Method for machine learning aided design of structural alloy for fourth-generation thorium-based molten salt reactor

Abstract

The invention relates to a method for machine learning aided design of a structural alloy for a fourth-generation thorium-based molten salt reactor, and belongs to the technical field of artificial intelligence auxiliary material design. The method comprises the steps that S1, an alloy component machine learning model for phase composition prediction is established, specifically, based on known alloy sample data, the sample data comprise alloy element content and corresponding phase composition information, and the machine learning model used for predicting alloy phase composition is established; s2, screening an initial alloy component space through phase composition constraint conditions; s3, secondary screening of alloy components based on thermodynamic simulation; s4, alloy sample preparation and solid solution homogenization treatment; s5, carrying out a mechanical property test; and S6, experimental result feedback and machine learning model closed-loop updating are carried out. According to the method, the artificial intelligence technology and the traditional alloy design theory are organically fused, an efficient and reliable new path is provided for research and development of key structural materials for the fourth-generation thorium-based molten salt reactor, and the important scientific research value and the engineering application prospect are achieved.

Description

Machine learning-aided design method for structural alloys used in fourth-generation thorium-based molten salt reactors Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted material design technology, and relates to a method for machine learning-assisted design of structural alloys for fourth-generation thorium-based molten salt reactors. Background Technology

[0002] Thorium-based molten salt reactors (TMSRs) are attracting widespread attention due to their inherent safety, atmospheric pressure operation, high-temperature output characteristics, and waterless cooling. However, the commercial application of molten salt reactors still faces a series of stringent technical challenges, among which the performance of structural materials is one of the core bottlenecks determining reactor safety and economic lifespan. Unlike traditional pressurized water reactors, the structural components of thorium-based molten salt reactors must operate for extended periods in complex environments with the combined effects of multiple extreme factors. This triple extreme condition of high temperature, strong corrosion, and high neutron radiation places almost demanding performance requirements on structural alloys. Since increasing the reactor outlet temperature is the most direct and effective way to improve the thermoelectric conversion efficiency of molten salt reactors, developing novel structural alloys that can withstand long-term stable operation at 800 °C, strong neutron radiation, and molten salt corrosion has become a core material bottleneck driving thorium-based molten salt reactors towards commercial application.

[0003] In general, while existing nickel-based alloy systems each have their advantages, they still cannot directly meet the comprehensive requirements of fourth-generation thorium-based molten salt reactors at 800 °C for structural alloys in terms of high-temperature strength, mid-temperature plasticity, resistance to molten salt corrosion, and resistance to neutron irradiation. Therefore, developing a structural alloy for MSR that truly meets the 800 °C service temperature requirement and possesses excellent comprehensive performance based on novel alloy design concepts has significant scientific and engineering application value.

[0004] Traditional alloy design methods rely primarily on trial and error and experience-driven composition adjustments, often resulting in development cycles that can last for decades. This inefficient development model is ill-suited to the urgent needs of today's rapidly evolving advanced nuclear energy systems, necessitating the introduction of new design paradigms to accelerate the development of alloy materials.

[0005] In recent years, with the rapid development of artificial intelligence technology, the use of machine learning to assist in materials design has gradually attracted widespread attention from academia and industry. Leveraging the powerful data processing and pattern recognition capabilities of machine learning algorithms, researchers can deeply explore and quantitatively characterize the complex elemental interactions, multi-scale microstructure evolution, and structure-property relationships in alloy systems. This allows them to overcome the limitations of traditional empirical models and achieve a shift from "trial and error" to "on-demand design." Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a machine learning-aided design method for structural alloys used in fourth-generation thorium-based molten salt reactors. This method constructs a closed-loop optimization program of "machine learning prediction - phase composition constraint screening - thermodynamic simulation - experimental verification - model feedback update". With machine learning algorithms such as Support Vector Machine (SVM) as the core, it integrates CALPHAD thermodynamic calculation and physical metallurgical constraints, and achieves efficient screening and performance control of alloy composition through multiple rounds of iterative optimization, providing a new data-driven path for the research and development of high-temperature resistant structural alloys for fourth-generation MSRs.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for machine learning-assisted design of structural alloys for fourth-generation thorium-based molten salt reactors, the method specifically including the following steps: S1, establishing a machine learning model for alloy composition prediction based on phase composition: based on known alloy sample data, the sample data including the content of alloy elements and their corresponding phase composition information, a machine learning model for predicting the phase composition of the alloy is constructed; S2, screening the initial alloy composition space through phase composition constraints; S3, secondary screening of alloy composition based on thermodynamic simulation; S4, alloy sample preparation and solution homogenization treatment; S5, conducting mechanical property tests; S6, experimental results feedback and closed-loop update of the machine learning model, and subsequent design based on the model.

[0008] Furthermore, in step S1, firstly, based on metallurgical principles and the service requirements of molten salt reactor structure alloys, the composition space of the alloy design is determined; the composition space alloy system uses Ni as the matrix, and the alloying elements are selected from at least five of Cr, W, Mo, Al, Nb, V, Ta, Zr, C, and B. To ensure that the alloy is Co-free, Co is strictly excluded from the composition; preliminary literature review is conducted to obtain multiple sets of known alloy sample data, which include the content information of each alloying element and its corresponding phase composition information. A machine learning prediction model is established based on feature vectors to output the phase composition results corresponding to the candidate alloy composition. The prediction model is preferably a support vector machine model, which is used to establish a nonlinear mapping relationship between alloy composition features and phase composition.

[0009] Furthermore, in step S1, the process of constructing a machine learning model for predicting the alloy phase composition includes: data preprocessing, model building, and model hyperparameter optimization; the data preprocessing includes: feature construction: using the content of each alloy element as the basic input feature, and further introducing derived features such as atomic size mismatch parameters, electronegativity difference, mixing enthalpy, and valence electron concentration to form a feature vector: (1); among which Indicates the content of the nth alloying element; Feature scaling: To eliminate the influence of different dimensions on model training, min-max normalization is applied to each feature dimension. (2); among which These are the original eigenvalues; and These are the minimum and maximum values ​​of the feature in the sample set, respectively. These are the normalized feature values, limited to the range [0,1]; Feature selection: To reduce the impact of redundant features on model stability, the variance of each feature is further calculated: (3) Remove low-information features with variance below the preset threshold, and at the same time remove highly correlated features.

[0010] Furthermore, the model construction specifically includes: using a radial basis function kernel support vector machine for phase composition classification prediction, with the objective function being: (4); The constraints are satisfied: ; Where ω is the weight vector and b is the bias term. Let C be a slack variable and C be the error penalty factor. The feature mapping function is defined as follows: The radial basis kernel function is defined as: (5); among which This is the kernel width parameter.

[0011] Furthermore, the model hyperparameter optimization specifically includes: determining the optimal parameter combination using grid search combined with K-fold cross-validation. (6); where CV represents the cross-validation classification accuracy; the parameter search range is preferably set to ; C is the error penalty factor.

[0012] Furthermore, in step S2, the initial alloy composition space is screened by phase composition constraints, which specifically includes: based on the machine learning model described in step S1, batch predictions are performed on preset alloy composition combinations to obtain the phase composition results corresponding to each candidate composition; at the same time, phase composition constraints are set, including the upper limit of brittle phase content, the volume fraction range of strengthening phase, and the matrix phase maintaining an fcc structure; the prediction results are compared with the phase composition constraints, and only alloy compositions that meet the constraints are retained as the initial design composition space.

[0013] Furthermore, in step S3, the secondary screening of alloy composition based on thermodynamic simulation specifically includes: performing thermodynamic phase equilibrium simulation on the initial design composition space obtained in step S2, obtaining the phase composition evolution law of each candidate alloy composition in the preset temperature range, and further eliminating alloy compositions with brittle phases existing in the service temperature range or strengthening phases dissolved at the service temperature based on the simulation results, thereby obtaining an optimized set of candidate alloy compositions.

[0014] Further, in step S4, the alloy sample preparation and solution homogenization treatment specifically include: weighing high-purity metal raw materials according to the candidate alloy composition ratio obtained in step S3, with each raw material having a purity of not less than 99.9 wt.%; melting the metal mixture under an inert gas protective atmosphere, preferably high-purity argon, and removing active impurities (oxygen, nitrogen, etc.) from the atmosphere by introducing an active metal for gas purification before melting, preferably titanium; using arc melting or equivalent high-energy melting to completely melt the metal raw materials to form a molten pool, with the single melting temperature being more than 100 ℃ higher than the alloy liquidus temperature; to promote the chemical homogeneity of the alloy composition, repeatedly turning and remelting the obtained ingot, with no less than 5 repetitions; placing the sample in a high-temperature heat treatment furnace for solution homogenization treatment at a temperature of 1100-1250 ℃ and a holding time of 2-6 minutes. h, to eliminate as-cast segregation and promote full diffusion of alloying elements, thereby achieving uniform microstructure; after heat treatment, furnace cooling, air cooling or water cooling is used for cooling; then the completely molten alloy liquid is poured into a water-cooled metal mold for rapid solidification, the mold being preferably a water-cooled copper mold, in order to obtain a dense as-cast structure.

[0015] Furthermore, in step S5, the mechanical property testing specifically includes: processing the homogenized alloy sample into tensile specimens using electrical discharge machining or equivalent precision cutting according to national or industry standards; placing the obtained tensile specimens in a high-temperature tensile testing device and conducting uniaxial tensile tests under different temperature conditions, the test temperatures including at least room temperature, 600 ℃, 700 ℃, and 800 ℃; obtaining the yield strength, tensile strength, and elongation after fracture of each candidate alloy component at different temperatures through the high-temperature tensile test, which is used to evaluate the high-temperature strength and medium-temperature plasticity of the alloy.

[0016] Furthermore, in step S6, the experimental result feedback and machine learning model closed-loop update specifically include: after completing the thermodynamic simulation screening and experimental verification, the obtained real phase composition and mechanical property data are used as new samples to re-input into the support vector machine model to update the training set. ; The model is retrained to form a closed-loop optimization procedure of "machine learning prediction - phase composition constraint screening - thermodynamic simulation - experimental verification - model feedback update", so that the deviation between the prediction results and the experimental results gradually converges, and the alloy composition, phase composition and performance are synergistically optimized.

[0017] The beneficial effects of this invention are as follows: Through machine learning-assisted alloy composition optimization, the fourth-generation structural alloy designed for MSR (Metallurgical Sequencer) achieves significant improvements over existing molten salt reactor structural materials in several key performance indicators. It not only overcomes the plasticity trough problem of existing Ni-W-Cr alloys (such as GH3539) in the 650-750 ℃ ​​mid-temperature range, but also breaks through the upper limit of the service temperature of existing MSR structural alloys, achieving a measured tensile strength of approximately 400 MPa at 800 ℃. This meets the 800 ℃ service temperature requirement for thorium-based molten salt reactors to achieve high-efficiency energy conversion such as high-temperature hydrogen production, providing a feasible material solution for the design and selection of subsequent high-power molten salt reactors. Furthermore, this invention forms a closed-loop optimization model of "machine learning prediction - phase composition constraint screening - thermodynamic simulation - experimental verification - model feedback update," significantly shortening the traditional "trial and error" alloy development cycle and reducing experimental costs. This data-driven materials design paradigm is not only applicable to the research and development of molten salt reactor structural alloys, but can also be extended to the composition screening and performance optimization of advanced materials for other extreme service environments, and has certain demonstration value and prospects for promotion and application.

[0018] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0019] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 is a flowchart of the method of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] Figure 1 is a flowchart of the method of the present invention. As shown in the figure, the method provided by the present invention specifically includes the following steps: S1, establishing a machine learning model for alloy composition prediction based on known alloy sample data, the sample data including the content of alloy elements and their corresponding phase composition information, constructing a machine learning model for predicting the phase composition of alloys; S2, screening the initial alloy composition space through phase composition constraints; S3, secondary screening of alloy composition based on thermodynamic simulation; S4, alloy sample preparation and solid solution homogenization treatment; S5, conducting mechanical property tests; S6, experimental results feedback and closed-loop update of the machine learning model, and subsequent design based on the model.

[0022] In this embodiment, data from over 200 sets of nickel-based superalloy samples were first collected, covering the Ni-Cr-W-Mo-Al-Ti-C system. The sample data included elemental content, phase composition, and corresponding heat treatment regimes.

[0023] A support vector machine (SVM) model for predicting phase composition was established using elemental content, atomic size mismatch parameters, electronegativity difference, enthalpy of mixing, and valence electron concentration as input features. Model parameters were determined through five-fold cross-validation, and the final model achieved an accuracy of over 92% in predicting brittle phases. Based on this, batch predictions were performed on the designed composition space, and the following phase composition constraints were introduced: 1. Brittle phase volume fraction < 1%; 2. Strengthening precipitate system fraction 5-35%; 3. The matrix maintains an fcc structure.

[0024] Through preliminary screening, 28 initial candidate alloy compositions were obtained. Subsequently, thermodynamic phase equilibrium simulations were performed on the above candidate compositions at 600-1600℃. Compositions with stable precipitation of σ phase or Laves phase were eliminated, and the width of the solid-liquid coexistence region was further limited to no more than 80℃. Finally, 3 optimized compositions were retained.

[0025] The experimental results are as follows: Three groups of samples were prepared. Metal raw materials with a purity ≥ 99.9 wt.% were weighed according to the composition, and arc melting was carried out in a high-purity argon protective atmosphere purified by titanium absorption. Each sample was flipped and remelted 8 times to ensure chemical homogeneity. Subsequently, water-cooled copper molds were used for casting to obtain a size of 10 × 5 × 3 mm. 3 The cubic specimens were then subjected to solution homogenization at 1200℃ for 4 h and air-cooled.

[0026] Subsequently, high-temperature tensile specimens were prepared using electrical discharge machining according to national standards, and uniaxial tensile tests were conducted at 600℃, 700℃, and 800℃ to obtain the mechanical properties at these three temperatures. The microstructure of the alloy meeting service requirements was then characterized using microscopy, and the results were cross-validated with machine learning data.

[0027] Technical Performance Analysis: The alloy obtained in this embodiment maintains stable plasticity over a wide temperature range of 600-800℃, without exhibiting the plasticity trough commonly seen in traditional Ni-W-Cr alloys in the 650-750℃ range. In particular, its tensile strength reaches approximately 400MPa at 800℃, significantly exceeding the level of existing molten salt reactor structural materials, thus meeting the 800℃ service requirements for high-temperature hydrogen production and other applications.

[0028] The results validate the effectiveness of the collaborative design approach of "machine learning prediction - phase composition constraint - thermodynamic simulation - experimental verification" proposed in this invention.

[0029] In summary, the novel structural alloy developed using the method of this invention achieves a tensile strength exceeding 400 MPa at an service temperature of 800℃, meeting the mechanical performance requirements of high-temperature structural components in molten salt reactors. Simultaneously, this alloy exhibits excellent plasticity in the mid-temperature range of 650-750℃, overcoming the plasticity trough defect present in existing alloys within this temperature range. This invention organically integrates artificial intelligence technology with traditional alloy design theory, providing an efficient and reliable new path for the development of key structural materials for fourth-generation thorium-based molten salt reactors, possessing both significant scientific research value and promising engineering applications.

[0030] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention.

Claims

1. A method for machine learning-aided design of structural alloys for fourth-generation thorium-based molten salt reactors, characterized in that: The method specifically includes the following steps: S1, establishing a machine learning model for alloy composition prediction based on phase composition: based on known alloy sample data, including the content of alloy elements and their corresponding phase composition information, a machine learning model for predicting the phase composition of the alloy is constructed; S2, screening the initial alloy composition space through phase composition constraints; S3, secondary screening of alloy composition based on thermodynamic simulation; S4, alloy sample preparation and solution homogenization treatment; S5, mechanical property testing; S6, experimental results feedback and closed-loop update of the machine learning model, and subsequent design based on the model; in step S1, firstly, based on metallurgical principles and the service requirements of molten salt stack structure alloys, the composition space of the alloy design is determined; the composition space alloy system uses Ni as the matrix, and the alloying elements are selected from Cr, W, Mo, Al, Nb, V, and Ta. At least five of the following elements are selected: Zr, C, and B. To ensure the alloy is designed to be Co-free, Co is strictly excluded from the composition. Multiple sets of known alloy sample data are acquired, including the content information of each alloy element and its corresponding phase composition information. A machine learning prediction model is established based on feature vectors to output the phase composition results corresponding to the candidate alloy composition. The prediction model is preferably a support vector machine model, used to establish a nonlinear mapping relationship between alloy composition features and phase composition. In step S1, the process of constructing the machine learning model for predicting the alloy phase composition includes: data preprocessing, model construction, and model hyperparameter optimization. The data preprocessing includes: feature construction: using the content of each alloy element as the basic input feature, and further introducing derived features such as atomic size mismatch parameters, electronegativity difference, mixing enthalpy, and valence electron concentration to construct a feature vector. (1); where: For a single alloy sample, the input feature vector is... This indicates the content of the nth alloying element. This represents the number of element types contained in the alloy system. This is the atomic size mismatch parameter. It is a parameter for poor electronegativity. For the enthalpy of mixture, Valence electron concentration; Feature scaling: To eliminate the influence of different dimensions on model training, min-max normalization is applied to each feature dimension. (2); among which These are the original eigenvalues; and These are the minimum and maximum values ​​of the feature in the sample set, respectively. These are the normalized feature values, limited to the range [0,1]; Feature selection: To reduce the impact of redundant features on model stability, the variance of each feature is further calculated: (3); where: For the first The variance of each feature, For the first The sample at the th The values ​​of each feature For the first The mean of each feature across all samples The total number of samples is used; low-variance features are removed based on a variance threshold; simultaneously, the correlation coefficient between features is calculated, and highly correlated features are deredundant to retain the feature set that contributes more significantly to phase composition and performance prediction; the model construction specifically includes: using a radial basis function kernel support vector machine for phase composition classification prediction, with the objective function being: (4); The constraints are satisfied: ； ;in The weight vector of the classification hyperplane. For the bias term of the classification hyperplane, For the first The slack variables corresponding to each sample are used to allow some samples to have classification errors under soft-margin conditions; This is the error penalty factor, used to balance the relationship between maximizing the margin and minimizing the classification error; For the first The input feature vector of each sample, For the first The category labels of each sample; It is a nonlinear mapping function that maps the original input space to a high-dimensional feature space; This represents the number of training samples; Let represent the squared L2 norm of the weight vector, used to control model complexity; where, The larger the value, the stronger the model's penalty for misclassified samples, and the more it tends to improve the fit of the training set; When the value is small, the model emphasizes the smoothness of the classification margin and its generalization ability; the radial basis kernel function is defined as: (5); where: For the sample With sample The kernel function value; For two input samples whose similarity needs to be calculated; The squared Euclidean distance between the two samples; The kernel width parameter controls the extent to which a single sample influences the classification boundary. When the kernel size is large, the effective range of the kernel function is smaller, and the decision boundary of the model becomes more complex; When the kernel size is small, the kernel function has a larger range of influence, resulting in smoother model boundaries.

2. The method for machine learning-aided design of structural alloys for fourth-generation thorium-based molten salt reactors according to claim 1, characterized in that: The model hyperparameter optimization specifically includes: using grid search combined with K-fold cross-validation to optimize hyperparameters, with the objective expressed as: (6); where: The optimal error penalty factor is... The optimal kernel width parameter, In parameter combination The classification accuracy or comprehensive evaluation index obtained through K-fold cross-validation; This represents the combination of parameters that maximizes the objective function; the preferred parameter search range is: ； Discrete values ​​are obtained by using an exponential step size to improve optimization efficiency and parameter selection accuracy.

3. The method for machine learning-aided design of structural alloys for fourth-generation thorium-based molten salt reactors according to claim 2, characterized in that: In step S2, the initial alloy composition space is screened by phase composition constraints, which specifically includes: based on the machine learning model described in step S1, batch predictions are performed on preset alloy composition combinations to obtain the phase composition results corresponding to each candidate composition; at the same time, phase composition constraints are set, including the upper limit of brittle phase content, the volume fraction range of strengthening phase, and the matrix phase maintaining an fcc structure; the prediction results are compared with the phase composition constraints, and only alloy compositions that meet the constraints are retained as the initial design composition space.

4. The method for machine learning-aided design of structural alloys for fourth-generation thorium-based molten salt reactors according to claim 3, characterized in that: In step S3, the secondary screening of alloy composition based on thermodynamic simulation specifically includes: performing thermodynamic phase equilibrium simulation on the initial design composition space obtained in step S2, obtaining the phase composition evolution law of each candidate alloy composition in the preset temperature range, and further eliminating alloy compositions with brittle phases existing in the service temperature range or strengthening phases dissolved at the service temperature based on the simulation results, thereby obtaining an optimized set of candidate alloy compositions.

5. The method for machine learning-aided design of structural alloys for fourth-generation thorium-based molten salt reactors according to claim 4, characterized in that: In step S4, the alloy sample preparation and solution homogenization treatment specifically include: weighing high-purity metal raw materials according to the candidate alloy composition ratio obtained in step S3, with each raw material having a purity of not less than 99.9 wt.%; melting the metal mixture under an inert gas protective atmosphere, preferably high-purity argon, and removing active impurities in the atmosphere by introducing an active metal for gas purification before melting, preferably titanium; using arc melting or equivalent high-energy melting to completely melt the metal raw materials to form a molten pool, with the single melting temperature being more than 100 ℃ higher than the alloy liquidus temperature; repeatedly turning and remelting the resulting ingot multiple times, with no less than 5 repetitions; and placing the sample in a high-temperature heat treatment furnace for solution homogenization treatment at a temperature of 1100-1250 ℃ and a holding time of 2-6 hours. h; After heat treatment, the alloy is cooled by furnace cooling, air cooling or water cooling; then the completely molten alloy liquid is poured into a water-cooled metal mold for rapid solidification, preferably a water-cooled copper mold, to obtain a dense as-cast structure.

6. The method for machine learning-aided design of structural alloys for fourth-generation thorium-based molten salt reactors according to claim 5, characterized in that: In step S5, the mechanical property testing specifically includes: processing the homogenized alloy sample into tensile specimens using electrical discharge machining or equivalent precision cutting according to national or industry standards; placing the obtained tensile specimens in a high-temperature tensile testing device and conducting uniaxial tensile tests under different temperature conditions, the test temperatures including at least room temperature, 600 ℃, 700 ℃, and 800 ℃; obtaining the yield strength, tensile strength, and elongation after fracture of each candidate alloy component at different temperatures through the high-temperature tensile test, which is used to evaluate the high-temperature strength and medium-temperature plasticity of the alloy.

7. The method for machine learning-aided design of structural alloys for fourth-generation thorium-based molten salt reactors according to claim 6, characterized in that: In step S6, the experimental result feedback and machine learning model closed-loop update specifically include: after completing the thermodynamic simulation screening and experimental verification, the obtained real phase composition and mechanical property data are used as new samples to re-input into the support vector machine model to update the training set. ； ;in, 、 This represents the original training sample set before the closed-loop update. 、 This represents the newly obtained sample set after thermodynamic simulation screening and experimental verification. 、 This indicates the updated training sample set after introducing new samples; the model is retrained to form a closed-loop optimization procedure of "machine learning prediction - phase composition constraint screening - thermodynamic simulation - experimental verification - model feedback update", so that the deviation between the prediction results and the experimental results gradually converges, and the synergistic optimization of alloy composition, phase composition and performance is achieved.

Images