Titanium-manganese-based hydrogen storage alloy design method and system based on machine learning

Abstract

This invention belongs to the fields of materials science and artificial intelligence, specifically a machine learning-based design method and system for titanium-manganese hydrogen storage alloys. The method involves acquiring experimental data tagged with chemical composition, hydrogen storage performance, and spontaneous combustion characteristics; constructing a key feature pool including A / B atomic ratio and Cr / Mn atomic ratio, and selecting a subset of key features; using the subset of key features as input and hydrogen storage performance and spontaneous combustion characteristics as output to construct a multi-task learning model; employing a multi-objective optimization algorithm to optimize the model under constraints, aiming to maximize hydrogen storage performance and non-spontaneous combustion probability, and outputting the Pareto optimal alloy composition as a candidate formulation; preparing the alloy according to the candidate formulation and testing its actual performance, and feeding back the experimental results to update the experimental data for model iteration. By using "non-spontaneous combustion probability" as a parallel optimization objective, and combining data-driven and mechanism-constrained approaches, the method achieves synergistic optimization of hydrogen storage performance and safety in titanium-manganese hydrogen storage alloys, significantly reducing R&D costs.

Description

Technical Field

[0001] This invention belongs to the fields of materials science and artificial intelligence, specifically a design method and system for titanium-manganese hydrogen storage alloys based on machine learning. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] The core bottleneck for the large-scale application of hydrogen energy lies in safe and efficient hydrogen storage technology. Among solid-state hydrogen storage, titanium-manganese hydrogen storage alloys, with their hydrogen storage capacity of approximately 2 wt% at room temperature, low raw material costs, and good kinetic performance, show broad application prospects in fields such as on-board hydrogen storage and distributed energy storage.

[0004] However, titanium-manganese hydrogen storage alloys contain highly reactive metallic elements such as Ti and Mn, requiring processing into small-particle-size powders to ensure hydrogen absorption and desorption performance. During the crushing preparation process and repeated hydrogen absorption and desorption, the alloy continuously pulverizes, resulting in a dramatic increase in specific surface area, leading to an extremely high risk of spontaneous combustion of the powder. This risk persists throughout the entire lifecycle, from preparation and storage to transportation, use, and disposal. Existing technologies primarily rely on passive control through physical protection methods such as inert gas encapsulation and explosion-proof crushing devices. However, these measures significantly increase industrialization costs and fail to address the safety issue at the material level.

[0005] For titanium-manganese based hydrogen storage alloys, obtaining different properties through compositional control still relies heavily on experience and limited experimental trial and error, resulting in high R&D costs. While some data-driven machine learning methods have aided materials development, most focus on optimizing single hydrogen storage capacity or simply classifying and predicting known alloys, lacking a systematic and proactive design approach that explicitly optimizes "non-spontaneous combustion" in conjunction with hydrogen storage performance. More importantly, the development of titanium-manganese based hydrogen storage alloys faces unique challenges such as high-dimensional small sample sizes, complex and difficult-to-quantify spontaneous combustion mechanisms, and inherent conflicts between hydrogen storage capacity and safety, making it difficult for general machine learning methods to be directly applied. Summary of the Invention

[0006] This invention provides a machine learning-based design method and system for titanium-manganese hydrogen storage alloys. By using "non-spontaneous combustion resistance" and hydrogen storage performance as parallel optimization objectives, it efficiently screens alloy formulations with both high hydrogen storage capacity and intrinsic safety characteristics in the composition space, solving the problems of long cycle, high cost and lack of safety design in traditional trial-and-error methods.

[0007] The first aspect of this invention discloses a machine learning-based design method for titanium-manganese hydrogen storage alloys, comprising the following steps: Obtain multiple experimental data points for titanium-manganese hydrogen storage alloys. Each data point should include at least the chemical composition, hydrogen storage performance parameters, and auto-ignition characteristic label of the titanium-manganese hydrogen storage alloy. The auto-ignition characteristic label should be a binary label of "flammable" or "non-flammable". Based on the chemical composition in the experimental data, a feature pool containing key atomic ratio features was constructed, and a subset of key features was selected; among them, the key atomic ratio features include at least the A / B atomic ratio and the Cr / Mn atomic ratio; A multi-task learning model is constructed using the selected subset of key features as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets. A multi-objective optimization algorithm is used to optimize the multi-task learning model under set constraints. The optimization objectives are hydrogen storage performance and non-spontaneous combustion probability. The Pareto optimal alloy composition is output as a candidate formulation. Alloys were prepared based on candidate formulations, and their actual hydrogen storage performance and spontaneous combustion characteristics were tested. The experimental results were then updated to the experimental data for iterative optimization of the machine learning model.

[0008] Furthermore, the criteria for determining the spontaneous combustion characteristic label are as follows: a powder sample of titanium-manganese hydrogen storage alloy is dropped from a set height onto a non-flammable surface, and the presence of flames is observed during the drop process and within a set time period after landing; if flames appear, it is determined to be "flammable", otherwise it is determined to be "non-flammable".

[0009] Furthermore, key atomic ratio characteristics also include one or more of the following: (Cr+Fe) / Mn atomic ratio, the product of Fe atomic percentage and V atomic percentage, and valence electron concentration.

[0010] Furthermore, a subset of key features was selected. Specifically, correlation analysis and feature importance assessment methods were used, combined with prior knowledge of the physicochemical properties of titanium-manganese hydrogen storage alloys, to select features from the feature pool that are significantly correlated with hydrogen storage performance and spontaneous combustion characteristics and have interpretable physical meaning as the subset of key features.

[0011] Furthermore, the multi-task learning model is constructed using ensemble learning algorithms, including one or more of random forest, extreme gradient boosting, and stacking ensemble. By sharing the underlying feature layer, the multi-task learning model simultaneously handles the task of predicting continuous values ​​of hydrogen storage performance parameters and the task of classifying spontaneous combustion characteristic labels.

[0012] Furthermore, the constraints include one or more of the following: A / B atomic ratio constraint, Cr / Mn atomic ratio constraint, element content range constraint, and the sum of the atomic percentages of each element being 100%.

[0013] Furthermore, the multi-objective optimization algorithm is the non-dominated sorting genetic algorithm (NSGA-II); the output Pareto optimal alloy composition is used as a candidate formulation, specifically: several candidate compositions that achieve a balance between hydrogen storage performance and non-spontaneous combustion probability are selected from the Pareto front.

[0014] Furthermore, its actual spontaneous combustion characteristics were tested, specifically: room temperature air exposure test, powder drop impact test, and post-cycle spontaneous combustion verification test.

[0015] Furthermore, the experimental results are updated to the experimental data for iterative optimization of the machine learning model. Specifically, the actual hydrogen storage performance and spontaneous combustion characteristics measured in the experiment, along with the corresponding alloy composition, are added to the experimental dataset. When the amount of new data accumulated reaches a preset threshold, the feature selection and model training steps are re-executed to incrementally update or retrain the model.

[0016] A second aspect of the present invention discloses a machine learning-based titanium-manganese hydrogen storage alloy design system, comprising: The data acquisition and processing module is configured to: acquire multiple experimental data of titanium-manganese hydrogen storage alloys, each data point including at least the chemical composition, hydrogen storage performance parameters, and self-ignition characteristic label of the titanium-manganese hydrogen storage alloy; the self-ignition characteristic label is a binary label of "flammable" or "non-flammable"; The feature engineering and screening module is configured to: construct a feature pool containing key atomic ratio features based on the chemical composition in the experimental data, and screen out a subset of key features; wherein, the key atomic ratio features include at least the A / B atomic ratio and the Cr / Mn atomic ratio; The machine learning model building and performance prediction module is configured to: take the selected subset of key features as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets to build a multi-task learning model. The multi-objective optimization and alloy composition design module is configured to: use a multi-objective optimization algorithm to optimize the multi-task learning model under set constraints, with the optimization objectives being hydrogen storage performance and non-spontaneous combustion probability, and output the Pareto optimal alloy composition as a candidate formulation; The verification and feedback module is configured to: obtain the actual hydrogen storage performance and spontaneous combustion characteristics of alloys prepared based on candidate formulations, update the experimental results to the experimental data, and use them to iteratively optimize the machine learning model.

[0017] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects: By introducing "non-spontaneous combustion" as a quantifiable binary classification label into a machine learning model, a multi-task learning framework was constructed with key features such as the A / B atomic ratio and the Cr / Mn atomic ratio as core inputs. This enabled the simultaneous prediction of hydrogen storage capacity and spontaneous combustion risk of titanium-manganese hydrogen storage alloys. Through constraint-guided multi-objective optimization algorithms, the Pareto optimal alloy composition was searched within a physicochemically reasonable composition space, solving the dilemma in traditional methods where hydrogen storage performance and safety are difficult to balance, and optimization results easily deviate from physical realizability. Furthermore, experimental verification results were fed back to the dataset to form a closed-loop iteration, realizing a shift from "experience-based trial and error" to "data-driven" approaches. Compared to traditional trial-and-error methods that rely on numerous "preparation-testing-adjustment" cycles, this invention significantly reduces the candidate range for experimental verification through computer pre-screening, substantially lowering the time and economic costs of alloy development. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 A flowchart illustrating a machine learning-based titanium-manganese hydrogen storage alloy design method provided for one or more embodiments of the present invention; Figure 2 A schematic diagram of the correlation coefficient matrix provided for one or more embodiments of the present invention; Figure 3 This is a schematic diagram comparing the performance of different models provided in one or more embodiments of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] As described in the background section, traditional alloy development relies on a repetitive cycle of "preparation-testing-adjustment," with each round of trial and error requiring a complete process including smelting, crushing, and performance testing, resulting in extremely high time and economic costs. More importantly, the assessment of spontaneous combustion risk must be conducted in a real powder state, further increasing the complexity and danger of the experiment. This solution, however, constructs a machine learning model to solidify the mapping relationship between alloy composition and target performance within the model. This allows researchers to quickly evaluate the hydrogen storage capacity and spontaneous combustion tendency of various virtual formulations on a computer, only using a few "highly probable" candidate formulations output by the model for experimental verification. This "computer pre-screening + experimental confirmation" model shifts the trial and error process from the "laboratory" to the "computer," fundamentally compressing the trial-and-error space and significantly reducing the time and economic costs of development.

[0023] Specifically, this solution provides a machine learning-based design method and system for titanium-manganese hydrogen storage alloys. It encodes materials science mechanisms into computable features and constructs a multi-task learning model with Cr / Mn atomic ratio, A / B atomic ratio, etc. as core inputs. It simultaneously predicts hydrogen storage capacity and spontaneous combustion risk. With phase stability, key atomic ratio, etc. as hard constraints, it guides a multi-objective optimization algorithm to search for Pareto optimal composition within a reasonable physicochemical space, achieving synergistic optimization of safety and performance.

[0024] Example 1: A machine learning-based design method for titanium-manganese hydrogen storage alloys includes the following steps: S1. Data Acquisition and Processing: Acquire multiple experimental data of titanium-manganese hydrogen storage alloys. Each data point includes the alloy's chemical composition, hydrogen storage performance parameters, and auto-ignition characteristic label. The auto-ignition characteristic label is a binary label of "flammable" or "non-flammable". S2. Feature Engineering and Screening: Based on the chemical composition of the alloy, a feature pool containing key atomic ratio features is constructed, which includes at least the A / B atomic ratio and the Cr / Mn atomic ratio; through correlation analysis and feature importance assessment, a subset of key features is screened from the feature pool. S3. Machine Learning Model Construction and Performance Prediction: Using the selected subset of key features as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets, a multi-task learning model is constructed; cross-validation is used to evaluate the model accuracy, and the optimal model is selected as the performance predictor. S4. Multi-objective optimization and alloy composition design: The optimal performance predictor is encapsulated as a fitness evaluation function. Composition constraints, including A / B atomic ratio ≤ 0.5, are set. A multi-objective optimization algorithm is used to search within the composition space that meets the constraints. The optimization objective is to simultaneously maximize the predicted hydrogen storage performance and the probability of non-spontaneous combustion, and output the Pareto optimal alloy composition. S5. Experimental Verification and Feedback: Alloys were prepared based on the Pareto optimal alloy composition, and their actual hydrogen storage performance and spontaneous combustion characteristics were tested. The experimental results were fed back to the dataset in S1 for iterative optimization of the machine learning model.

[0025] The following is combined Figure 1 The specific process will be explained in detail.

[0026] S1. Data Acquisition and Processing.

[0027] By constructing a high-quality, standardized training dataset, a data foundation is provided for the model to learn the mapping relationship between "composition-performance-safety". Specifically, multiple experimental data of titanium-manganese hydrogen storage alloys are obtained. Each data includes the chemical composition of the alloy, hydrogen storage performance parameters, and self-ignition characteristic labels. The self-ignition characteristic labels are binary labels of "flammable" or "non-flammable".

[0028] The detailed process is as follows: (1) Data Collection: Collect experimental data on multi-element titanium-manganese hydrogen storage alloys containing elements such as Ti, Zr, Mn, Cr, V, and Fe from published academic papers or historical experimental records of the research team. Each data point should include at least: Chemical composition: atomic percentage or mass percentage of each element; Hydrogen storage performance parameters: at least one of the following: maximum hydrogen storage capacity, effective hydrogen storage capacity, plateau pressure, PCT curve slope, and hysteresis; Spontaneous combustion characteristic label: a binary label of "flammable" or "non-flammable" based on experimental or theoretical criteria.

[0029] (2) Standardization of Spontaneous Combustion Characteristic Labeling: To ensure label consistency and repeatability, the determination of spontaneous combustion characteristics adopts existing standards (such as GB / T21850). The standard defines spontaneous combustion as "the generation of flame without triggering under normal conditions." The specific operation includes: dropping 1 ml or 2 ml of powder sample from a height of 1 m onto a non-combustible surface, observing the drop process and whether a flame appears within 5 minutes after landing, repeating 6 times; if a flame appears, it is determined to be "flammable," otherwise it is "non-flammable." Samples in historical data that are not labeled according to this standard need to be re-evaluated and labeled based on experimental records.

[0030] (3) Data preprocessing: The collected data is formatted, outliers are removed, missing values ​​are processed, and the data is divided into training set and validation set.

[0031] S2. Feature Engineering and Screening.

[0032] This step transforms the microscopic mechanisms of materials science into numerical features that can be processed by machine learning. By reducing dimensionality through feature selection, overfitting is avoided, ensuring that the laws learned by the model have physical meaning and interpretability. Specifically, based on the chemical composition of the alloy, a feature pool containing key atomic ratio features is calculated and constructed. The key atomic ratio features include at least the A / B atomic ratio and the Cr / Mn atomic ratio. Through correlation analysis and feature importance assessment, a subset of key features is selected from the feature pool.

[0033] The detailed process is as follows: (1) Based on the chemical composition of the alloy and known physicochemical principles, an initial feature pool is constructed. Feature types include basic composition features, key atomic ratio features, interaction term features, and electronic structure features.

[0034] The basic composition is characterized by the atomic percentage of each element.

[0035] Key atomic ratio characteristics are features based on materials science mechanisms and include: A / B atomic ratio: The atomic ratio of A-site elements (Ti, Zr) to B-site elements (Mn, Cr, V, Fe), used to characterize the stability of the AB2 type Laves phase. Limiting A / B ≤ 0.5 is a key structural parameter for forming a stable main phase. In this scheme, the A / B atomic ratio = the atomic ratio of the A-side (Ti+Zr) to the B-side (V+Cr+Mn+Fe+Al+Ni). Cr / Mn atomic ratio: The atomic ratio of the protective element Cr to the risky element Mn, used to quantify the relative advantages of Cr and Mn in competitive diffusion to form a surface oxide layer during the initial stage of oxidation; (Cr+Fe) / Mn atomic ratio: the ratio of protective elements to risky elements.

[0036] Regarding the "Cr / Mn atomic ratio" and "(Cr+Fe) / Mn atomic ratio", when a formula contains only Cr and no Fe, the "Cr / Mn atomic ratio" can be used. If Fe is present, since Fe has a significant impact on material properties, the "(Cr+Fe) / Mn atomic ratio" needs to be considered.

[0037] Interaction term features are used to capture nonlinear synergistic effects between elements, such as Fe atomic percentage × V atomic percentage.

[0038] Electronic structure characteristics are defined by valence electron concentration (e / a), which is calculated by multiplying the number of valence electrons of each element by its atomic percentage and summing the results. This is used to correlate the phase stability of alloys and the thermodynamic behavior of hydrides.

[0039] (2) Using correlation analysis and feature importance assessment methods (such as random forest feature importance and recursive feature elimination), combined with prior knowledge of materials science, a subset of key features that have the greatest impact on the performance of the predicted target and have clear physical meaning is selected from the initial feature pool.

[0040] The screening criteria include that the features are significantly correlated with the target variable (e.g., Figure 2 (As shown), low collinearity among features, and features with interpretable physicochemical significance.

[0041] S3. Machine Learning Model Building and Performance Prediction.

[0042] This step establishes a mapping model of "composition characteristics → hydrogen storage performance + spontaneous combustion risk" as a fitness evaluation function for subsequent optimization search, enabling rapid prediction of the performance of new formulations. Specifically, a multi-task learning model is constructed using a selected subset of key features as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets. Cross-validation is used to evaluate the model accuracy, and the optimal model is selected as the performance predictor.

[0043] The detailed process is as follows: (1) A multi-task learning model is constructed using the selected key feature subset as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets. This model solves the problem of large differences in data distribution between the two types of tasks by sharing the underlying feature layer and simultaneously handling continuous value prediction tasks (hydrogen storage capacity) and classification tasks (spontaneous combustion characteristics). It also improves the generalization ability under small sample conditions by utilizing the correlation between tasks.

[0044] (2) Employing ensemble learning algorithms, and considering the characteristics of small sample sizes and high interpretability requirements in materials research and development, multiple highly interpretable models are selected and integrated, such as: Logistic regression: used to verify the sign and significance of key feature coefficients, providing linear, quantifiable mechanistic verification; Decision trees: used to discover clear component threshold rules (such as Cr / Mn ratio > 1.8), and output decision rules that can directly guide engineering design; Random Forest: Provides robust feature importance ranking for cross-validation of key features; Extreme Gradient Boosting (XGBoost): Used to improve prediction accuracy.

[0045] (3) To address the class imbalance problem where there are far fewer “non-flammable” samples than “flammable” samples in the spontaneous combustion characteristic label, a weighted cross-entropy loss function is adopted to assign a higher weight to the “non-flammable” category, thereby improving the model’s ability to identify safe alloys.

[0046] (4) Model training and evaluation: k-fold cross-validation was used to evaluate the model's prediction accuracy. Evaluation metrics included mean squared error (MSE) and coefficient of determination (R²). 2 The model was used for hydrogen storage performance prediction, and accuracy, recall, and F1 score were used for spontaneous combustion characteristic classification. The model with the best overall performance was selected as the performance predictor. Figure 3 The performance comparison of different models is shown.

[0047] The amount of data for "spontaneous combustion" samples is very small, leading to class imbalance during training. To address this issue, a weighted cross-entropy loss function is used when training the spontaneous combustion characteristic classification model. The formula is weight = total number of samples / (number of classes × number of samples in that class). Higher penalty weights are assigned to the minority class ("non-flammable"), forcing the model to prioritize learning the features of safe alloys. Furthermore, during model evaluation, precision, recall, and F1 score (especially the F1 score for the "non-flammable" class) are emphasized over pure accuracy to ensure the model's ability to identify safe formulations. Simultaneously, the ensemble algorithms used, such as Random Forest and XGBoost, and their inherent bootstrapping mechanisms, naturally alleviate the problem of uneven data distribution to some extent.

[0048] (5) Introduce physical rule constraints in model training to ensure that the model output conforms to the laws of materials science. For example, when the A / B atomic ratio is ≤0.5, the model should tend to predict the formation of a stable Laves phase and avoid outputting prediction results that contradict the known phase diagram.

[0049] This embodiment employs a multi-task learning ensemble model to simultaneously handle two tasks: hydrogen storage performance (regression) and spontaneous combustion characteristics (classification) by sharing underlying features. Specifically, it utilizes ensemble algorithms such as Random Forest, XGBoost, and Stacking for construction.

[0050] Regarding the loss function: Regression task (hydrogen storage performance): using mean squared error (MSE).

[0051] Classification task (spontaneous combustion characteristics): Using the weighted cross-entropy loss function. To address the issue of insufficient "non-flammable" samples, a higher weight is assigned to the "non-flammable" category, based on the total number of samples / (number of categories × number of samples in that category), to improve the model's ability to identify safe alloys.

[0052] Weighting criteria: Model fusion weights: assigned based on the performance of each base learner in cross-validation (such as R² and F1 scores).

[0053] Physical constraint weights: A regularization term (such as forcing the A / B atomic ratio to be ≤0.5) is added to the loss function, and its intensity coefficient is determined through cross-validation to ensure that the prediction conforms to the laws of materials science.

[0054] S4. Multi-objective optimization and alloy composition design.

[0055] This step uses the prediction model as the optimization objective function, and performs a reverse search within a physicochemically reasonable composition space to automatically generate the optimal alloy formulation that balances hydrogen storage performance and safety. Specifically, the optimal performance predictor is encapsulated as a fitness evaluation function, and compositional constraints, including an A / B atomic ratio of ≤0.5, are set. A multi-objective optimization algorithm is used to search within the composition space that meets the constraints. The optimization objective is to simultaneously maximize the predicted hydrogen storage performance and the probability of non-spontaneous combustion, and output the Pareto optimal alloy composition.

[0056] The detailed process is as follows: (1) The optimal performance prediction model obtained in step S3 is encapsulated into a fitness evaluation function. This function receives the alloy composition input and outputs the predicted hydrogen storage capacity and the probability of non-self-ignition.

[0057] (2) To ensure that the optimization results are physically feasible, multi-level constraints are set, including component boundary constraints, structural stability constraints, safety threshold constraints and element type constraints.

[0058] Compositional boundary constraints: Based on process feasibility and phase diagram knowledge, the content ranges of each element are set (e.g., Ti: 20-35%, Mn: 25-33%, Cr: 20-32%, V: 8-10%, Fe: 0-6%); reflected in the characteristics, this is: the hydrogen storage alloy is Ti a Zr b Mn c Cr d V e Fe f M g , where 25≤a≤35, 2≤b≤6, 25≤c≤33, 20≤d≤32, 8≤e≤10, 0≤f≤6, 0≤g≤4; M is any one or a combination of several of Fe, Cu, Co, Mo, Al, La, Ni, Ce, and Nb; a+b+c+d+e+f+g=100.

[0059] Structural stability constraints are used to limit the A / B atomic ratio to ≤0.5, ensuring that the optimization results tend to form a stable AB2-type Laves phase main phase.

[0060] The safety threshold constraint is based on the pattern discovered by the model, which forces the Cr / Mn atomic ratio to be greater than 1.8 to ensure that the candidate components have high potential for resisting spontaneous combustion.

[0061] Element type constraints are used to limit the alloy to include Ti and Mn as main elements, and may include one or more of Zr, V, Cr, and Fe, controlling the complexity of the alloy system within an experimentally verifiable range.

[0062] (3) Multi-objective optimization search: The non-dominated sorting genetic algorithm (NSGA-II) is used to search within the component space that satisfies the constraints. The optimization objective is: Objective 1: Maximize the predicted hydrogen storage capacity (wt%). Objective 2: Maximize the predicted probability of non-spontaneous combustion (or equivalently minimize the spontaneous combustion risk score).

[0063] This scheme generates the next generation of population through selection, crossover, and mutation operations. After multiple generations of iteration, it outputs a Pareto optimal alloy composition that achieves the best balance between hydrogen storage performance and safety.

[0064] (4) Select 3-5 representative candidate components from the Pareto front, prioritizing those that perform well in terms of hydrogen storage capacity and low probability of spontaneous combustion and have moderate component complexity, as candidate formulations for subsequent experimental verification.

[0065] S5. Experimental verification and feedback.

[0066] This step verifies the actual performance of the computer-recommended formulation through experiments, feeding the experimental results back to the database to form a closed-loop iteration, continuously improving the model's predictive ability and design reliability. Specifically: alloys are prepared based on Pareto optimal alloy compositions, their actual hydrogen storage performance and auto-ignition characteristics are tested, and the experimental results are fed back to the S1 dataset for iterative optimization of the machine learning model.

[0067] The detailed process is as follows: (1) Alloy preparation: Based on the optimal composition obtained in step S4, alloy preparation is carried out by vacuum non-consumable arc melting or vacuum induction melting. The raw materials selected are industrial-grade sponge titanium, sponge zirconium, electrolytic manganese, metallic chromium, ferrovanadium alloy, etc. Manganese is prone to burn-off during the melting process, and a small amount needs to be added extra during weighing to compensate for the burn-off.

[0068] (2) Hydrogen storage performance test: After the oxide layer is removed by grinding the surface of the prepared hydrogen storage alloy ingot, it is mechanically crushed. 2-3g of powder particles are randomly weighed and placed into a high pressure gas adsorption analyzer (such as PCTPro). After vacuuming and calibrating the volume, about 3MPa of hydrogen with a purity of 99.99% is introduced at 25℃ for the first activation. Then, the hydrogen adsorption and desorption PCT curves of the alloy are tested at 25℃ and 40℃ respectively, and the hydrogen storage capacity, plateau pressure, hysteresis and other performance parameters are recorded.

[0069] (3) Spontaneous combustion characteristic test: The spontaneous combustion characteristic test shall be conducted in accordance with GB / T-21850 standard, specifically including: The room temperature air exposure test is as follows: the sample is placed in a vacuum environment to degas (60-80℃, 2-3h) to fully release the hydrogen adsorbed in the alloy. The powder sample is then placed on a non-combustible surface and left to stand for ≥72 hours. The results are observed for smoke, local reddening, combustion and abnormal temperature changes. The powder drop impact test is as follows: alloy powder is dropped freely from a height of 1m onto a non-combustible surface. The falling process and whether there is any luminescence, sparks, ignition or combustion within 30 minutes after landing are observed. The test is repeated at least 6 times. Alloy powder samples that have completed 100 hydrogen absorption and desorption cycles were selected, and the above two test experiments were repeated to verify the stability of the alloy's non-self-ignition properties after long-term cycling.

[0070] (4) Add the experimentally measured actual hydrogen storage performance and spontaneous combustion characteristics, along with the corresponding alloy composition, to the dataset of step S1. After accumulating a sufficient amount of new data, repeat steps S2 to S3 to incrementally train or retrain the model to achieve continuous optimization of the design closed loop.

[0071] This method utilizes atomic engineering to control the alloy composition and structure, constructing a multi-scale collaborative defense system to achieve a balance between hydrogen storage performance and non-spontaneous combustion characteristics. At the bulk level, with an A / B atomic ratio of ≤0.5 as the structural basis, the alloy forms a stable AB2-type Laves phase as the main phase, reducing the intrinsic chemical reactivity of the alloy. At the surface level, by increasing the content of Cr and Fe and their atomic ratio relative to Mn, Cr and Fe are guided to preferentially diffuse to the alloy surface, forming a dense and continuous protective layer of Cr2O3 and iron oxides. This effectively inhibits the outward diffusion of active Mn to form loose and harmful oxides, physically blocks oxygen from penetrating into the alloy interior, and significantly improves the alloy's combustion activation energy. At the thermodynamic level, by optimizing the concentration of valence electrons and the synergistic regulation of V element, the alloy can obtain a smooth PCT hydrogen absorption and desorption curve and low hysteresis characteristics, thereby reducing the risk of heat accumulation during hydrogen storage cycle.

[0072] Through the above-mentioned multi-scale synergistic effect, the spontaneous combustion chain reaction of hydrogen release-oxidation exothermic is blocked from the surface, bulk phase and thermodynamic levels, so as to achieve synergistic optimization of hydrogen storage performance and safety performance.

[0073] Example 2: A machine learning-based titanium-manganese hydrogen storage alloy design system, including; The data acquisition and processing module is configured to: acquire multiple experimental data of titanium-manganese hydrogen storage alloys, each data point including at least the chemical composition, hydrogen storage performance parameters, and self-ignition characteristic label of the titanium-manganese hydrogen storage alloy; the self-ignition characteristic label is a binary label of "flammable" or "non-flammable"; The feature engineering and screening module is configured to: construct a feature pool containing key atomic ratio features based on the chemical composition in the experimental data, and screen out a subset of key features; wherein, the key atomic ratio features include at least the A / B atomic ratio and the Cr / Mn atomic ratio; The machine learning model building and performance prediction module is configured to: take the selected subset of key features as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets to build a multi-task learning model. The multi-objective optimization and alloy composition design module is configured to: use a multi-objective optimization algorithm to optimize the multi-task learning model under set constraints, with the optimization objectives being hydrogen storage performance and non-spontaneous combustion probability, and output the Pareto optimal alloy composition as a candidate formulation; The verification and feedback module is configured to: prepare alloys according to candidate formulations, test their actual hydrogen storage performance and spontaneous combustion characteristics, update the experimental results to the experimental data, and use them to iteratively optimize the machine learning model.

[0074] By introducing "non-spontaneous combustion" as a quantifiable binary classification label into a machine learning model, a multi-task learning framework was constructed with key features such as the A / B atomic ratio and the Cr / Mn atomic ratio as core inputs. This enabled the simultaneous prediction of hydrogen storage capacity and spontaneous combustion risk of titanium-manganese hydrogen storage alloys. Through constraint-guided multi-objective optimization algorithms, the Pareto optimal alloy composition was searched within a physicochemically reasonable composition space, solving the dilemma in traditional methods where hydrogen storage performance and safety are difficult to balance, and optimization results easily deviate from physical realizability. Furthermore, experimental verification results were fed back to the dataset to form a closed-loop iteration, realizing a shift from "experience-based trial and error" to "data-driven" approaches. Compared to traditional trial-and-error methods that rely on numerous "preparation-testing-adjustment" cycles, this invention significantly reduces the candidate range for experimental verification through computer pre-screening, substantially lowering the time and economic costs of alloy development.

[0075] The above are merely preferred embodiments of this solution and are not intended to limit the solution. Various modifications and variations can be made to this solution by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this solution should be included within the scope of protection of this solution.

Claims

1. A machine learning-based design method for titanium-manganese hydrogen storage alloys, characterized in that, Includes the following steps: Obtain multiple experimental data points for titanium-manganese hydrogen storage alloys. Each data point should include at least the chemical composition, hydrogen storage performance parameters, and auto-ignition characteristic label of the titanium-manganese hydrogen storage alloy. The auto-ignition characteristic label should be a binary label of "flammable" or "non-flammable". Based on the chemical composition in the experimental data, a feature pool containing key atomic ratio features was constructed, and a subset of key features was selected; among them, the key atomic ratio features include at least the A / B atomic ratio and the Cr / Mn atomic ratio; A multi-task learning model is constructed using the selected subset of key features as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets. A multi-objective optimization algorithm is used to optimize the multi-task learning model under set constraints. The optimization objectives are hydrogen storage performance and non-spontaneous combustion probability. The Pareto optimal alloy composition is output as a candidate formulation. Alloys were prepared based on candidate formulations, and their actual hydrogen storage performance and spontaneous combustion characteristics were tested. The experimental results were then updated to the experimental data for iterative optimization of the machine learning model.

2. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, The criteria for determining the spontaneous combustion characteristic label are as follows: a powder sample of titanium-manganese hydrogen storage alloy is dropped from a set height onto a non-flammable surface, and the presence of flames is observed during the drop process and within a set time period after landing; if flames appear, it is determined to be "flammable", otherwise it is determined to be "non-flammable".

3. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, Key atomic ratio characteristics also include the (Cr+Fe) / Mn atomic ratio, the product of the Fe atomic percentage and the V atomic percentage, and one or more of the valence electron concentrations.

4. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, A subset of key features was selected by means of correlation analysis and feature importance assessment, combined with prior knowledge of the physicochemical properties of titanium-manganese hydrogen storage alloys, to select features from the feature pool that are significantly correlated with hydrogen storage performance and spontaneous combustion characteristics and have interpretable physical meaning as the subset of key features.

5. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, The multi-task learning model is constructed using ensemble learning algorithms, including one or more of random forest, extreme gradient boosting, and stacking ensemble. By sharing the underlying feature layer, the multi-task learning model can simultaneously handle the task of predicting continuous values ​​of hydrogen storage performance parameters and the task of classifying spontaneous combustion characteristic labels.

6. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, The constraints include one or more of the following: A / B atomic ratio constraint, Cr / Mn atomic ratio constraint, content range constraint for each element, and the sum of the atomic percentages of each element being 100%.

7. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, The multi-objective optimization algorithm is a non-dominated sorting genetic algorithm; the Pareto optimal alloy composition is output as a candidate formulation, specifically: several candidate compositions that achieve a balance between hydrogen storage performance and non-spontaneous combustion probability are selected from the Pareto front.

8. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, Its actual spontaneous combustion characteristics were tested, specifically: room temperature air exposure test, powder drop impact test, and post-cycle spontaneous combustion verification test.

9. The machine learning-based titanium-manganese hydrogen storage alloy design method as described in claim 1, characterized in that, The experimental results are updated to the experimental data for iterative optimization of the machine learning model. Specifically, the actual hydrogen storage performance and spontaneous combustion characteristics measured in the experiment, along with the corresponding alloy composition, are added to the experimental dataset. When the amount of new data accumulated reaches a preset threshold, the feature selection and model training steps are re-executed to incrementally update or retrain the model.

10. A machine learning-based titanium-manganese hydrogen storage alloy design system, characterized in that, include; The data acquisition and processing module is configured to: acquire multiple experimental data of titanium-manganese hydrogen storage alloys, each data point including at least the chemical composition, hydrogen storage performance parameters, and self-ignition characteristic label of the titanium-manganese hydrogen storage alloy; the self-ignition characteristic label is a binary label of "flammable" or "non-flammable"; The feature engineering and screening module is configured to: construct a feature pool containing key atomic ratio features based on the chemical composition in the experimental data, and screen out a subset of key features; wherein, the key atomic ratio features include at least the A / B atomic ratio and the Cr / Mn atomic ratio; The machine learning model building and performance prediction module is configured to: take the selected subset of key features as input and hydrogen storage performance parameters and spontaneous combustion characteristic labels as output targets to build a multi-task learning model. The multi-objective optimization and alloy composition design module is configured to: use a multi-objective optimization algorithm to optimize the multi-task learning model under set constraints, with the optimization objectives being hydrogen storage performance and non-spontaneous combustion probability, and output the Pareto optimal alloy composition as a candidate formulation; The verification and feedback module is configured to: obtain the actual hydrogen storage performance and spontaneous combustion characteristics of alloys prepared based on candidate formulations, update the experimental results to the experimental data, and use them to iteratively optimize the machine learning model.

Images