This invention constructs a fully intelligent iterative system encompassing "theoretical simulation - high-throughput preparation - high-throughput characterization - model update": In the theoretical simulation stage, the machine learning model uses multi-scale data from first-principles calculations and molecular dynamics simulations to initially screen preparation parameters; in the high-throughput preparation stage, a reinforcement learning agent collects parameters such as temperature and pH in real time, automatically adjusting them when they deviate from the optimal values; in the high-throughput characterization stage, a CNN model resolves XRD patterns and inverts lattice parameters within one hour; characterization data is fed back to the database, triggering incremental model training and efficiently shortening the R&D iteration cycle;

[0030] 2. This invention establishes a dedicated database of cathode material genes encrypted with AES-256, storing DFT / MD original files, machine learning model weights, preparation process parameters, and characterization data. It supports multi-dimensional retrieval of "material type-performance index-process parameters" with a response time of less than 1 second. By identifying the causes of material cycle degradation through k-means clustering and combining data mining technology, it can both predict performance based on material structure and infer structure and process from target performance, providing data support for material optimization.

[0031] 3. This invention targets the characteristics of cathode materials such as NCM, LFP, and LCO, and uses a co-precipitation method combined with a multi-channel reactor. Machine learning models are used to optimize parameters such as source solution concentration and reaction temperature, and the product particle size variation coefficient is controlled to be ≤8%, ensuring consistency in batch preparation. Characterization methods such as XPS, XRD, and electrochemical testing are integrated, and machine learning is used to correlate characterization data with battery PACK-level industrialization performance, reducing the conversion cycle and cost from laboratory results to industrial applications.

[0032] 4. This invention employs transfer learning technology to transfer the GNN model of the polyethylene cracking system to the design of Fe-V diatomic catalysts for lithium-sulfur batteries. Only 200 sets of target system data are required for fine-tuning, the convergence speed is increased by 50% and the data requirement is reduced by 40%. The machine learning model uses DFT electronic structure features and MD mesoscopic parameters as core inputs, rather than single material-specific parameters, which can be adapted to the development of cathode materials and lithium-sulfur battery catalysts, significantly improving the flexibility of system application. Attached Figure Description

[0033] Figure 1 This is a flowchart of the simulation system for the research and development of new lithium battery materials according to the present invention. Detailed Implementation

[0034] This invention provides, for example Figure 1 The simulation system shown is used for the research and development of new lithium battery materials. It includes three core modules, each of which deeply integrates machine learning (ML) and multi-scale simulation technologies. The specific process is as follows:

[0035] (I) S1: Construction of a cathode material gene database integrating multi-scale simulation and ML

[0036] Multi-scale simulation data acquisition

[0037] DFT calculations: Using VASP software, based on the generalized gradient approximation (GGA)-PBE functional, the crystal structure and electronic structure (density of states DOS, Bader charge analysis) of cathode materials (NCM, LFP, LCO) and lithium-sulfur battery catalysts (Fe-V diatomic) were calculated. + Migration paths and energy barriers; for example, calculating the effect of Ni doping ratio (80%-90%) in NCM on Li. + The effect of diffusion barrier is shown, and barrier data for different doping ratios are output (e.g., barrier 0.45eV for 80% Ni and 0.52eV for 90% Ni).

[0038] MD simulation: LAMMPS software was used, with ReaxFF force field (simulating polysulfide diffusion at the cathode-electrolyte interface in lithium-sulfur batteries), PCFF force field (simulating LFP particle growth dynamics), and CVFF force field (simulating LCO grain boundary migration) selected respectively. The ion diffusion coefficients were output (e.g., LFP diffusion coefficient 1.2e-10cm at 25℃). 2 / s), interfacial reaction kinetic parameters (such as SEI film growth rate of 0.02 nm / h).

[0039] Cross-scale correlation analysis: A mapping relationship is established between the DFT energy barrier (atomic level), MD diffusion coefficient (mesoscopic level), and macroscopic capacity retention rate using a gradient boosting tree model. 2 (Determination coefficient) > 0.93.

[0040] ML Model Training and Optimization

[0041] Supervised learning model: Predicting Li using a random forest model (based on scikit-learn). + The migration barrier was determined using input features including atomic electronegativity, bond length, and band width. The training set consisted of 80% DFT data, and the test set consisted of 20%. The cross-validation error was <3%. The SVM model was used to classify the cycle stability of cathode materials ("excellent / good / poor"), with input features including MD diffusion coefficient and the full width at half maximum (FWHM) of XRD characteristic peaks. The classification accuracy was >92%.

[0042] Graph Neural Network (GNN) Model: An atomic graph model is constructed based on the DGL library. The node features are atom type and charge density, and the edge features are bond energy and bond angle. It is used to predict the polysulfide conversion efficiency of Fe-V catalysts in lithium-sulfur batteries. The model training adopts transfer learning (the idea of ​​transferring the polyethylene cracking model to the glucose catalytic system, and then to catalyst design), which reduces the data requirement by 40%.

[0043] Reinforcement learning model: A preparation parameter optimization agent is constructed based on the DQN algorithm (PyTorch framework). The state is the real-time preparation temperature and pH value, and the reward is the material performance score (comprehensive particle size uniformity and capacity retention). It converges after 500 iterations, and the parameter optimization efficiency is improved by 80%.

[0044] Encrypted database construction

[0045] The database stores the following: DFT / MD original calculation files (VASPINCAR, LAMMPSinput), ML model weight files (random forest.pkl, GNN.pt), preparation process parameters (coprecipitation temperature, sintering time), and characterization data (XRD patterns, charge-discharge curves). It uses the AES-256 encryption algorithm to ensure data security, supports fast retrieval by "material type-performance index", and has a response time of <1s.

[0046] (II) S2: Optimization of high-throughput preparation of cathode materials driven by ML

[0047] Preliminary screening of preparation parameters

[0048] Historical data of target materials (such as NCM and LFP) were retrieved from the S1 database. Parameter ranges were screened based on Bayesian optimization algorithm (combined with random forest model predictions). For example, the parameter ranges for LFP coprecipitation process were: source solution concentration 0.8-1.5 mol / L, reaction temperature 40-60℃, pH value 9-11, and stirring speed 500-800 rpm. After initial screening, the number of parameter combinations was reduced from 120 to 20.

[0049] Real-time dynamic optimization

[0050] The high-throughput preparation equipment (multi-channel reactor, high-temperature sintering furnace) is linked with the DQN agent to collect temperature fluctuations (±0.5℃), pH changes (±0.1), and stirring torque (±5N·m) in real time during the preparation process. When the parameters deviate from the ML-predicted optimal value, the agent automatically adjusts (e.g., when the pH value is 0.2 lower than the optimal value, the precipitant drop acceleration rate is increased by 0.5mL / min).

[0051] For the sintering process of Fe-V catalysts for lithium-sulfur batteries, the DQN intelligent agent adjusts the heating rate (5-15℃ / min) and holding time (2-6h) in real time to ensure that the number of active sites of the catalyst (detected by in-situ Raman) reaches more than 95% of the predicted value.

[0052] Multi-scale simulation verification

[0053] For the prepared intermediate samples, the mesoscopic structure (such as LFP particle dispersibility, CV value <5% is acceptable) is verified using GROMACS software (MD simulation), and the electronic structure (such as the deviation of the peak position of the NCM density of states from the predicted value <0.1eV) is calculated using VASP. Data of unacceptable samples are automatically fed back to the S1 database, triggering incremental training of the ML model (new data accounts for 10%).

[0054] (III) S3: ML-enhanced high-throughput characterization and performance prediction

[0055] Multidimensional representation data collection

[0056] Structural characterization: XPS (characterizing surface elemental composition, such as the Li / O ratio on LCO surface), XRD (characterizing crystal phase structure, such as the intensity of the (003) peak of NCM), and X-ray fluorescence (characterizing elemental distribution uniformity) were used, with 10 sets of data collected per batch of samples.

[0057] Electrochemical characterization: Using a multi-functional battery test channel (25℃ constant temperature environment), the charge and discharge specific capacity, cycle life (1000 times), and rate performance were tested in the mode of "constant current and constant voltage 30min charging (0.1C) - constant current discharging (0.1C / 1C / 5C)". The test time for each sample was shortened from the traditional 24h to 8h (based on ML prediction to terminate invalid tests early).

[0058] ML-driven data parsing

[0059] Structural analysis: XRD patterns are processed using a CNN model (PyTorch framework). The input is diffraction peak data of 20 (10-80°), and the output is the lattice constant (e.g., LFP). The resolution time has been reduced from the traditional 24 hours to 1 hour, with an error of <0.1%.

[0060] Performance Prediction: A multi-task learning model is used to simultaneously predict cycle lifetime and rate performance. The input feature is the elemental content of XPS (e.g., Ni in NCM). 3+ / Ni 2+ The prediction bias was <3% for the ratio and MD diffusion coefficient; the k-means clustering algorithm (unsupervised learning) was used to classify the cyclic decay data and identify the causes of decay (phase transformation / SEI film rupture), with a classification accuracy of >92%.

[0061] Closed-loop feedback and model update

[0062] The characterization analysis results (e.g., "LFP particle size CV = 6%, capacity retention rate of 85% after 1000 cycles") are compared with the ML prediction values. If the deviation is >5%, the full process data (DFT / MD / preparation / characterization) of the sample is added to the S1 database, triggering the ML model update (random forest retraining, GNN weight fine-tuning). After the update, the model prediction accuracy is improved by 5%-8%.

[0063] The simulation system implementation example is as follows:

[0064] Example 1:

[0065] Random forest optimization of LFP cathode coprecipitation preparation process

[0066] Objective: To verify whether the ML model can output LFP materials that meet performance requirements by optimizing the parameters under the condition of "selecting all lower limits of preparation parameters". The lower limits of parameters are: source solution concentration 0.8 mol / L, reaction temperature 40℃, pH value 9, and stirring rate 500 rpm.

[0067] step:

[0068] (1) Data Acquisition and Model Training: 200 sets of LFP coprecipitation data were retrieved from the S1 database, of which 50 sets were data near the lower limit of parameters (concentration 0.8-1.0 mol / L, temperature 40-42℃, pH 9-9.2, stirring 500-520 rpm). A random forest model was used to establish a parameter-performance mapping. The input features included concentration, temperature, pH, and stirring rate. The output features were particle size CV value (target <8%) and capacity retention rate after 1000 cycles (target >80%). Model cross-validation R 2 =0.94, Root Mean Square Error (RMSE) =0.02.

[0069] (2) Parameter fine-tuning prediction: The model predicts based on the lower limit of parameters (0.8 mol / L, 40℃, pH 9, 500 rpm), with initial performance of particle size CV = 11.2% (exceeding the standard) and capacity retention rate = 76% (not meeting the standard). Through Bayesian optimization, the model outputs fine-tuning parameters: concentration 0.85 mol / L (still lower than the median value of 1.15 mol / L), temperature 41℃ (lower than the median value of 50℃), pH 9.1 (lower than the median value of 10), stirring speed 510 rpm (lower than the median value of 650 rpm), with predicted performance of particle size CV = 7.8% (meeting the standard) and capacity retention rate = 81% (meeting the standard).

[0070] (3) Multi-scale simulation verification:

[0071] DFT calculation: LFP crystal structure prepared by fine-tuning parameters was calculated using VASP, Li + The migration barrier is 0.38 eV, with an error of <2.6% compared to the model prediction (0.39 eV);

[0072] MD simulation: LFP particle growth was simulated using LAMMPS (PCFF force field), with a diffusion coefficient of 1.1e-10 cm⁻¹ at 25°C. 2 / s, which meets the requirements of the model input features.

[0073] (4) Experimental verification: High-throughput preparation was carried out with fine-tuned parameters (multi-channel reactor, 3 sets of parallel samples), and the test results were as follows:

[0074] Structural characterization: XRD showed LFP pure phase (no impurity peaks), particle size CV = 7.6% (0.2% deviation from the predicted value of 7.8%);

[0075] Electrochemical performance: The initial capacity at 25℃ and 1C rate is 158mAh / g, the capacity retention after 1000 cycles is 81.5% (0.5% deviation from the predicted value of 81%), and the capacity retention at 5C rate is 68% (better than the initial value of 60% at the lower limit of the parameter).

[0076] Compared to the traditional trial-and-error method (requiring 30 experiments and taking 21 days), the ML model only requires 5 experiments and 5 days to complete the optimization, improving efficiency by 4 times; and it still achieves performance targets even when the parameters are close to the lower limit, breaking through the empirical understanding that "low parameters must have low performance".

[0077] Example 2:

[0078] GNN optimizes the Ni doping process for NCM90100 cathode.

[0079] Objective: To verify whether optimizing the atomic-level structure using a GNN model can reduce the Li doping density under the condition of "maximum Ni doping ratio (90%)". + Migration energy barrier, improving the rate performance of NCM90100 (traditionally, it is believed that Ni>85% will lead to Li) + Diffusion is hindered.

[0080] step:

[0081] (1) Basic data for DFT calculation: The NCM crystal structure with Ni doping ratio of 80%-90% (step size 2%) was calculated using VASP, with a focus on analyzing the atomic arrangement when Ni is 90% doped—discovering Ni 3+ Enrichment on the (003) crystal plane leads to Li + The diffusion path is narrow, and the initial migration barrier is 0.52 eV (higher than 0.45 eV for 80% of Ni), consistent with conventional wisdom.

[0082] (2) GNN model construction and training:

[0083] Atomic diagram design: Taking NCM crystal as the object, the node features include the charge density of Ni / Co / Mn atoms (Bader charge analysis results) and ionic radius, and the edge features are bond length and bond angle (DFT optimized data);

[0084] Transfer learning: Drawing on the idea of ​​"transferring the polyethylene cracking model to the glucose catalytic system", the NCMGNN model with Ni 80%-88% (training set of 1000 atomic diagrams) was transferred to the Ni 90% system. Only 200 Ni 90% atomic diagrams were added for fine-tuning, and the model convergence speed was improved by 50%.

[0085] Structural optimization prediction: GNN model outputs the optimal atomic arrangement of 90% Ni – introducing 0.5% Al doping into the (003) crystal plane (suppressing Ni) 3+ (Enrichment), predicting Li + The migration barrier drops to 0.40 eV (lower than 0.42 eV for Ni 85%).

[0086] (3) MD simulation verification: LAMMPS (ReaxFF force field) was used to simulate NCM particles containing 90% Ni (including 0.5% Al) at 25℃. + The diffusion coefficient is 1.3e-10 cm⁻¹ 2 / s, superior to 90% of the original Ni structure (1.0e-10cm) 2 / s), consistent with the energy barrier reduction trend predicted by GNN; at the same time, simulating the phase transformation during the cycle, it was found that Al doping increased the H2 phase → H3 phase transformation temperature from 200℃ to 250℃, suppressing the collapse of the high-temperature structure.

[0087] (4) Experimental preparation and testing:

[0088] Preparation process: NCM90100 (0.5% Al doped) was prepared by co-precipitation according to the atomic structure predicted by GNN, with a sintering temperature of 1000℃ and a holding time of 6h, and the Li / M ratio was controlled at 1.12 (slightly higher than the conventional 1.05).

[0089] Structural characterization: XPS showed that Al was uniformly distributed on the (003) crystal plane, and Ni... 3+ The content was reduced by 15%; XRD showed that the (003) / (104) peak intensity ratio was 1.8 (better than the original Ni 90% 1.2), and the crystal plane order was improved;

[0090] Electrochemical performance: At 25℃, the initial capacity at 1C is 205mAh / g (higher than 195mAh / g of Ni85%), the capacity retention rate after 1000 cycles is 75% (on par with Ni85% and better than 65% of the original Ni90%), and the capacity retention rate at 5C is 72% (better than 68% of Ni85%), breaking through the traditional limitation that "Ni90% must be low-rate".

[0091] Atomic-level structure optimization was achieved using a GNN model, under the condition that the Ni doping ratio reached its upper limit, Li + The migration barrier is reduced by 23%, and the rate performance surpasses that of materials with medium and low doping ratios; the calculation cycle of DFT+GNN is shortened from the traditional 30 days to 10 days, providing a new path for the research and development of high NiNCM.

[0092] Example 3:

[0093] DQN optimizes the sintering process of Fe-V dual-atom catalyst for lithium-sulfur batteries.

[0094] Objective: (Sintering temperature 900℃, holding time 4h, heating rate 10℃ / min) To improve the conversion efficiency of catalyst for polysulfides (LiPSs) by dynamically optimizing the process through DQN reinforcement learning.

[0095] step:

[0096] (1) Parameter range and initial settings: The intermediate values ​​for the sintering process parameters are: temperature 900℃ (range 800-1000℃), holding time 4h (range 2-6h), and heating rate 10℃ / min (range 5-15℃ / min). The initial settings are based on the intermediate values. DFT calculations show that the Fe-V active site density is 2.5×10⁻⁶. 14 cm -2 The predicted conversion efficiency of LiPSs is 85%.

[0097] (2) Deployment of DQN agents:

[0098] State space: Real-time acquisition of sintering furnace temperature (±1℃), holding time (±5min), and exhaust gas composition (SO2 concentration, reflecting the intermediate state of LiPSs conversion);

[0099] Reward function: Combining active site density (40%) and LiPSs conversion efficiency (60%), the target reward value is >90;

[0100] Iterative optimization: A total of 300 iterations were performed, and convergence was achieved in the 250th iteration. The optimized intermediate parameters were output as follows: temperature 905℃ (close to the median value), holding time 4.2h (close to the median value), heating rate 9.8℃ / min (close to the median value), predicted reward value 92, and LiPSs conversion efficiency 92%.

[0101] (3) DFT-MD co-validation:

[0102] DFT calculation: The Fe-V bond energy of the optimized catalyst was calculated using VASP (2.8 eV), which is lower than the initial intermediate value (3.0 eV), indicating that the active sites are more likely to adsorb LiPSs.

[0103] MD simulation: The diffusion and decomposition of Li2S6 on the catalyst surface were simulated using LAMMPS (ReaxFF force field). The decomposition rate was 0.03 mol / (L·h), which is higher than the initial median value of 0.02 mol / (L·h) and consistent with the DQN prediction.

[0104] (4) Experimental verification:

[0105] Catalyst preparation: Fe-V / C catalyst was sintered according to the optimized intermediate parameters. The elemental distribution was characterized by X-ray fluorescence. The Fe / V ratio was 1:1 (design value), and there was no agglomeration.

[0106] Battery performance testing: The assembled lithium-sulfur battery (cathode contains 5% Fe-V / C catalyst) has an initial capacity of 1280 mAh / g at 25℃ and 0.2C rate. After 500 cycles, the capacity retention rate is 85% (better than the initial median value of 78%), and the LiPSs shuttle current decreases by 30% (results from electrochemical impedance spectroscopy).

[0107] The DQN model achieves performance breakthroughs through dynamic fine-tuning based on intermediate parameters, avoiding the empirical misconception that "the intermediate value is the optimal value." Compared with traditional orthogonal experiments (18 sets of experiments, taking 14 days), the ML method only requires 8 sets of experiments and 7 days to complete the optimization, and the LiPSs conversion efficiency is improved by 9%.

[0108] Example 4:

[0109] CNN+ multi-task learning optimizes LCO cathode characterization and performance prediction

[0110] Objective: Based on intermediate characterization parameters (XRD2θ = 10-80°, charge / discharge rate 0.5C), a rapid inversion of the crystal phase structure and cycle life prediction of LCO cathode can be achieved through CNN and multi-task learning model, replacing the traditional time-consuming refinement and long-cycle testing.

[0111] step:

[0112] (1) Data acquisition and preprocessing:

[0113] Characterization data: XRD patterns (2θ=10-80°, step size 0.02°), XPS patterns (Li1s, Co2p, O1s), and charge-discharge curves (0.5C, 25℃, voltage range 3.0-4.3V) of 1000 LCO samples were collected, of which 500 samples were intermediate performance samples (capacity retention of 80%-85% after 1000 cycles).

[0114] Data labeling: Lattice constants (a, c) and grain size were labeled using XRD refinement software (MAUD), and capacity retention was labeled using long-cycle testing (1000 cycles) as model labels.

[0115] (2) Model training:

[0116] CNN crystal inversion model: A 3-layer convolutional + 2-layer fully connected network is built based on PyTorch. The input is an XRD pattern (flattened into a 1D vector), and the output is the lattice constants a, c and the grain size. The training set is 80% and the test set is 20%. The lattice constant error after training is... Grain size error <5nm, resolution time reduced from 24h to 1h;

[0117] Multi-task learning prediction model: using the structural parameters (a, c, grain size) output by the CNN + the Co of XPS 3+ / Co 2+ The model takes the ratio as input and outputs a capacity of 0.5C and a capacity retention rate after 1000 cycles. It employs a shared feature layer + task-specific layer architecture. R... 2The values ​​are 0.95 (capacity) and 0.93 (retention rate), respectively, with a prediction error of <2%.

[0118] (3) Real-time feedback optimization:

[0119] For newly prepared LCO samples (sintering temperature 850℃, holding time 3h, intermediate process), CNN was used to invert the lattice constant. (MAUD refinement value) error The multi-task model predicts a capacity retention rate of 83% after 1000 cycles.

[0120] Actual test results: The initial capacity at 0.5C is 145mAh / g, and the capacity retention rate after 1000 cycles is 82.5% (0.5% deviation from the predicted value). If the predicted value is lower than the target of 80%, it will be automatically fed back to the S2 module to adjust the sintering temperature (e.g., increase it to 860℃).

[0121] Achieving "characterization as prediction" – performance prediction can be completed in just 1 hour of XRD / XPS testing without waiting for long-cycle testing, shortening the R&D cycle by 80%; and the multi-task model outputs multiple performance indicators at the same time, avoiding the limitations of a single model and providing real-time guidance for LCO process optimization.

[0122] This invention utilizes molecular dynamics (MD), first-principles calculations (DFT), and machine learning (ML) techniques to construct a full-scale data system covering the atomic, mesoscopic, and macroscopic levels. By performing DFT calculations using VASP software, the crystal structure, electronic structure (density of states (DOS), Bader charge distribution), and Lic of cathode materials (such as NCM and LFP) can be accurately obtained. + Key atomic-level data, such as migration barriers, are collected. MD simulations using LAMMPS software (with force fields such as ReaxFF and PCFF) can capture mesoscopic dynamic parameters such as ion diffusion coefficients and electrode-electrolyte interface reaction kinetics. Simultaneously, using this multi-scale physical data as core input features, ML models such as random forests and graph neural networks (GNNs) can be trained to establish a direct correlation between the material's microstructure and macroscopic performance (such as cycle life and rate performance). For example, in the design of Fe-V diatomic catalysts for lithium-sulfur batteries, the GNN model, based on atomic graph features (atomic electronegativity, bond length, and bond energy), can accurately predict the conversion efficiency of polysulfides (LiPSs) with a prediction error controlled within 5%. In the LFP cathode material Li... + In the optimization of migration barriers, the random forest model, combined with DFT and MD data, can quickly lock the barrier range corresponding to the optimal preparation parameters, ensuring that the ion transport performance of the material meets the standards and significantly improving the scientificity and accuracy of material design.

[0123] This invention forms a fully intelligent iterative system encompassing "theoretical simulation - high-throughput preparation - high-throughput characterization - model update". In the theoretical simulation stage, the ML model outputs preliminary screening results of material design schemes and process parameters based on multi-scale data. In the high-throughput preparation stage, the reinforcement learning (DQN) agent can collect key parameters such as temperature fluctuations, pH changes, and stirring torque in real time. When parameters deviate from the predicted optimal values, the process parameters are automatically adjusted (e.g., dynamically increasing the precipitant drop rate by 0.5 mL / min when the pH deviates by 0.2) to ensure the stability of the preparation process and the consistency of the product. In the high-throughput characterization stage, a CNN model is used to quickly analyze XRD patterns, completing accurate inversion of lattice constant and grain size within one hour (traditional methods require 24 hours). The multi-task learning model simultaneously inputs XPS elemental content and charge / discharge data to achieve real-time prediction of cycle life and rate performance. Simultaneously, characterization data and performance test results are automatically fed back to the theoretical simulation database, triggering incremental training of the ML model (e.g., adding 100 sets of data can improve prediction accuracy by 3%-5%), forming a self-optimizing R&D closed loop and significantly shortening the iteration cycle from material design to verification.

[0124] This invention establishes an encrypted dedicated database for cathode material genes. The system stores original DFT / MD calculation files (VASPINCAR, LAMMPS input parameters), ML model weight files (random forest.pkl, GNN.pt), high-throughput preparation process parameters (co-precipitation temperature, sintering time, stirring rate), and high-throughput characterization data (XRD patterns, XPS elemental analysis reports, charge-discharge curves). It supports rapid retrieval by multiple dimensions: "material type - performance index - process parameters," with a response time of less than 1 second. Unsupervised learning (k-means clustering) is used to analyze the cyclic decay numbers in the database. According to the classification, the core reasons for the degradation of material performance (such as crystal phase transformation and SEI film rupture) can be accurately identified, with a classification accuracy of over 92%. At the same time, based on the massive data accumulated in the database, data mining methods are developed, which can not only predict the performance of new materials based on the material structure, but also reverse the material structure and preparation process based on the target performance, providing data support for subsequent material performance optimization. For example, based on the XRD and cycling data of LCO cathode materials in the database, the correlation between the peak intensity ratio of (003) / (104) and cycle life can be mined, which can guide the optimization of the crystal plane order of new batches of LCO materials and further improve the cycle stability.

[0125] In the high-throughput preparation stage, this invention, targeting the characteristics of different cathode materials (such as NCM, LFP, and LCO), combines the advantages of preparation processes such as co-precipitation. By optimizing key parameters such as source solution concentration, reaction temperature, pH value, and stirring rate through ML models, materials with specified composition, size, and surface state can be prepared. The products exhibit good reproducibility and high consistency in batch preparation, meeting the requirements of material uniformity for industrial production. In the high-throughput characterization stage, XPS, XRD, X-ray fluorescence, and multifunctional battery testing channels are integrated to simultaneously perform comprehensive detection of material surface composition, crystal structure, elemental distribution, and electrochemical performance (lithium-ion diffusion coefficient, interfacial charge transfer rate, capacity, and potential). Furthermore, by using ML models, the characterization data is correlated with industrial application performance (such as battery pack-level cycle life and rate discharge stability), ensuring that the newly developed materials can be directly adapted to the needs of industrial production, reducing the adaptation cycle and cost of transforming laboratory results into industrial applications.

[0126] This invention employs transfer learning technology to transfer a well-trained ML model (such as the GNN model for the polyethylene pyrolysis system) to the development of new lithium battery materials (such as the glucose catalytic system and lithium-sulfur battery catalyst design). Only a small amount of target system data (approximately 200 sets) is needed to fine-tune the model, significantly reducing the model training cost and data requirements for new system development. Simultaneously, the ML model uses multi-scale physical data as its core feature, rather than material-specific parameters, making it adaptable to the development needs of various lithium battery materials, including NCM, LFP, LCO cathode materials, and lithium-sulfur battery catalysts—for example, for NCM cathodes using Li… + The random forest model for predicting migration barriers can be applied to the performance prediction of LFP cathodes by adjusting only some feature weights; the GNN model for predicting the conversion efficiency of LiPSs in lithium-sulfur batteries can be extended to the performance prediction of other catalytic systems by optimizing the definition of atomic graph nodes and edge features, significantly improving the system's adaptability and flexibility in the research and development of various lithium battery materials.

Claims

1. A simulation system for the research and development of new lithium battery materials, characterized in that: It includes a theoretical simulation calculation module, a machine learning model module, a high-throughput preparation module, a high-throughput characterization module, a closed-loop feedback module, and a dedicated database module for cathode material genes; The theoretical simulation calculation module performs first-principles calculations and molecular dynamics simulations to obtain the atomic-level characteristics and mesoscopic dynamic parameters of lithium battery cathode materials. The machine learning model module constructs and trains a model based on the aforementioned parameters to predict material properties and optimize preparation parameters; The high-throughput preparation module prepares cathode materials according to the optimized parameters output by the model; The high-throughput characterization module collects material structure and electrochemical performance data; The closed-loop feedback module feeds back the characterization data to the theoretical simulation calculation module and the machine learning model module, updating the data and model parameters to form a closed loop for the entire R&D process.

2. The simulation system for the research and development of new lithium battery materials according to claim 1, characterized in that: In the theoretical simulation module, the DFT uses VASP software based on GGA-PBE functionals to calculate the crystal structure, electronic structure, and Li of the cathode material. + Migration barrier; MD uses LAMMPS software to select ReaxFF, PCFF or CVFF force fields depending on the material type, and outputs the ion diffusion coefficient and SEI film growth rate.

3. The simulation system for the research and development of new lithium battery materials according to claim 1, characterized in that: The machine learning model module includes supervised learning, GNN, DQN and unsupervised learning sub-models; The supervised learning sub-model uses the random forest algorithm, taking DFT electronic structure features and MD mesoscopic parameters as input, and outputs Li + Migration barrier; The GNN sub-model is based on the DGL library, takes an atomic graph as input, and outputs polysulfide conversion efficiency or cycle lifetime. Dynamically optimize the preparation parameters of the DQN sub-model; The unsupervised learning sub-model uses k-means clustering to identify the causes of cyclic decay.

4. The simulation system for the research and development of new lithium battery materials according to claim 1, characterized in that: The dedicated database module for cathode material genes is encrypted with AES-256 and stores DFT / MD raw files, machine learning model weight files, preparation process parameters and characterization data. It supports multi-dimensional retrieval and has a response time of less than 1 second.

5. The simulation system for the research and development of new lithium battery materials according to claim 1, characterized in that: The high-throughput preparation module adopts the co-precipitation method, equipped with a multi-channel reactor and a high-temperature sintering furnace. The parameter control range is: source solution concentration 0.8-1.5 mol / L, reaction temperature 40-60℃, pH value 9-11, stirring speed 500-800 rpm, sintering temperature 800-1000℃, holding time 2-6 h, and product particle size variation coefficient ≤8%.

6. The simulation system for the research and development of new lithium battery materials according to claim 1, characterized in that: The high-throughput characterization module includes structural and electrochemical characterization units; The structural characterization units employed XPS, XRD, and X-ray fluorescence. XRD data was analyzed using a CNN model, and the lattice constant and grain size were inverted within one hour, with errors of [missing information]. <5nm; Electrochemical characterization units are tested for charge-discharge and cycle performance in a specific mode.

7. The simulation system for the research and development of new lithium battery materials according to claim 3, characterized in that: The machine learning model module also includes a multi-task learning sub-model, which takes XPS element content and MD diffusion coefficient as input and outputs capacity retention rate and 5C rate capacity retention rate after 1000 cycles with a deviation of ≤3%.

8. The simulation system for the research and development of new lithium battery materials according to claim 3, characterized in that: The GNN sub-model employs transfer learning, transferring the GNN model of the polyethylene pyrolysis system to the design of Fe-V diatomic catalysts for lithium-sulfur batteries. This requires only 200 fine-tuning of the target system atomic diagrams, increasing the convergence speed by 50% and reducing the data requirement by 40%.

9. The simulation system for the research and development of new lithium battery materials according to claim 1, characterized in that: The feedback process of the closed-loop feedback module is as follows: when the deviation between the characterization data and the model prediction value is >5%, the entire sample process data is added to the database. For every 100 new data sets added to the database, incremental training of the model is triggered, improving prediction accuracy by 3%-5%.

10. The simulation system for the research and development of new lithium battery materials according to claim 1, characterized in that: The simulation system is adapted for the research and development of NCM, LFP, LCO cathode materials and Fe-V biatomic catalysts for lithium-sulfur batteries.

Images