Alloy design method and system based on expert optimized trajectory learning framework, AlCoCrFeNi high-entropy alloy and preparation method thereof

Abstract

This invention discloses an alloy design method and system based on an expert-optimized trajectory learning framework, as well as an AlCoCrFeNi high-entropy alloy and its preparation method, belonging to the field of artificial intelligence-assisted material design. The method includes: S1 integrating experimental records of the same composition under different process conditions into ordered trajectories according to predefined performance indicators and performing state encoding; S2 constructing an Alloy Trajectory Optimization Network (ATON), capturing optimization history through a state-aware encoder, dynamically modulating compositional features based on process conditions and trajectory states using an adaptive gating mechanism, and independently predicting ultimate tensile strength and elongation using task-specific prediction heads; S3 quantifying prediction uncertainty using the Monte Carlo Dropout method, providing the prediction mean and confidence interval; and S4 performing multi-objective optimization recommendation based on the qNEHVI acquisition function. This invention can accurately predict alloy composition and process conditions, and achieve synergistic optimization of high strength and high plasticity, thereby accelerating the research and development process of advanced alloys.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted material design, specifically to an alloy design method and system based on an expert-optimized trajectory learning framework, which is particularly suitable for the rapid screening and recommendation of alloy composition and thermomechanical processing conditions for high strength and high plasticity synergistic optimization. Background Technology

[0002] With the rapid development of aerospace, automobile manufacturing, and other fields, the demand for high-performance alloy materials is becoming increasingly urgent. Traditional alloy design mainly relies on trial and error and the accumulated experience of researchers. This traditional R&D model based on physical metallurgy is not only time-consuming and costly, but also falls short when faced with the vast space of compositional design, severely restricting the development process of new high-performance alloys.

[0003] In recent years, machine learning technology, with its powerful high-dimensional data modeling capabilities, has been widely introduced into the field of materials informatics, providing a new data-driven paradigm for alloy design. By mining the implicit "composition-property" mapping relationship in existing experimental data, machine learning models can quickly predict the mechanical properties of unknown compositions, significantly narrowing the scope of experimental verification. For example, Chinese patent CN 115579091 A discloses a multi-property collaborative optimization method for high-entropy alloy composition design based on machine learning. After establishing a machine learning model, the Pareto front is determined by calculating the expected improvement value of each property in the composition space, and then the cluster centers are obtained through cluster analysis, thereby achieving efficient composition design for multi-property collaborative optimization of high-entropy alloys.

[0004] However, the final properties of a material depend not only on its chemical composition but also heavily on processing conditions. While existing machine learning techniques have made some progress in predicting single properties, current algorithms still have significant limitations, failing to simultaneously and accurately predict alloy composition and processing conditions. Specifically, the shortcomings of existing technologies are mainly reflected in the following aspects: First, existing machine learning models often only establish unidirectional positive predictions of "composition-property," and few algorithms can achieve multi-variable joint predictions including alloy processing conditions (such as rolling deformation, annealing temperature, and time). Second, the strength and plasticity of alloys typically exhibit an inverse relationship of mutual constraint in physical metallurgy. Current algorithms, when performing multi-objective optimization, often get trapped in local optima due to the lack of an effective fusion mechanism for the material's intrinsic physical mechanisms.

[0005] In summary, overcoming the shortcomings of existing machine learning algorithms and constructing a design method that can comprehensively consider and accurately predict alloy composition and process conditions, and achieve synergistic optimization of high strength and high plasticity, is a technical challenge that urgently needs to be solved in the field of alloy material research and development. Summary of the Invention

[0006] This invention provides an alloy design method and system based on an expert-optimized trajectory learning framework, aiming to construct a design method that can comprehensively consider and accurately predict alloy composition and process conditions, and achieve synergistic optimization of high strength and high plasticity.

[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0008] An alloy design method based on an expert-optimized trajectory learning framework includes the following steps:

[0009] S1: Trajectory-based data representation method: for a given alloy composition For each of the x constituent elements, collect all experimental records with the same composition and organize them into an experimental trajectory according to predefined performance indicators.

[0010] S2: Building the Alloy Trajectory Optimization Network (ATON): Building a trajectory-based tabular learning framework called the Alloy Trajectory Optimization Network (ATON).

[0011] S3: Bayesian uncertainty quantification using the Monte Carlo Dropout method: This method provides confidence intervals along with point predictions to recommend candidates.

[0012] Furthermore, in step S1, the trajectory-based data representation method includes the following steps:

[0013] S11: Trajectory Construction;

[0014] S12: Trajectory status coding.

[0015] According to a specific embodiment of the present invention, in step S11, the trajectory construction defines the comprehensive performance index as the product of ultimate tensile strength and elongation:

[0016]

[0017] in Indicates ultimate tensile strength. This represents the elongation rate. For a fixed component C, the corresponding experimental curves can be constructed as an ordered sequence.

[0018]

[0019] in Indicates the first Step processing parameters, This indicates the measured mechanical properties. This is a pseudo-time index. The trajectories are arranged in the following order:

[0020]

[0021] According to a specific embodiment of the present invention, in step S12, the trajectory state is encoded.

[0022] At each pseudo-time step t, the optimization history process can be summarized by a compact state vector:

[0023]

[0024] The four components represent: up to step The optimal performance achieved, the performance value of the previous step, the performance increment change, and the current position along the trajectory.

[0025] Furthermore, in step S2, the construction of the Alloy Trajectory Optimization Network (ATON) includes the following steps:

[0026] S21: State-aware trajectory coding method: It can capture optimization history through features such as best performance, recent improvement effects and trajectory position;

[0027] S22: Adaptive gating mechanism: Dynamically adjusts the combination features according to process conditions and trajectory status, reflecting the physical principle that "the same components will produce different properties under different thermomechanical treatments";

[0028] S23: Task prediction head for ultimate tensile strength and elongation: enables the model to learn independent feature representations corresponding to different performances in these mechanisms.

[0029] Furthermore, in step S21, the state-aware trajectory encoding method, the Alloy Trajectory Optimization Network (ATON), uses an encoder to process heterogeneous feature modes, thereby enabling effective representation learning at the levels of composition, process, testing, and trajectory state information.

[0030] Component encoder. Normalized component vectors. ( Mapped to a latent representation via a multilayer perceptron:

[0031]

[0032] Process condition encoder. Process conditions include categorical variables. (Preparation methods, processing techniques, and casting conditions) and continuous variables (For example, temperature, time, and deformation). These components are coded individually and then combined sequentially:

[0033]

[0034]

[0035] in, Indicates the relationship with the first Embedding layers related to categorical variables, This indicates a feature splicing operation.

[0036] Test conditions encoder. Mechanical performance test parameters. (Including strain rate, test conditions, and test temperature) are all coded into corresponding numerical values:

[0037]

[0038] State encoder. Trajectory state vector Embed in the following ways:

[0039]

[0040] According to a specific embodiment of the present invention, in step S22, the adaptive gating mechanism is used by the Alloy Trajectory Optimization Network (ATON) to capture the strong dependence of material properties on processing history. This mechanism can adjust the compositional characteristics of the material according to the process conditions and processing trajectory status.

[0041] Gated network. Based on the processing results and state representation, the element-wise gating vector is calculated:

[0042]

[0043] In the formula It is the sigmoid activation function.

[0044] Gated combinational properties. Combinatorial embedding is modulated via element-wise multiplication:

[0045]

[0046] in This represents the Hadamard product. This operation can selectively enhance or weaken component features depending on the processing context.

[0047] Feature fusion. Gated combined features are fused with the test condition embedding vector through a shared transformation:

[0048]

[0049] Furthermore, in step S23, the task-specific prediction head is used to model the ultimate tensile strength of the material respectively. ) and plasticity (δ):

[0050]

[0051]

[0052] The final model output is provided by Provided.

[0053] Furthermore, in step S3, the recommendation method uses the Monte Carlo Dropout method to estimate the posterior distribution of each candidate alloy and process condition, and uses the noisy expected hypervolume improvement (qNEHVI) acquisition function to balance the ultimate tensile strength and elongation for recommendation.

[0054] Furthermore, in step S3, the prediction uncertainty is quantified using the Monte Carlo Dropout method:

[0055] During inference, the Dropout layer remains active and performs N random forward propagations to generate a set of predictions:

[0056]

[0057] The predicted average and uncertainty estimates are as follows:

[0058]

[0059]

[0060] Assuming approximate normality, the 95% confidence interval is estimated as follows: .

[0061] Furthermore, the alloy design method based on the expert-optimized trajectory learning framework also includes: experimentally verifying the recommended composition and thermomechanical processing parameters;

[0062] This invention also provides an alloy design system based on an expert-optimized trajectory learning framework. The system employs the method described above to achieve joint optimization design of alloy composition and thermomechanical processing parameters. The system includes:

[0063] The acquisition module is used to acquire training data, which includes the alloy composition, thermomechanical processing parameters, test parameters, and the corresponding measured values ​​of ultimate tensile strength and elongation.

[0064] The trajectory construction module is used to sort experimental records of the same composition under different process conditions from low to high according to comprehensive performance index to construct an ordered trajectory, and calculate the state vector of each trajectory point. The trajectory data containing alloy composition, process parameters, test parameters, state vector and corresponding measured mechanical properties are output to the training module.

[0065] The training module is used to input the trajectory data into the Alloy Trajectory Optimization Network (ATON) for training to obtain a trained machine learning model; and output the ATON model parameters after training to the evaluation module and the prediction module simultaneously.

[0066] The evaluation module is used to evaluate the robustness of the established machine learning model;

[0067] The prediction module receives ATON model parameters from the training module and virtual composition space data of the alloy to be designed from external input. It performs multiple random forward propagations on each candidate sample in the virtual composition space and outputs the predicted mean and 95% confidence interval of the ultimate tensile strength and elongation of each candidate scheme to the recommendation module.

[0068] The recommendation module receives the predicted mean and 95% confidence interval of each candidate scheme, sorts the candidate conditions, and recommends the combination of alloy composition and thermomechanical processing parameters that have both high ultimate tensile strength and high elongation.

[0069] Furthermore, the system also includes:

[0070] The verification module performs experimental preparation and mechanical property testing according to the alloy composition and thermomechanical processing parameters output by the recommendation module. It compares the measured values ​​with the predicted confidence intervals and feeds the verification results back to the training module to trigger incremental learning and iterative optimization of the model.

[0071] Third, this invention also provides a method for preparing an AlCoCrFeNi high-entropy alloy, characterized by comprising the following steps:

[0072] Using the alloy design method based on the expert-optimized trajectory learning framework, recommended alloy compositions and corresponding thermomechanical processing parameters are selected from the composition space of Al-Co-Cr-Fe-Ni high-entropy alloys.

[0073] The alloy is prepared according to the recommended alloy composition, melted and cast in a protective atmosphere using a vacuum arc melting method to obtain a cast alloy.

[0074] The as-cast alloy is rolled, deformed, and annealed according to the recommended thermomechanical processing parameters to obtain the finished alloy.

[0075] Furthermore, the selected recommended alloy composition, by mass percentage, is Al: 5~8 wt.%, Co: 15~18 wt.%, Cr: 15~18 wt.%, Fe: 15~18 wt.%, Ni: balance; the recommended thermomechanical processing parameters are: rolling deformation: 65~85%, annealing temperature: 600~800 ℃, annealing time: 0.5-4 h, strain rate: 1×10⁻⁶ -3 s -1 .

[0076] Fourth, the present invention also provides a high-entropy alloy, characterized in that it is prepared by the above-mentioned high-entropy alloy preparation method.

[0077] Compared with the prior art, the present invention has the following beneficial technical effects:

[0078] 1) By using a multi-encoder architecture to take alloy composition, thermomechanical processing parameters (including categorical and continuous variables) and test conditions as joint inputs, synchronous and accurate prediction of alloy composition and process conditions is achieved in the same framework, enabling rapid and accurate prediction of the composition and process conditions of high-strength-ductile alloys.

[0079] 2) By representing the performance evolution of the same component under different process conditions through trajectory data, the model organizes it into an ordered trajectory, enabling it to learn the physical metallurgical laws inherent in the optimization path "from low performance to high performance." Simultaneously, the task prediction head allows ultimate tensile strength and elongation to learn independent feature representations, and the adaptive gating mechanism dynamically adjusts the contribution of component features according to the process state. The synergistic effect of these three elements allows the model to effectively balance strength and plasticity and perform collaborative optimization, avoiding getting trapped in local optima.

[0080] 3) The Monte Carlo Dropout method is used to provide a complete posterior distribution estimate and a 95% confidence interval during the prediction phase, enabling researchers to quantitatively assess the predictive risk of each candidate solution. Combined with the qNEHVI acquisition function's multi-objective optimization strategy under uncertainty, the recommended candidate solutions take into account both the prediction mean and prediction variance, significantly improving the relevance and success rate of experimental validation. Attached Figure Description

[0081] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings are provided to further explain the invention and constitute a part 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. In the drawings:

[0082] Figure 1This is a flowchart of the alloy design method based on the expert-optimized trajectory learning framework according to an embodiment of the present invention;

[0083] Figure 2 This is an engineering stress-strain curve of a high-entropy alloy according to an embodiment of the present invention. Detailed Implementation

[0084] To make the technical solutions and effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments; based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0085] This invention provides a trajectory-based table learning framework alloy design method, such as... Figure 1 As shown, the method includes the following steps:

[0086] S1: Establish a dataset of mechanical properties for thermomechanical processing of alloys.

[0087] A literature review was conducted on target alloy systems (including but not limited to high-entropy alloys, titanium alloys, aluminum alloys, magnesium alloys, and nickel-based superalloys), collecting composition, thermomechanical processing parameters, test parameters, and mechanical property data for different alloy systems to establish an alloy thermomechanical processing mechanical property dataset. The composition includes the types and amounts of elements constituting the alloy, with a total of 22 element types. The thermomechanical processing parameters include rolling process parameters, heat treatment process parameters, and other thermomechanical processing conditions. The test parameter is the strain rate. The mechanical properties include ultimate tensile strength and elongation. That is, for a given alloy composition... This corresponds to the atomic fraction of the 22 constituent elements.

[0088] S2: Training and evaluating the performance of a machine learning model based on the Alloy Trajectory Optimization Network (ATON) algorithm: Using the aforementioned alloy thermomechanical properties dataset, a machine learning model was established and trained using the Alloy Trajectory Optimization Network (ATON) algorithm, with alloy composition, thermomechanical processing parameters, and test parameters as input data. Root mean square error (RMSE) and coefficient of determination (R²) were used. 2 We will use this to evaluate the robustness of the model before proceeding to the next step.

[0089] S3: Design a virtual composition space, predict its composition using a trained machine learning model, and rank candidate conditions based on the posterior predictions generated by the Alloy Trajectory Optimization Network (ATON) using a noisy expected hypervolume improvement (qNEHVI) acquisition function. Perform qNEHVI optimization via the Bgolearn API to achieve efficient multi-objective optimization under prediction uncertainty, recommending alloys with both high ultimate tensile strength and high elongation; the virtual composition space is a sample group consisting of alloys to be designed whose compositions meet the predetermined range.

[0090] S4: Experimentally verify the recommended alloy composition and thermomechanical processing parameters.

[0091] Example 1

[0092] This invention provides an alloy design method based on an expert-optimized trajectory learning framework, the method comprising the following steps:

[0093] S1: A literature review was conducted on high-entropy alloys, collecting data on the composition, thermomechanical processing parameters, test parameters, and mechanical properties of different high-entropy alloy systems. A total of 1000 samples were collected to establish a dataset of high-entropy alloy thermomechanical processing mechanical properties. The composition includes the types and amounts of elements that make up the alloy; the thermomechanical processing parameters include rolling process parameters, heat treatment process parameters, and other thermomechanical processing conditions; the test parameter is the strain rate; and the mechanical properties include ultimate tensile strength and elongation.

[0094] S2: The dataset is divided into an 80% training set and a 20% test set. The composition of all elements in the dataset, thermomechanical processing parameters, and strain rates are used as input data. A machine learning model is trained using the Alloy Trajectory Optimization Network (ATON) algorithm. Root mean square error (RMSE) and coefficient of determination (R²) are used. 2 We will then evaluate the model's performance. Next steps will follow.

[0095] S3: A virtual composition space was constructed to explore potential novel alloys. This virtual composition space consisted of compositions observed in the training dataset; process parameters were systematically varied within experimentally feasible limits. Specifically, rolling deformation varied from its minimum to its maximum observed value in 1% increments. Annealing temperature was scanned from its minimum to its maximum value in 5°C increments, and annealing time varied in 0.5-h increments. The strain rate was fixed at 1 × 10⁻⁶. -3 s -1This method generates a virtual composition space containing 20,000 potential high-entropy alloys and uses the established Alloy Trajectory Optimization Network (ATON) machine learning model to predict ultimate tensile strength and elongation. Then, based on the posterior predictions generated by ATON, a noisy expected hypervolume improvement (qNEHVI) acquisition function is used to rank the candidate conditions. qNEHVI optimization is performed via the Bgolearn API, thereby achieving efficient multi-objective optimization under prediction uncertainty and recommending high-entropy alloys with both high ultimate tensile strength and high elongation.

[0096] Based on the recommendations, the preferred alloy composition and thermomechanical processing parameters are as follows:

[0097] Al: 5~8 wt. %, Co: 15~18 wt. %, Cr: 15~18 wt. %, Fe: 15~18 wt. %, Ni: balance.

[0098] Rolling deformation: 65~85%, annealing temperature: 600~800 ℃, annealing time: 0.5 - 4 h.

[0099] Example:

[0100] Based on the recommended results of steps S1-S3 above, this embodiment selects a high-entropy alloy with the following composition (by mass percentage): 8% Al, 17% Co, 15% Cr, 16% Fe, with the balance being Ni (i.e., 8Al-17Co-15Cr-16Fe-Ni) for verification. The recommended thermomechanical processing technology is: rolling deformation of 65%, annealing temperature of 800℃, and annealing time of 1 hour.

[0101] High-purity (≥99.9 wt%) Al, Co, Cr, Fe, and Ni metal particles were used as raw materials and accurately weighed according to the above-mentioned component ratios. Melting was carried out in a vacuum arc melting furnace under a high-purity argon protective atmosphere. To ensure the uniformity of the alloy composition, the resulting alloy ingot was flipped and remelted, a process repeated eight times. Subsequently, the uniform alloy melt was drawn into a water-cooled copper mold using a suction casting method to obtain a cast alloy ingot.

[0102] The cast alloy ingots were cut into plates of suitable size and cold rolled at room temperature, with the total deformation precisely controlled at 65%. The rolled plates were then placed in a box-type resistance furnace preheated to 800°C for annealing. After holding at that temperature for 1 hour, they were removed and quickly water-quenched to obtain a recrystallized structure and a good balance of strength and plasticity.

[0103] Standard "dog bone" (plate-shaped) tensile specimens are machined from rolled and annealed sheet metal using wire cutting. The surface of the specimens is then polished sequentially with silicon carbide sandpaper, from coarse to fine (up to 2000 grit), to remove machining marks.

[0104] Uniaxial tensile tests were performed on tensile specimens at room temperature using a universal testing machine, with the initial strain rate set to 1 × 10⁻³ s⁻¹ to 1 × 10⁻³ s⁻¹. To ensure data reliability, three parallel specimens of the same alloy condition were tested, and their average values ​​were calculated.

[0105] Figure 2 The engineering stress-strain curves of the alloy in this embodiment are shown. Test results demonstrate that the alloy possesses excellent mechanical properties: a yield strength of 1.3 GPa, an ultimate tensile strength of 1.6 GPa, and an elongation after fracture of 17.6%. This result perfectly achieves the synergistic optimization of high strength and high plasticity.

[0106] The measured performance data was compared with the predicted mean and 95% confidence interval given in step S33. The measured values ​​all fell well within the predicted interval, which fully verifies the accuracy and reliability of the alloy design method based on the expert optimization trajectory learning framework proposed in this invention. This method effectively shortens the cycle from composition design to experimental verification, providing a powerful data-driven tool for the rapid development of advanced high-performance alloys.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An alloy design method based on an expert-optimized trajectory learning framework, characterized in that, Includes the following steps: S1: Constructing a trajectory-based alloy process optimization dataset: For a given alloy composition vector Where x is the number of constituent elements, all experimental records with the same composition and corresponding to different thermomechanical processing conditions are collected, organized into ordered experimental trajectories according to predefined performance indicators, and state encoding is performed on each trajectory point to obtain a trajectory state vector containing optimization history information. S2: Construct and train the alloy trajectory optimization network ATON: The ATON network is trained under supervision by taking the alloy composition, thermomechanical processing parameters, test parameters, and trajectory state vector of the trajectory described in step S1 as inputs, and the ultimate tensile strength and elongation as output targets. The ATON network includes a state-aware trajectory encoder, an adaptive gating mechanism, and a task prediction head for ultimate tensile strength and elongation. S3: Uncertainty quantification based on Monte Carlo Dropout method: During the inference phase, the Dropout layer of the ATON network described in step S2 is kept active. Multiple random forward propagations are performed on each input sample of the candidate alloy to obtain the predicted mean and standard deviation of the ultimate tensile strength and elongation, and the prediction confidence interval is given accordingly. S4: Multi-objective optimization recommendation: Using the predicted mean and standard deviation obtained in step S3 as input, the noisy expected hypervolume improved acquisition function is used to perform multi-objective optimization in the joint search space of alloy composition and thermomechanical processing parameters, and recommend candidate alloy compositions with both high strength and high plasticity and corresponding thermomechanical processing parameter combinations.

2. The alloy design method based on an expert-optimized trajectory learning framework according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11: Trajectory Construction: The product of ultimate tensile strength and elongation is defined as a comprehensive performance index. ,in, Indicates the first The ultimate tensile strength of this experiment Represents elongation; For a fixed component C, all experimental records under different thermomechanical processing conditions were constructed into an ordered sequence: ,in, Indicates the first Step processing parameters, This indicates the measured mechanical properties. It is a pseudo-time index; the trajectory is based on Values ​​sorted in ascending order: This causes the trajectory to evolve from low performance to high performance; S12: Trajectory State Encoding: At each pseudo-time step t, the optimization history is encoded into a four-dimensional state vector: The four components represent: up to step The optimal performance achieved, the performance value of the previous step, the performance increment change, and the current position along the trajectory.

3. The alloy design method based on an expert-optimized trajectory learning framework according to claim 1, characterized in that, In step S2, the ATON network includes four encoders: a component encoder, a process condition encoder, a test condition encoder, and a status encoder. The component encoder, after normalizing the component vector Mapped to a latent representation via a multilayer perceptron: The process condition encoder includes categorical variables for the process conditions. and continuous variables These components are encoded individually and then combined sequentially: in, Indicates the relationship with the first Embedding layers related to categorical variables, Indicates feature concatenation operation The test conditions encoder, mechanical performance test parameters The strain rate, test conditions, and test temperature are all encoded as corresponding values: The state encoder, trajectory state vector Embed in the following ways: 。 4. The alloy design method based on an expert-optimized trajectory learning framework according to claim 3, characterized in that, In step S2, the adaptive gating mechanism takes the trajectory state vector constructed in step S1 and the process conditions in step S3 as inputs, and dynamically modulates the response of the component features to different process conditions, specifically including: S21. Gated Network: Based on the processing results and state representation, calculate the element-wise gating vector: In the formula It is the sigmoid activation function; S22. Gated Combination: Combinatorial embedding is modulated by element-wise multiplication: in Represents the Hadamard product; S23. Feature Fusion: The modulated component features are fused with the test condition embedding vector through a shared transformation. 。 5. The alloy design method based on an expert-optimized trajectory learning framework according to claim 1, characterized in that, In step S2, the task prediction head measures the ultimate tensile strength of the modeling material. With plasticity Make a prediction: The final model output is provided by Provided.

6. The alloy design method based on an expert-optimized trajectory learning framework according to claim 1, characterized in that, In step S3, quantization is performed using the Monte Carlo Dropout method: During inference, the Dropout layer remains active and performs N random forward propagations to generate a set of predictions: The predicted average and uncertainty estimates are as follows: Normality, 95% confidence interval estimate is .

7. The alloy design method based on an expert-optimized trajectory learning framework according to claim 1, characterized in that, The noisy expected supervolume improvement acquisition function uses ultimate tensile strength and elongation as two optimization objectives, and takes the predicted mean and predicted standard deviation obtained in step S3 as inputs. It calculates the expected supervolume improvement value of each candidate alloy composition and thermomechanical processing parameter combination, and selects the top M candidate schemes as recommended outputs after sorting them in descending order of the improvement value, so as to balance the utilization of prediction performance and the exploration of high uncertainty regions.

8. The alloy design method based on an expert-optimized trajectory learning framework according to any one of claims 1-7, characterized in that, It also includes step S5: according to the alloy composition and thermomechanical processing parameters recommended in step S4, perform alloy melting, thermomechanical processing and mechanical property testing, and compare the measured mechanical properties with the confidence interval predicted in step S3 for verification; if the measured value falls within the predicted confidence interval, the verification is successful; if the measured value deviates from the predicted confidence interval, the measured data is fed back to the dataset for model iteration and update.

9. An alloy design system based on an expert-optimized trajectory learning framework, characterized in that, The system employs the method described in any one of claims 1 to 8 to achieve joint optimization design of alloy composition and thermomechanical processing parameters, and the system comprises: The acquisition module is used to acquire training data, which includes the alloy composition, thermomechanical processing parameters, test parameters, and the corresponding measured values ​​of ultimate tensile strength and elongation. The trajectory construction module is used to sort experimental records of the same composition under different process conditions from low to high according to the comprehensive performance index to construct an ordered trajectory, calculate the state vector of each trajectory point, and output the trajectory data containing alloy composition, process parameters, test parameters, state vector and corresponding measured mechanical properties to the training module. The training module is used to input the trajectory data into the Alloy Trajectory Optimization Network (ATON) for training to obtain a trained machine learning model; and simultaneously output the ATON model parameters after training to the evaluation module and the prediction module. The evaluation module is used to evaluate the robustness of the established machine learning model; The prediction module receives ATON model parameters from the training module and virtual composition space data of the alloy to be designed from external input. It performs multiple random forward propagations on each candidate sample in the virtual composition space and outputs the predicted mean and 95% confidence interval of the ultimate tensile strength and elongation of each candidate scheme to the recommendation module. The recommendation module receives the predicted mean and 95% confidence interval of each candidate scheme, sorts the candidate conditions, and recommends the combination of alloy composition and thermomechanical processing parameters that have both high ultimate tensile strength and high elongation.

10. The alloy design system based on an expert-optimized trajectory learning framework according to claim 9, characterized in that, Also includes: The verification module performs experimental preparation and mechanical property testing according to the alloy composition and thermomechanical processing parameters output by the recommendation module. It compares the measured values ​​with the predicted confidence intervals and feeds the verification results back to the training module to trigger incremental learning and iterative optimization of the model.

11. A method for preparing an AlCoCrFeNi high-entropy alloy, characterized in that, Includes the following steps: Using the alloy design method based on the expert-optimized trajectory learning framework as described in any one of claims 1 to 8, recommended alloy compositions and corresponding thermomechanical processing parameters are screened from the composition space of Al-Co-Cr-Fe-Ni system high-entropy alloys. The alloy is prepared according to the recommended alloy composition, melted and cast in a protective atmosphere using a vacuum arc melting method to obtain a cast alloy. The as-cast alloy is rolled, deformed, and annealed according to the recommended thermomechanical processing parameters to obtain the finished alloy.

12. The method for preparing a high-entropy alloy according to claim 11, characterized in that, The selected recommended alloy composition, by mass percentage, is Al: 5~8 wt.%, Co: 15~18 wt.%, Cr: 15~18 wt.%, Fe: 15~18 wt.%, Ni: balance; the recommended thermomechanical processing parameters are: rolling deformation: 65~85%, annealing temperature: 600~800℃, annealing time: 0.5-4h, strain rate: 1×10⁻⁶. -3 s -1 .

13. A high-entropy AlCoCrFeNi alloy, characterized in that, It is prepared by the method for preparing high-entropy alloy according to any one of claims 11 or 12.

Images