Artificial intelligence-based Li7P3S11-doped solid electrolyte design and optimization method

Abstract

The invention provides a Li7P3S11 doped solid electrolyte design and optimization method based on artificial intelligence. According to the method, at least 50 variant crystal structures containing Cl, Br and other doped elements are generated through a GAN or diffusion model; synchronously predicting an ionic conductivity sigma and an air stability score S by using a GNN model of a CGCNN or M3GNet architecture; and screening at least five optimal structures according to a'double-high 'principle for experimental synthesis, updating the model through data closed-loop feedback, and achieving a performance target through 3-5 rounds of iteration. The method breaks through the limitation of the traditional technology, is high in prediction precision, and shortens the research and development period from 3-5 years to 6 months. The room temperature ionic conductivity gt of the prepared solid electrolyte; the capacity retention rate gt after being exposed for 48 hours at 25 DEG C under the conditions of 20 mS / cm, 50% relative humidity and 20 mS / cm; the method is suitable for all-solid-state lithium batteries, and the performance of the batteries is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of solid-state battery materials technology, specifically to an artificial intelligence-based doped Li7P3S... 11 Solid electrolyte design and optimization methods. Background Technology

[0002] Solid-state batteries, with their higher energy density and superior safety, have become a research hotspot and development direction in the field of new energy batteries. Among them, sulfide solid electrolytes are considered one of the key materials for realizing high-performance all-solid-state batteries due to their high ionic conductivity.

[0003] Li7P3S 11 As a typical sulfide solid electrolyte, Li7P3S exhibits a room-temperature ionic conductivity of approximately 17 mS / cm, far exceeding that of mainstream oxide electrolytes (such as LLZO, which has a conductivity of only 0.5 mS / cm), theoretically possessing broad application prospects. However, Li7P3S... 11 The key technical drawback is its extreme air sensitivity. When exposed to an environment with a relative humidity greater than 30% for more than 2 hours, its surface will rapidly react to form insulating layers such as Li2S and H3PO4, causing a sharp drop in ionic conductivity, which in turn leads to catastrophic failure of battery performance and severely limits its practical application.

[0004] To address the aforementioned issues, various technical solutions have been proposed in related fields. Institutions such as Toyota and Samsung SDI have employed materials such as Al2O3 or LiNbO3 to synthesize Li7P3S... 11 Surface coating protection is an option, but it only provides surface protection and cannot address the material's stability at its core. Furthermore, the coating process is complex, costly, and may introduce additional interfacial impedance. The MIT Materials Project proposed a high-throughput DFT screening method, but it can only screen known structures, cannot create new structures, and has high computational costs. Most laboratories employ traditional trial-and-error methods for doping research, which are inefficient, have a large parameter space, and require 3-5 years of development time. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose an artificial intelligence-based method for doping Li7P3S. 11 Solid electrolyte design and optimization methods.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an artificial intelligence-based doped Li7P3S 11 Solid electrolyte design and optimization methods include the following steps: S1: Using a generative AI model, create variant crystal structures of doped elements: Using a generative AI model, input Li7P3S 11 Based on the crystal structure data, at least 50 variant crystal structures with doped elements are generated; the doped elements are selected from halogens (Cl, Br, I) and metals (Ge, Ti, Sn), with a doping concentration of 0.1%-5% atomic ratio; The generative AI model is a branch of artificial intelligence technology based on machine learning models, which generates new content by mimicking data distribution; The generative AI model selected is either a Generative Adversarial Network (GAN) model or a diffusion model. The adversarial network GAN model includes a generator G, a discriminator D, and a loss function, and the loss function includes physical constraint terms to ensure reasonable atomic spacing and charge balance; S2: Use the Graph Neural Network (GNN) model to predict key performance indicators across multiple tasks; The crystal graph convolutional neural network (CGCNN) model or the M3GNet architecture in the graph neural network (GNN) model is selected to convert the variant crystal structure generated in step S1 into a graph structure representation; the graph structure representation uses atoms as nodes and the inter-atomic connections as edges. The crystal graph convolutional neural network CGCNN model adopts a multi-task learning architecture, which simultaneously predicts two key performance indicators, avoiding the efficiency loss of training a single-task model separately. The key performance indicators include ionic conductivity σ and air stability score S; S3: Synthesis and performance testing of optimal candidate experiments; Based on the ionic conductivity σ and air stability score S predicted in step S2, at least 5 variant crystal structures of doped elements with the best performance were selected according to the dual principle of "high ionic conductivity σ + high air stability score S", and then experimentally synthesized and tested. S4: Closed-loop feedback of experimental data and model update; The experimental data obtained in step S3 is fed back into the model in steps S1 and S2 to retrain and update the parameters of the model. The experimental data from step S3 are cleaned to remove outliers, including performance deviations caused by errors in the synthesis operation, and retain valid data, including synthesis parameters, measured ionic conductivity σ, measured air stability score S, and prediction error. The effective experimental data is used as new training samples to supplement the training sets of the generative AI model in step S1 and the graph neural network (GNN) model in step S2, and the models are retrained and the parameters are updated. S5: Multiple iterations until the performance target is achieved; Repeat steps S1 to S4 for at least 2 rounds, with 3-5 rounds being the preferred option, to continuously optimize the model's predictive ability and material properties through multiple iterations.

[0007] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention employs a generative AI model (GAN or diffusion model), which can extract data from Li7P3S... 11 Starting from the baseline crystal structure, novel and undiscovered doped variant crystal structures are generated, breaking through the limitation of existing technologies that can only screen known structures and greatly expanding the space for materials research and development. Furthermore, by utilizing the multi-task learning architecture of graph neural networks (GNN), the ionic conductivity and air stability are predicted simultaneously, with an efficiency improvement of more than 50% compared to training two independent single-task models, and high prediction accuracy (σ prediction MAE < 2 mS / cm; S prediction AUC > 0.85), ensuring the reliability of the screening.

[0008] Secondly, this invention constructs a closed-loop optimization system of "generation, prediction, experiment, verification, feedback, and iteration". Through active learning and adaptive weight adjustment mechanisms, the prediction accuracy of the AI ​​model is continuously improved. After 3-5 rounds of iteration, the prediction accuracy can be improved to over 85%, significantly shortening the R&D cycle from the traditional 3-5 years to 6 months. Furthermore, this invention optimizes the bulk properties of the material through the rational selection and concentration control of doping elements, resulting in doped Li7P3S 11 Solid electrolytes combine high ionic conductivity (>20 mS / cm, preferably >22 mS / cm) with excellent air stability (>80% capacity retention after 48 hours, preferably >90% capacity retention after 72 hours), eliminating the need for surface coatings and avoiding problems such as increased interfacial impedance, complex processes, and high costs associated with coatings. The method of this invention has good repeatability and scalability. The generative AI model and GNN model can be updated and optimized by continuously adding new experimental data. It is applicable to the study of various doping elements and different doping ratios, providing a new technical path for the efficient development of sulfide solid electrolytes. At the same time, the solid electrolyte prepared based on the method of this invention has excellent performance. When applied to all-solid-state lithium batteries, it can significantly improve the energy density, cycle stability and safety of the battery, and has broad market application prospects. The corresponding intelligent design and manufacturing system realizes the automated closed loop of material design, synthesis, testing and optimization, further improving R&D efficiency and reducing R&D costs. Attached Figure Description

[0009] Figure 1 AI-driven doped Li7P3S of this invention 11 A flowchart illustrating the solid electrolyte generation design and closed-loop optimization method; Figure 2 For the Li7P3S in Embodiment 1 of the present invention 10.8 Cl 0.2 With pure Li7P3S 11 A comparison chart of capacity retention rates; Figure 3 This is a graph showing the trend of ionic conductivity variation of the optimal sample during the three iterations in Embodiment 2 of the present invention. Detailed Implementation

[0010] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0011] Example 1: AI-generated Li7P3S 10.8 Cl 0.2 And verify: 1. AI Generation and Prediction: Generative Adversarial Networks (GANs) are used as generative AI models, with Li7P3S as the input. 11 Based on the reference crystal structure data (CIF file, space group P-1), Cl was selected as the dopant element to generate Li7P3S. 11-X Cl X A series of 100 candidate structures (x=0.1, 0.2, 0.5, 1.0) were generated. A crystal graph convolutional neural network (CGCNN) was used as a graph neural network (GNN) to convert the generated candidate structures into graph structure representations (node ​​features include atomic number, ionic radius, electronegativity and oxidation state; edge features are atomic spacing). A multi-task learning architecture was adopted, and the loss function was L= MSE(σ) +BCE(S), which simultaneously predicted the ionic conductivity σ and the air stability score S.

[0012] Prediction results show that Li7P3S 10.8 Cl 0.2 The best performance is achieved, with σ prediction = 22.5 mS / cm and S = 0.87.

[0013] Experimental synthesis: Experimental synthesis was performed using mechanical ball milling: Li₂S, P₂S₅, and LiCl raw materials were accurately weighed in an argon glove box, following the method described for Li₇P₃S… 10.8 Cl 0.2 The stoichiometric ratios were mixed and placed in a planetary ball mill, where the mixture was milled at 500 rpm for 8 hours. The milled product was then placed in a quartz boat and annealed at 300°C for 2 hours under an argon atmosphere to obtain the target product.

[0014] Performance testing: Ionic conductivity test: Electrochemical impedance spectroscopy (EIS) was used for testing at room temperature (25°C) and a frequency range of 0.1Hz–1MHz. The measured ionic conductivity σ = 23.1 mS / cm. Air stability test: The sample was placed in an environment with 50% relative humidity and 25°C for air exposure experiments, and the capacity retention rate was tested at different time points.

[0015] The results showed that after 168 hours of exposure, the sample volume retention rate still reached 94%. Comparative experiment: Pure Li7P3S 11 After 72 hours of exposure under the same conditions, the volume retention rate was <60%.

[0016] Conclusion: The AI ​​predictions and experimental results show high consistency (error <3%). The Li7P3S prepared in this embodiment... 10.8 Cl 0.2 Solid electrolytes are significantly superior to pure Li7P3S in both conductivity and air stability. 11 .

[0017] Example 2: Closed-loop optimization results after 3 rounds of iteration: Statistical analysis of the iterations yields the following results: 1. First iteration: The generative AI model (GAN) generates 100 candidate structures. The GNN predicts and selects 10 doped variants with the best performance for experimental verification. The optimal sample composition is Li7P3S. 10.8 Cl 0.2 The measured ionic conductivity was 23.1 mS / cm, the prediction error was ±2.3 mS / cm, and the air stability score (measured) was 0.94. The experimental data were fed back to the AI ​​model to update the model parameters.

[0018] 2. Second iteration: The generative AI model generates 150 candidate structures based on the updated parameters, selects 10 for experimental verification, and the optimal sample composition is Li7P. 2.95 Ti 0.05 S 10.9 Br 0.1 The measured ionic conductivity was 24.8 mS / cm, the prediction error was ±1.5 mS / cm, and the air stability score (measured) was 0.91. The experimental data were fed back again to update the model parameters.

[0019] 3. Third iteration: The generative AI model generated 200 candidate structures, selected 8 for experimental verification, and the optimal sample composition was Li7P. 2.9 Ge 0.1 S 10.8 Cl 0.2 The measured ionic conductivity was 26.3 mS / cm, the prediction error was ±0.8 mS / cm, and the air stability score (measured) was 0.96.

[0020] Observation results: AI accuracy improved: After three rounds of closed-loop feedback, the GNN prediction error decreased from ±2.3 mS / cm to ±0.8 mS / cm, and the prediction accuracy was significantly improved; Performance was optimized round by round: the ionic conductivity of the optimal sample was gradually increased from 23.1 mS / cm to 26.3 mS / cm, while the air stability remained at a high level; Synergistic effect: Co-doping with Ge (P-site) and Cl (S-site) exhibits a synergistic enhancement effect. Analysis shows that Ge substitution at the P-site reduces the Li... + The migration barrier, with Cl substitution at the S site improving grain boundary stability and broadening the Li migration barrier, was addressed. + aisle.

[0021] Example 3: Comparative Experiment (Proving Superiority): Select pure Li7P3S 11 A performance comparison was made between the Cl-doped sample from the literature, the Toyota Al2O3 coated sample, and the sample prepared by the AI ​​method of this invention. The performance of the doped Li7P3S prepared by the AI ​​method of this invention was compared. 11 Solid electrolytes are significantly superior to existing technologies in terms of ionic conductivity and air stability, and their development cycle is greatly shortened, demonstrating obvious technological advantages and application value.

[0022] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. An AI-based doped Li7P3S 11 Solid electrolyte design and optimization method, characterized by: Includes the following steps: S1: Using a generative AI model, create variant crystal structures of doped elements: Using a generative AI model, input Li7P3S 11 Based on the crystal structure data, at least 50 variant crystal structures with doped elements are generated; the doped elements are selected from halogens (Cl, Br, I) and metals (Ge, Ti, Sn), with a doping concentration of 0.1%-5% atomic ratio; The generative AI model is a branch of artificial intelligence technology based on machine learning models, which generates new content by mimicking data distribution; The generative AI model selected is either a Generative Adversarial Network (GAN) model or a diffusion model. The adversarial network GAN model includes a generator G, a discriminator D, and a loss function, and the loss function includes physical constraint terms to ensure reasonable atomic spacing and charge balance; S2: Use the Graph Neural Network (GNN) model to predict key performance indicators across multiple tasks; The crystal graph convolutional neural network (CGCNN) model or the M3GNet architecture in the graph neural network (GNN) model is selected to convert the variant crystal structure generated in step S1 into a graph structure representation; the graph structure representation uses atoms as nodes and the inter-atomic connections as edges. The crystal graph convolutional neural network CGCNN model adopts a multi-task learning architecture, which simultaneously predicts two key performance indicators, avoiding the efficiency loss of training a single-task model separately. The key performance indicators include ionic conductivity σ and air stability score S; S3: Synthesis and performance testing of optimal candidate experiments; Based on the predicted ionic conductivity σ and air stability score S obtained in step S2, at least 5 variant crystal structures of doped elements with the best performance were selected according to the dual-excellence principle of "high ionic conductivity σ + high air stability score S", and then experimentally synthesized and tested. S4: Closed-loop feedback of experimental data and model update; The experimental data obtained in step S3 is fed back into the model in steps S1 and S2 to retrain and update the parameters of the model. The experimental data from step S3 are cleaned to remove outliers, including performance deviations caused by errors in the synthesis operation, and retain valid data, including synthesis parameters, measured ionic conductivity σ, measured air stability score S, and prediction error. The effective experimental data is used as new training samples to supplement the training sets of the generative AI model in step S1 and the graph neural network (GNN) model in step S2, and the models are retrained and the parameters are updated. S5: Multiple iterations until the performance target is achieved; Repeat steps S1 to S4 for at least 2 rounds, with 3-5 rounds being the preferred option, to continuously optimize the model's predictive ability and material properties through multiple iterations.

2. The artificial intelligence-based doped Li7P3S as described in claim 1 11 Solid electrolyte design and optimization method, characterized by: The loss function is L=MSE(σ)+BCE(S), which combines the loss values ​​of the two predicted key performance indicators. MSE(σ) is the mean square error of the ionic conductivity prediction, and BCE(S) is the binary cross-entropy of the air stability score prediction. The model parameters are optimized through this function to improve the prediction accuracy.

3. The artificial intelligence-based doped Li7P3S as described in claim 1 11 Solid electrolyte design and optimization method, characterized by: In step S2, the specific steps for predicting key performance indicators are as follows: S21: Data Preprocessing and Model Training S211: Data Preprocessing: The prerequisite for prediction is having a high-quality labeled dataset for training, validating, and optimizing the GNN model. The data processing flow is as follows: Integrating three core datasets: the PNNL sulfide electrolyte dataset, data from the Materials Projec open-source materials database, and laboratory-tested doped Li7P3S. 11 Sample data, including crystal structure and measured performance; Invalid data, including samples with incomplete structure and abnormal performance test results, are removed. Samples with clear crystal structure parameters, namely atomic type, coordinates, lattice parameters, dopant element / concentration / site, and corresponding measured performance, namely ionic conductivity σ and air stability score S, are retained to form a dataset of no less than 300 valid samples. S212: GNN Model Construction and Training: The crystal diagram convolutional neural network (CGCNN) or M3GNet architecture is selected, and the model construction and training steps are as follows: S2121: Graph Structure Transformation: Transform each crystal structure in the dataset into an input graph for the GNN. Node definition: Each atom is a node, and the node feature vector contains four core types of information: atomic number (Li=3, P=15, S=16, Cl=17), ionic radius, electronegativity, and oxidation state (Li). + =+1、P 5+ =+5、S 2- =-2); Edge definition: When the distance between two atoms is less than a preset threshold, usually 1.2 times the sum of the covalent radii of atoms in the crystal, it is defined as the existence of an edge connection. The edge characteristic is the actual distance between the two atoms. S2122: The model output layer is designed with a dual-branch structure to simultaneously learn the predictive mapping relationship between ionic conductivity σ and air stability score S; S2123: Loss function definition: The model is optimized using a combined loss function to balance the prediction accuracy of the two indicators. Loss term for ionic conductivity σ: mean square error (MSE). The formula is MSE(σ) = (1 / n)×Σ(σ) pred -σ true )², Where, σ pred For the predicted value, σ true The values ​​are the measured values, and n is the sample size. The loss term for the air stability score S is the binary cross-entropy (BCE). The formula is BCE(S) = -Σ[S] true ×ln(S pred )+(1-S true )×ln (1-S pred )]; Among them, S pred S is the predicted value. true These are measured values; Total loss function: L = MSE(σ) + BCE(S), minimize the total loss value through backpropagation; S2124: Training parameter configuration: The optimizer is Adam, the initial learning rate is set to 1e-3, and dynamically adjusted according to the performance of the validation set; the number of training epochs is set to 500-1000 to avoid overfitting; S2125: Model Accuracy Validation: After training, evaluate model performance using the test set. The mean absolute error (MAE) of the σ prediction curve must be <3 mS / cm (preferably <2 mS / cm), and the area under the S prediction curve (AUC) must be >0.

85. Only after meeting these accuracy requirements can the prediction be used for candidate structures.

4. The artificial intelligence-based doped Li7P3S as described in claim 1 11 Solid electrolyte design and optimization method, characterized by: The generative artificial intelligence model in step S1 is a diffusion model, which generates new doped variant structures by adding noise to the baseline structure and gradually removing the noise.

5. The artificial intelligence-based doped Li7P3S as described in claim 1 11 Solid electrolyte design and optimization method, characterized by: The experimental synthesis in step S3 was carried out using the sol-gel method or mechanical ball milling, and annealed at a temperature of 250-350°C.

6. The artificial intelligence-based doped Li7P3S as described in claim 1 11 Solid electrolyte design and optimization method, characterized by: The method also includes an interpretability analysis step: using SHAP values, LIME, or attention mechanisms to identify the contribution of dopant elements and sites to performance, in order to guide subsequent optimization directions.

7. A doped Li7P3S 11 Solid electrolyte, characterized in that: Prepared by the method according to any one of claims 1 to 6, satisfying the following: room temperature ionic conductivity > 20 mS / cm; capacity retention > 80% after exposure at 50% relative humidity and 25°C for > 48 hours; the doping element is selected from at least one of Cl, Br, I, Ge, Ti, and Sn, with a doping concentration of 0.1%-5%.

8. A fully solid-state lithium battery, characterized in that: It includes the solid electrolyte as described in claim 7 as an ion-conducting layer.

9. A solid electrolyte intelligent design and manufacturing system, integrating the AI ​​optimization method according to any one of claims 1-6, characterized in that: include: The system includes an AI model training and generation module, an automated synthesis equipment interface, an automatic performance test data acquisition module, and a closed-loop feedback control system for updating the AI ​​model.

Images