A method for designing high-stability multi-element lithium alloy by artificial intelligence

Abstract

The application provides a method for designing high-stability multi-element lithium alloy through artificial intelligence, aiming to optimize the calendar life and interface stability of lithium metal batteries. By constructing a multi-dimensional database containing alloy components, physical and chemical characteristics, and experimental characteristics, machine learning algorithms are used for feature selection and prediction model optimization. In particular, Pearson correlation coefficient and random forest regression are used for preliminary feature selection, combined with forward feature selection optimization to identify key influencing factors. High-throughput calculation is used to predict the calendar life, and the composition space is further optimized through the expected improvement maximization strategy. The model forms a closed loop through continuous iteration and experimental data feedback, improving prediction accuracy and design precision. This design method not only effectively prolongs the calendar life of lithium alloy anodes, but also ensures their high performance and industrial feasibility.

Description

Technical Field

[0001] This invention relates to the field of lithium alloy design technology for lithium batteries, and more specifically to a method for designing highly stable multi-element lithium alloys using artificial intelligence. Background Technology

[0002] Lithium alloy design involves the targeted optimization of mechanical, physical, and chemical properties through the rational selection and control of alloy composition, microstructure, and fabrication processes to meet the complex requirements of high-energy-density lithium metal batteries (LMBs) for industrial applications. Its core objective is to achieve optimal matching under multiple performance constraints, driving technological upgrades in lithium alloy anode materials. However, traditional lithium alloy development methods heavily rely on trial and error, requiring extensive experiments and repeated verification and adjustments to alloy composition and processes to approximate target performance. This approach suffers from inherent drawbacks such as long development cycles, high costs, low efficiency, and difficulty in achieving multi-objective optimization, especially when dealing with high-dimensional compositional spaces and complex structure-property relationships, where its limitations become even more pronounced.

[0003] Lithium metal batteries undergo calendar aging during prolonged periods of quiescent storage, a process that leads to performance degradation. Calendar aging is typically caused by internal chemical reactions that occur even without external circuit current. Despite its crucial impact on battery performance, relevant research is relatively scarce in the literature, particularly in the field of lithium metal batteries. The significant capacity degradation of lithium metal batteries during quiescent storage is primarily due to the persistence and severity of lithium metal interface corrosion. Studies have shown that rationally designed lithium alloys can effectively mitigate this problem, not only improving the uniformity of lithium deposition / stripping processes and suppressing dendrite growth and localized corrosion, but also stabilizing the solid electrolyte interface (SEI), reducing side reactions, and preventing irreversible consumption of active lithium, thereby significantly extending the calendar life of lithium metal batteries. However, calendar aging tests are often lengthy and resource-intensive, severely limiting the widespread implementation of related research.

[0004] In recent years, the rapid development of artificial intelligence and materials computing technologies has provided new opportunities for alloy design. Machine learning, high-throughput computing, and multi-objective optimization algorithms can efficiently uncover the intrinsic relationships between alloy composition, structure, processing, and performance, enabling accurate performance prediction, intelligent component selection, and iterative process optimization. This new approach significantly shortens the R&D cycle, reduces experimental costs, and provides a feasible path for the rational design of complex lithium alloys. Therefore, developing a highly stable multi-element lithium alloy design method based on artificial intelligence has become an urgent technical challenge. This method can break through the bottleneck of traditional trial-and-error models, enabling a shift from experience-based exploration to precise design, thereby meeting the demand for high-performance lithium alloys in new lithium batteries. Summary of the Invention

[0005] To address the problems in existing technologies, such as the reliance on trial and error in lithium alloy design, long R&D cycles, high costs, difficulty in achieving multi-objective synergistic optimization, and difficulties in predicting the calendar life of lithium metal batteries and long experimental verification cycles, this invention provides a highly stable multi-element lithium alloy design method based on artificial intelligence, so as to achieve efficient and accurate design of lithium alloy composition and performance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for designing highly stable multi-element lithium alloys using artificial intelligence includes the following steps: S1. Construct a multi-dimensional relational database of lithium alloys: Experimental and physicochemical data of multi-component lithium alloys were collected to establish a dataset covering "composition-physicochemical characteristics-experimental characteristics"; the composition is mainly lithium, and also contains one or more metallic or non-metallic elements, including Al, Ag, Zn, Mg, Cu, Sn, Na, etc., and non-metallic elements including N, F, Si, etc.

[0007] The physicochemical characteristics include: average atomic radius r, atomic size difference δ r Valence electron concentration VEC, electronegativity difference Δχ, enthalpy of mixing ΔH, entropy of mixing ΔS, free electron concentration e / a, average binding energy Ec, lattice distortion energy Eμ, shear modulus G, melting point T, electron affinity EA, first ionization energy FIE, cohesive energy COH, and work function WF.

[0008] The experimental feature is the calendar life of the lithium alloy battery.

[0009] S2, Secondary Feature Filtering: After feature construction is complete, all features need to be relevance-based to identify highly correlated features and reduce data noise. For this purpose, the Pearson correlation coefficient (PCC) is used, with a threshold of 0.95, to remove one feature from a pair of highly correlated features; if the PCC values ​​of two features are higher than 0.95, the feature with higher importance in the dataset is retained. The importance of two correlated features is calculated using a random forest regressor (RF regressor).

[0010] S3. Model Building and Evaluation: Using alloy composition and physicochemical characteristics as input variables and experimental characteristics as output targets, a machine learning or deep learning prediction model is constructed. The prediction model is selected from at least one of K-Nearest Neighbors (KNN), Gradient Boosting Regression (GBR), Random Forest (RF), Multinomial Kernel Support Vector Regression (SVR-P), and Radial Basis Kernel Support Vector Regression (SVR-R).

[0011] To alleviate the skewness of the data points, the dataset is normalized using standard normalization, which scales and normalizes the feature data of each dimension to a distribution with a mean of 0 and a standard deviation of 1.

[0012] In the model evaluation phase, 5-fold cross-validation (CV) and bootstrapping with replacement are used to evaluate model performance, achieve accurate mapping and quantitative prediction of alloy composition, physicochemical characteristics and experimental characteristics, obtain a high-confidence performance prediction model, and obtain the uncertainty distribution of the predicted values.

[0013] Model evaluation and feature selection employ two error metrics: the coefficient of determination (R²). 2 ) and root mean square error (RMSE).

[0014] After the first round of feature removal using Pearson correlation coefficient (PCC) combined with random forest (RF), the forward feature selection (FFS) algorithm is further employed to more efficiently select the optimal subset of features that can describe the target variable (calendar life).

[0015] The forward feature selection (FFS) process is as follows: First, each feature is selected individually to train the model, and the single feature that gives the highest cross-validation score is selected as the starting point; then, the remaining features are added sequentially, and at each step, the feature that maximizes the cross-validation score is retained.

[0016] S4. Multi-objective optimization and constraint setting: With the goal of maximizing calendar life, engineering indicators such as alloy melting point, structural stability, element cost, preparation feasibility and phase composition controllability are introduced as constraints to exclude composition ranges that are not industrializable or thermodynamically unstable.

[0017] S5. Inverse Component Design and Screening: A virtual composition space for multi-component lithium alloys containing more than 700,000 different components was constructed.

[0018] The virtual component space is constructed within the following range: the Li content varies in increments of 1 at% within the range of 50–99 at%, and one or more elements among Al, Ag, Zn, Mg, Cu, Sn, Na, N, F, and Si vary in increments of 1 at% within the range of 0–50 at%.

[0019] Subsequently, all components in this virtual component space are input into the best-performing machine learning model to predict calendar lifetimes. After the prediction is complete, all results are sorted from longest to shortest calendar lifetime.

[0020] Finally, an expected improvement (EI) maximization strategy is adopted to select the optimal candidate alloy in the virtual composition space from all high prediction results.

[0021] The formula for calculating the expected improvement maximization strategy is as follows: ； Where φ(z) is the standard normal density function, and Φ(z) is the cumulative distribution function; (y-μ) * P(y|x′) represents the potential increase in the predicted value; P(y|x′) represents the distribution of the predicted value y (assuming the predicted value y, calendar lifetime, follows a normal distribution); z=(μ-μ * ) / σ, where μ * The maximum value of the calendar lifetime in the training set is given by μ, and μ and σ are the predicted mean and standard deviation, respectively.

[0022] Finally, after sorting the alloys according to their EI values, the optimal candidate alloys were selected.

[0023] S6. Experimental Verification and Feedback: Multi-element lithium alloy samples were prepared according to the screening results using a melting-rolling method. Electrochemical tests were then performed on the samples to obtain their actual calendar lifetime.

[0024] The positive electrode was NCM811, and the negative electrode was a multi-element lithium alloy sample (thickness 60μm). The electrolyte used was 1.2M lithium hexafluorophosphate (LiPF6) dissolved in ethylene carbonate (EC) and ethyl methyl carbonate (EMC) (mass ratio 3:7), with 2% vinylene carbonate (VC) added as an additive.

[0025] For the fabrication of coin cells (CR2032), all samples must maintain consistent electrode fabrication processes, electrolyte usage, and packaging techniques to avoid individual variations affecting test results. At least three parallel samples should be prepared for each test condition to ensure data repeatability.

[0026] All batteries were formed in a constant temperature environment of 25℃ using constant current and constant voltage (CC-CV) mode. The specific parameters were: constant current charging at 0.1C to the cutoff voltage (4.35V), constant voltage charging until the current drops to 0.05C, and resting for 1 hour; then constant current discharging at 0.1C to the cutoff voltage (3.0V), repeating 2 cycles; finally, the batteries were charged at 0.1C to 100% SOC, and the initial OCV value was measured after resting for 1 hour.

[0027] During open-circuit storage (25℃), the decay rate of open-circuit voltage (OCV) is used as the criterion for calendar life: when the battery OCV drops by more than 0.2 V from the initial value, or the daily average decay rate is greater than 0.002 V / d, the battery is judged to have failed the calendar life.

[0028] Experimental verification data is fed back to the database to correct, retrain, and update the artificial intelligence prediction model and optimization objectives, forming a closed-loop design system of data-model-design-preparation-verification-feedback, which continuously improves the design accuracy and reliability of multi-element lithium alloys.

[0029] Compared with existing technologies, this invention, based on artificial intelligence, constructs a multi-element lithium alloy design method, which has significant advantages in design concept, optimization path, and engineering implementation. Its beneficial effects and principles are explained below: 1. This invention constructs a multidimensional database of "composition-physicochemical characteristics-performance" and utilizes machine learning models to establish nonlinear mapping relationships, essentially approximating complex structure-property relationships as high-dimensional functions. Compared to traditional methods that rely on experience and experimental iterations, it can quickly locate the region of optimal performance in a high-dimensional composition space, thereby significantly shortening the R&D cycle and reducing experimental costs.

[0030] 2. This invention uses Pearson correlation coefficient (PCC) and random forest importance assessment for initial screening, followed by forward feature selection (FFS) for secondary optimization. This approach eliminates redundant features from two dimensions: statistical relevance and model contribution. The principle behind this is to reduce multicollinearity among input variables, decrease the risk of model overfitting, and retain intrinsic descriptive parameters that have a key impact on the target variable (calendar lifetime), thereby improving prediction stability and interpretability.

[0031] 3. In the reverse engineering stage, this invention not only relies on the predicted mean but also obtains the predicted distribution through bootstrapping, thereby quantifying model uncertainty and introducing the Expected Improvement (EI) function for candidate selection. Essentially, within the Bayesian optimization framework, it simultaneously considers "exploitation" and "exploration," avoiding local optima and overfitting problems caused by solely relying on point predictions, and improving the global optimal probability of candidate alloys.

[0032] 4. This invention expands the optimization objective from a single calendar lifetime to multi-objective collaborative optimization, introducing constraints such as melting point, structural stability, elemental cost, preparation feasibility, and phase composition controllability. Its principle lies in embedding thermodynamic stability (such as mixing enthalpy and phase stability range) and engineering constraints (such as cost and processing window) into the optimization process, achieving integrated "performance-manufacturing-cost" design, thus avoiding obtaining material solutions that are only theoretically excellent but cannot be industrialized.

[0033] 5. This invention constructs a closed-loop system of "data-model-design-preparation-verification-feedback," continuously feeding experimental results back into model training. Its essence is an active learning process, enabling the model to be continuously corrected and strengthened in key data areas, thereby gradually improving prediction accuracy and design reliability, and achieving long-term self-optimization capabilities for complex material systems.

[0034] 6. This method does not depend on a specific alloy system and can be extended to different element combinations and multi-element systems through feature engineering and model training. It can effectively address the combinatorial explosion problem caused by high-dimensional composition space and provide a general technical path for the design of new high-performance lithium alloys and even other functional materials.

[0035] In summary, this invention has significant advantages in both theoretical mechanism and engineering implementation. It not only improves the efficiency and accuracy of lithium alloy design, but also takes into account performance and industrial feasibility, and plays an important supporting role in the practical application of high-energy-density lithium metal batteries. Detailed Implementation

[0036] 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.

[0037] Example 1: A Li-based approach based on low-cost constraints 92 Mg6Al2 ternary lithium alloy.

[0038] This embodiment aims to provide a lithium alloy anode material that balances low cost, structural stability, and interface stability, making it suitable for large-scale industrial production scenarios.

[0039] During the AI ​​screening phase, the following engineering constraints are set: the formation energy (ΔH) is negative and the absolute value is moderate (to avoid brittle intermetallic compounds).

[0040] Initial feature screening was performed using Pearson correlation coefficient (PCC) and random forest regression, followed by further optimization with forward feature selection (FFS). The model identified the key features affecting calendar lifetime as: mixture entropy (ΔS) and atomic size difference (δ). r The mechanism is that a moderate atomic size difference can introduce controllable lattice distortion and suppress local stress concentration; a higher mixing entropy helps to improve phase stability and suppress element segregation.

[0041] The prediction results show that the predicted calendar lifetime is approximately 550 days, and the prediction uncertainty σ is low, located in the high-density region of the training data, indicating that this component belongs to the high-confidence optimization region (exploitation-dominated).

[0042] Multi-element lithium alloy samples were prepared according to the screening results using a melting-rolling method. Electrochemical tests were then performed on the samples to obtain their actual calendar lifetimes. A medium-frequency induction furnace was used for melting, and a precision twin-roll cold rolling mill was used for rolling, with a thickness control accuracy of ±1 μm.

[0043] Specifically, Li (99.9%), Mg (99.99%), and Al (99.99%) are weighed in an argon glove box and mixed according to the atomic ratio Li:Mg:Al = 92:6:2.

[0044] Place the above raw materials in a stainless steel crucible under a high-purity argon atmosphere (99.999%). Heating process: 300–350℃: melt lithium, finally raise to 550℃, hold for 30–60 minutes to avoid excessive temperature causing Li volatilization, and use electromagnetic stirring to achieve homogenization of composition.

[0045] Molten alloy is poured into a mold to obtain a slab, which is then subjected to multi-pass precision rolling at room temperature: single-pass reduction: 8–12%, with short-term annealing allowed in between, with an annealing temperature ≤120°C to prevent cracking, ultimately yielding an alloy foil with a thickness of approximately 40 μm.

[0046] Assemble CR2032 button cells: The positive electrode uses NCM811, and the electrolyte is 1.2M LiPF6 / EC:EMC (3:7) + 2%VC.

[0047] Test procedure: The system is placed at 25°C and the OCV is recorded every 24 hours. The OCV decay rate is used as the calendar lifetime characterization index.

[0048] Test results show that the average OCV decay rate is 0.0003 V / d, significantly lower than that of the pure lithium control (typically around 0.001–0.002 V / d). Based on an empirical conversion model (based on a failure threshold of 0.2 V), the corresponding calendar lifetime is approximately 520–580 days. This deviates from the predicted 550 days by about 4.2%, indicating that the model prediction has high accuracy.

[0049] Example 2: A Lithium-based ... 96 Ag 1.5 Zn 2.5 Multi-component lithium alloy. This embodiment aims to provide a lithium alloy anode material that balances high lithium affinity, good interfacial stability, and long calendar life, making it particularly suitable for improving the long-term stability and safety of lithium metal batteries.

[0050] The following engineering constraints are set: the goal is to improve the uniformity of the lithium deposition / stripping process and extend calendar lifetime.

[0051] Initial feature screening was performed using Pearson correlation coefficient (PCC) and random forest regression (RF), followed by further optimization using forward feature selection (FFS). The model identified the key features affecting calendar lifetime as: electron affinity (EA) and lattice distortion energy (Eμ). This is mainly because Ag and Zn elements can form highly lithiophilic sites, reducing lithium deposition overpotential and promoting uniform lithium deposition. At the same time, lower lattice distortion energy helps improve the long-term stability of the structure.

[0052] The prediction results show that the predicted calendar lifetime is approximately 720 days, and the prediction uncertainty σ is low, indicating that this component belongs to the high-confidence optimization region (exploitation-dominated).

[0053] Weigh out Li (99.9%), Ag (99.99%), and Zn (99.99%) in an argon glove box and mix them in an atomic ratio of Li:Ag:Zn = 96:1.5:2.5.

[0054] Place the above raw materials in a stainless steel crucible under a high-purity argon atmosphere (99.999%). Heating process: 300–350℃: melt lithium, finally raise to 550℃, hold for 30–60 minutes to avoid excessive temperature causing Li volatilization, and use electromagnetic stirring to achieve homogenization of composition.

[0055] Molten alloy is poured into a mold to obtain a slab, which is then subjected to multi-pass precision rolling at room temperature: single-pass reduction: 8–12%, with short-term annealing allowed in between, with an annealing temperature ≤120°C to prevent cracking, ultimately yielding an alloy foil with a thickness of approximately 40 μm.

[0056] Battery assembly and testing: NCM811 was selected for the positive electrode, and the electrolyte was 1.2MLiPF6 / EC:EMC (3:7) + 2%VC. The battery was assembled into a CR2032 coin cell.

[0057] OCV tests were conducted at 25°C with the system left to stand, and data was recorded every 24 hours. The OCV decay rate was used as the characterization index for calendar lifetime.

[0058] After 180 days of testing, the average OCV decay rate was 0.00012V / d. In the control experiment, the OCV decay rate of the pure lithium alloy under the same conditions was 0.0012V / d, which translates to a calendar life of approximately 720 days. The deviation from the predicted value was only 4.2%, proving that the prediction accuracy was high. The experiment also verified the high lithium affinity and excellent interface stability of the alloy.

[0059] 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. A method for designing highly stable multi-element lithium alloys using artificial intelligence, characterized in that, Includes the following steps: S1. Construct a multi-dimensional correlation database for multi-element lithium alloys: collect experimental and physicochemical data of multi-element lithium alloys, and establish a dataset covering "composition-physicochemical characteristics-experimental characteristics"; the composition is mainly lithium, and the experimental characteristics are the calendar life of lithium alloy batteries. S2, Secondary Feature Selection: First, the Pearson correlation coefficient (PCC) combined with the importance assessment of random forest (RF) is used for the first round of feature elimination; then, the forward feature selection (FFS) algorithm is used to select the optimal feature subset; S3. Model Building and Evaluation: Establish a predictive model, use cross-validation and bootstrapping to quantitatively predict calendar lifetimes, and obtain the uncertainty distribution of the predicted values; S4. Multi-objective optimization and constraint setting: The goal is to maximize calendar lifetime, and engineering indicators such as alloy melting point, structural stability, element cost, preparation feasibility and phase composition controllability are introduced as constraints. S5. Inverse composition design and screening: Construct a virtual composition space for multi-element lithium alloys, input the composition in the space into the prediction model, and use the expected improvement (EI) maximization strategy to screen candidate alloys; S6. Experimental Verification and Feedback: Alloys are prepared according to the screening results and electrochemical tests are conducted. The actual calendar lifetime data obtained is fed back to the database to realize the closed-loop iterative update of the model.

2. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, The physicochemical characteristics described in step S1 include: average atomic radius, atomic size difference, valence electron concentration, electronegativity difference, enthalpy of mixing, entropy of mixing, free electron concentration, average binding energy, lattice distortion energy, as well as shear modulus, melting point, electron affinity, first ionization energy, cohesive energy, and work function.

3. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, In addition to lithium, the components mentioned in step S1 also include one or more of Al, Ag, Zn, Mg, Cu, Sn, Na, N, F, and Si.

4. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, The specific logic of the first round of feature removal in step S2 is as follows: set the PCC threshold to 0.

95. When the correlation coefficient between two features is higher than this threshold, use a random forest regressor to calculate the feature importance and retain the feature with higher importance.

5. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, The prediction model mentioned in step S3 is selected from at least one of K-Nearest Neighbors (KNN), Gradient Boosting Regression (GBR), Random Forest (RF), Multinomial Kernel Support Vector Regression (SVR-P), and Radial Basis Kernel Support Vector Regression (SVR-R).

6. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, In step S3, the dataset is first normalized to scale the feature data of each dimension to a distribution with a mean of 0 and a standard deviation of 1.

7. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, In step S3, the coefficient of determination R is used. 2 The root mean square error (RMSE) is used as an evaluation metric for model performance.

8. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, The construction range of the virtual component space mentioned in step S5 is as follows: the Li content varies in increments of 1 at% within the range of 50 to 99 at%, and one or more elements among Al, Ag, Zn, Mg, Cu, Sn, Na, N, F, and Si vary in increments of 1 at% within the range of 0 to 50 at%.

9. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, The formula for calculating the expected improvement maximization strategy in step S5 is EI=σ[φ(z)+zΦ(z)], Where z = (μ - μ) * ) / σ,μ * Let μ be the maximum value of the calendar lifetime in the training set, μ be the predicted mean, σ be the predicted standard deviation, φ(z) be the standard normal density function, and Φ(z) be the cumulative integral function.

10. The method for designing highly stable multi-element lithium alloys using artificial intelligence according to claim 1, characterized in that, The electrochemical test described in step S6 uses the open-circuit voltage (OCV) decay rate as the calendar life criterion: when the battery OCV drops by more than 0.2 V from the initial value, or the daily average decay rate is greater than 0.002 V / d, the battery is determined to have failed the calendar life test.