Molten salt eutectic formula design model and establishment method and application thereof

Abstract

The invention discloses a fused salt eutectic formula design model and an establishment method and application thereof, and belongs to the technical field of fused salt system design, the establishment method of the fused salt eutectic formula design model comprises the following steps: collecting known single-component salt and fused salt system data, establishing a database, and performing normalization processing on the data to obtain a fused salt eutectic formula design model; performing model training and verification by using a machine learning algorithm and data in the database, optimizing the model, extracting high-correlation characteristics, continuing model optimization, and finally evaluating and screening the optimized model to obtain a model with good prediction accuracy, namely the fused salt eutectic formula design model. The model is used for fused salt eutectic formula design, the problems that a traditional method is long in research time and low in efficiency are solved, prediction of the fused salt melting point is improved to be within dozens of seconds, and very high accuracy (gt; 90%) and universality.

Description

Technical Field

[0001] This invention relates to the field of molten salt system design technology, and in particular to a molten salt eutectic formulation design model, its establishment method, and its application. Background Technology

[0002] In recent years, with the rapid growth of installed capacity for photovoltaic, wind power, and solar thermal power, renewable energy power generation has increased rapidly. However, due to the influence of natural conditions, wind and solar power generation are highly volatile and intermittent, and large-scale grid connection can have a significant negative impact on the safety and stability of the power grid. Against this backdrop, vigorously developing corresponding supporting energy storage technologies can help solve the problem of intermittent and fluctuating inter-temporal, distributed, and high-efficiency utilization of solar, wind, electricity, and thermal energy sources.

[0003] Molten salt thermal energy storage boasts advantages such as high capacity density, excellent heat transfer and storage performance, long lifespan, low cost, good peak-shaving performance, and strong grid compatibility, making it a preferred energy storage technology for large-scale concentrated solar power (CSP), deep peak shaving in thermal power plants, renewable energy power generation storage, and multi-energy complementarity in integrated energy systems. However, the current nitrate system used for thermal energy storage suffers from limiting bottlenecks due to its high melting point and narrow temperature range, resulting in high CSP costs and severely restricting the development of high-capacity molten salt thermal energy storage systems. Therefore, there is an urgent need to design and develop low-melting-point molten salts.

[0004] Generally speaking, the trial-and-error method of mixing multi-component molten salts is a common method used by researchers to obtain low-melting-point, wide-temperature-range, high-temperature, and large-capacity heat storage materials. Commonly used design methods for low-melting-point molten salts include: (1) Trial-and-error method: By testing a large number of molten salts with different components in the same system, the molten salt formula with the lowest eutectic temperature is obtained; (2) Thermodynamic phase diagram calculation method: The phase diagram of the entire system is calculated by thermodynamic methods, thereby obtaining the molten salt formula with the lowest eutectic temperature; (3) Empirical formula method: Based on existing experimental data and practical experience, by studying and analyzing various properties of the molten salt system, some empirical formulas that can characterize the melting point and other properties of molten salts are summarized, and then low-melting-point molten salts are designed based on these formulas. Although the prediction speed is fast, the accuracy is poor, and the error can be as high as 100% or more.

[0005] The two methods (1) and (2) above usually require a lot of experiments or calculations to obtain the lowest eutectic point of a new system, which takes at least several days to more than a month. Both methods have the problems of long research time and low efficiency. Method (3) has the problem of large error. Summary of the Invention

[0006] To address the issues of long time or large error in the design of molten salt eutectic formulations, this invention provides a method for establishing a molten salt eutectic formulation design model. The molten salt eutectic formulation design model established by this method can be used to predict multi-component molten salt systems, which has the advantages of high efficiency, high accuracy and strong universality.

[0007] On one hand, this invention discloses a method for establishing a molten salt eutectic formulation design model, comprising the following steps:

[0008] S1. Establish the database:

[0009] Collect characteristic parameters of each atom and ion in the single-component salt, including structural characteristic parameters and / or interaction characteristic parameters of the atoms and ions; collect molten salt composition and / or phase transition temperature data of known multi-component molten salt systems composed of the single-component salts to form a dataset;

[0010] S2. Data Processing:

[0011] The feature parameters are normalized, which involves converting the data into dimensionless feature parameter values ​​in the range of 0-1.

[0012] S3. Model Training

[0013] The normalized dataset is divided into a training set, a validation set, and a test set;

[0014] The training set data is obtained, the feature parameters are used as input data, the molten salt composition and / or phase transition temperature data are used as output data, a model is established, the model is trained with several machine learning algorithms, and the model parameters are adjusted with an optimized loss function to obtain several training models.

[0015] Obtain the data from the validation set and the data from the test set. Use the data from the validation set to optimize the hyperparameters of the trained model, and use the data from the test set to evaluate the model performance, thereby obtaining the coefficient of determination R. 2 Screening for the coefficient of determination R 2 The model that reaches the threshold is the initial screening model;

[0016] S4. Feature Engineering

[0017] Principal component analysis is used to extract the feature parameters that are highly correlated with the output data in the initial screening model as highly correlated feature parameters. By transforming the feature parameters, combined feature parameters that can reflect the strength and type of inter-ion interactions in the multi-component molten salt system are obtained. The highly correlated feature parameters and combined feature parameters are used as input data for subsequent model optimization.

[0018] S5. Model Optimization

[0019] The parameters of each model in the preliminary screening model obtained above are optimized using the validation set, and the optimized model is evaluated by substituting the data from the test set. The model with the best prediction accuracy is selected, which is the molten salt eutectic formulation design model.

[0020] The above-mentioned method for establishing the molten salt eutectic formulation design model first involves collecting a certain amount of characteristic parameters that characterize the atomic and ionic properties of single-component salts from various types of literature, such as journal articles, master's and doctoral dissertations, patent documents, technical manuals, books, technical reports, and technical dictionaries. These parameters serve as feature values ​​for machine learning. Then, data on the molten salt combination composition and / or phase transition temperature (such as eutectic point ratio, minimum eutectic melting temperature, etc.) of the corresponding multi-component molten salt system are collected as target values ​​for machine learning, forming a database. After appropriate data cleaning of the collected feature values ​​and target values, they are used as input and output data, respectively, and input into the machine learning model for training to establish a prediction algorithm. Next, feature engineering is performed on the initial screening model to extract highly relevant feature parameters that reflect the combined feature parameters constituting the internal interactions of the multi-component molten salt system. The model is then optimized again, and finally, a model with accurate predictions is selected, which is the molten salt eutectic formulation design model.

[0021] When using this model to design molten salt eutectic formulations, the characteristic parameter values ​​of each single-component salt in the unknown system can be input into the trained model for prediction, thereby obtaining the eutectic formulation of the new molten salt system and the corresponding molten salt combination composition and / or phase transition temperature data.

[0022] In step S1 of some embodiments, the single-component salt has a melting point ≤1050℃, a boiling point ≥380℃, and a decomposition temperature ≥380℃.

[0023] In step S1 of some embodiments, the single-component salt includes at least one of NaNO3, KNO3, Ca(NO3)2, Na2CO3, K2CO3, NaCl, KCl, Na2SO4, and K2SO4.

[0024] In step S1 of some embodiments, the structural characteristic parameters include at least one of the following: atomic number Z, group F, period pe, block, number of protons p, number of electrons s, number of neutrons N, atomic mass M, mass number A, number of outer electrons V, atomic radius ra, covalent radius rco, van der Waals radius rf, metallic radius rm, ionic radius ri, atomic volume Va, and coordination number Cn.

[0025] In step S1 of some embodiments, the interaction characteristic parameters include at least one of: atomic electronegativity E, atomic ionization potential ip, n1 / 3ws value, Miedema electronegativity EM, ionization energy ie, atomic affinity a, atomic polarizability αa, ionic polarizability αi, ionic electronegativity Ei, and lattice energy E.

[0026] In step S1 of some embodiments, the molten salt composition and / or phase transition temperature data includes at least one of the following: minimum eutectic point ratio, minimum eutectic point melting temperature, minimum eutectic point density, and minimum eutectic point specific heat.

[0027] In step S1 of some embodiments, the molten salt eutectic formulation is a binary molten salt system composed of NaNO3 and Ca(NO3)2, and the molten salt composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M.

[0028] In step S1 of some embodiments, the molten salt eutectic formulation is a ternary molten salt system composed of NaNO3, KNO3 and KCl, and the molten salt composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M.

[0029] In step S2 of some implementation schemes, a data cleaning step is performed before data normalization. The data cleaning step includes a screening step and / or an optimization step. The screening step is to correct or remove erroneous data based on the reasonable range and pattern of each output data and input data. The optimization step is to count the frequency of each parameter value when different parameter values ​​of the same data are collected from different sources, and take the parameter value with the highest frequency as the data value of that data.

[0030] In step S2 of some implementation schemes, the screening step includes the following steps:

[0031] Analyze and judge each data point; if it meets the exclusion criteria, exclude the data; if it meets the correction criteria, correct the data.

[0032] The exclusion criteria are as follows: data will be excluded if it meets any of the following conditions:

[0033] Scenario 1: The data values ​​violate physical and / or chemical laws;

[0034] Scenario 2: The relevant variables in the data are missing;

[0035] Scenario 3: The data source indicates that the data is unreliable;

[0036] Scenario 4: Data that appears for the second or subsequent times;

[0037] Scenario 5: Extreme values ​​in statistics or physics, which are manually verified as erroneous;

[0038] Case 6: The sum of the components in a multi-component system deviates from 100% by more than 1%;

[0039] The correction criteria and methods are as follows: when data falls into the following categories, corrections shall be made according to the correction methods for those categories:

[0040] I. Unit Correction: When the unit of a data value is inconsistent with the standard unit of that type of data, it is converted to the corresponding value in the standard unit.

[0041] II. Systematic Error Correction: When the data source shows that the data from that source has a systematic bias, the data from that source will be corrected by deducting the offset.

[0042] III. Component Normalization Correction: When the sum of the components in a multi-component system deviates from 100% by ≤1%, normalization is performed according to the proportion of each component. After correction, the sum of the proportions of each component equals 100%.

[0043] In step S2 of some implementation schemes, the normalization process is performed by converting according to the following formula:

[0044]

[0045] in:

[0046] x represents the parameter value of one of the features in the feature parameters;

[0047] x min This represents the minimum value of the feature;

[0048] x max This represents the maximum value of the feature;

[0049] x norm This represents the normalized feature parameter values.

[0050] In step S3 of some implementation schemes, the machine learning algorithm includes: negative feedback artificial neural network algorithm, support vector machine algorithm, ridge regression algorithm, nearest neighbor analysis algorithm, decision tree algorithm, or random forest algorithm.

[0051] In step S3 of some implementation schemes, the optimized loss function is selected from at least one of mean absolute error, Huber loss, and mean squared error.

[0052] In step S3 of some implementation schemes, the determination coefficient R 2 Reaching the threshold refers to R 2 ≥0.8.

[0053] In step S3 of some embodiments, the molten salt eutectic formulation is a binary molten salt system composed of NaNO3 and Ca(NO3)2, the molten salt composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature, the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M, and the machine learning algorithm is a negative feedback artificial neural network algorithm.

[0054] In step S3 of some embodiments, the molten salt eutectic formulation is a ternary molten salt system composed of NaNO3, KNO3 and KCl, the molten salt composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature, the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M, and the machine learning algorithm is a negative feedback artificial neural network algorithm.

[0055] In step S4 of some implementation schemes, the principal component analysis is calculated using the following formula:

[0056] Xnew=X·V

[0057] in:

[0058] X represents the initial data matrix.

[0059] V represents a matrix composed of principal components, each of which is a linear combination of the initial data;

[0060] Xnew represents the data matrix after dimensionality reduction.

[0061] In step S4 of some implementation schemes, the highly correlated feature parameter meets the following condition:

[0062] I: Eigenvalue > 1;

[0063] II: The cumulative variance contribution rate of all highly correlated characteristic parameters is ≥70% to 80%;

[0064] III: Combined load factor >|0.7|.

[0065] In step S4 of some embodiments, the combined characteristic parameters include at least one of the following: ionic radius ratio, ionic radius to electronegativity ratio, and atomic coefficient to ionic radius ratio.

[0066] In step S5 of some implementation schemes, the optimization method is at least one of grid search and random search; preferably, grid search.

[0067] In step S5 of some implementations, the grid search is to systematically traverse the hyperparameter space to find the optimal combination of hyperparameters.

[0068] In step S5 of some implementations, the random search is to randomly sample the hyperparameter space to obtain the optimal combination of hyperparameters.

[0069] In step S5 of some implementation schemes, the model parameters include at least one of: solver, activation function, number of neural layers, number of nodes, error function, learning rate, and momentum.

[0070] In some implementations, the method for establishing the molten salt eutectic formulation design model further includes a model iteration step S6, in which steps S3-S5 are repeated to dynamically optimize the prediction effect and establish a prediction model with an accuracy greater than a threshold.

[0071] In step S6 of some implementations, the accuracy threshold is 90%.

[0072] On the other hand, the present invention also discloses a molten salt eutectic formulation design model, which is established by the above-mentioned method for establishing a molten salt eutectic formulation design model.

[0073] In some embodiments, the single-component salt includes at least one of NaNO3, KNO3, Ca(NO3)2, Na2CO3, K2CO3, NaCl, KCl, Na2SO4, and K2SO4.

[0074] In some embodiments, the molten salt eutectic formulation is a binary molten salt system composed of NaNO3 and Ca(NO3)2, and the molten salt composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M, lattice energy E, coordination number Cn and the melting point T of the single-component salt.

[0075] In some embodiments, the molten salt eutectic formulation is a ternary molten salt system composed of NaNO3, KNO3 and KCl, and the molten salt composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa, atomic mass M, lattice energy E and coordination number Cn.

[0076] In some implementations, the machine learning algorithm is a negative feedback artificial neural network.

[0077] In some of these implementations, the determination coefficient R 2 ≥0.8.

[0078] In some of these implementations, the accuracy rate is ≥90%.

[0079] On the other hand, the present invention also discloses a molten salt eutectic formulation design method based on machine learning, comprising the following steps: inputting the characteristic parameters of each single component salt in the multi-component molten salt system to be predicted into the above-mentioned molten salt eutectic formulation design model, calculating with the model, and outputting the predicted molten salt combination composition and / or phase transition temperature data of the multi-component molten salt system.

[0080] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of the present invention.

[0081] The reagents and raw materials used in this invention are all commercially available.

[0082] The positive and progressive effects of this invention are as follows:

[0083] The present invention discloses a method for establishing a molten salt eutectic formulation design model. This model utilizes machine learning training to establish a molten salt eutectic formulation design model. This model can accurately obtain the formulation and property data of eutectic salts using only simple and readily available parameters, with an accuracy rate higher than 90%, even reaching over 99%. Furthermore, it has strong predictive universality. The model maintains a prediction error of less than 10% for binary and ternary molten salts, as well as for molten salt combinations of different cations and anions.

[0084] Furthermore, compared to traditional research methods, the model of this invention is used for design prediction, and the prediction speed is only tens of seconds (e.g., less than 30 seconds), which significantly improves efficiency. Attached Figure Description

[0085] Figure 1 The method and steps for establishing the molten salt eutectic formulation design model for Example 1 are as follows.

[0086] Figure 2 The image shows the DSC scan of the binary molten salt system NaNO3-Ca(NO3)2 in Example 3.

[0087] Figure 3 This is a DSC scan of the ternary molten salt system NaNO3-KNO3-KCl in Example 4. Detailed Implementation

[0088] The present invention is further illustrated below by way of embodiments, but the invention is not limited to the scope of the embodiments described herein. Experimental methods in the following embodiments that do not specify specific conditions were performed according to conventional methods and conditions, or as selected according to the product instructions.

[0089] Example 1

[0090] A molten salt eutectic formulation design model, such as Figure 1 As shown, it is established through the following steps:

[0091] S1. Establish the database

[0092] Data is collected through various means such as literature, patents, journals, technical dictionaries and manuals, and experimental tests. The data mainly includes two categories: characteristic parameters that constitute the input data and multi-component molten salt system data that constitute the output data (such as molten salt composition and / or phase transition temperature parameters).

[0093] (1) Screening of single-component salts:

[0094] The screening of single-component salts used multiple indicators such as melting point, boiling point and decomposition temperature, price, and toxicity to initially select 38 single-component salts from a large number of salts. The screening conditions included: the highest melting point of the single-component salt must be less than 1050℃, the boiling point and decomposition temperature must be greater than 380℃, the price must not exceed 8 times that of reagent-grade nitrates, and the toxicity must be low.

[0095] Furthermore, by comprehensively considering energy storage density and energy storage cost, the following potential single-component salts were finally determined to be included in this embodiment: NaNO3, KNO3, Ca(NO3)2, Na2CO3, K2CO3, NaCl, KCl, Na2SO4, K2SO4, etc.

[0096] (2) Data collection for multi-component molten salt systems:

[0097] Collect molten salt composition and / or phase transition temperature data of known multi-component molten salts composed of the above-mentioned single-component salts. In this embodiment, the molten salt composition and / or phase transition temperature data include: minimum eutectic point ratio, minimum eutectic point melting temperature, minimum eutectic point density, and minimum eutectic point specific heat, which are used as target parameters for training, i.e., the output data of the model.

[0098] (3) Feature parameter collection:

[0099] Building molten salt eutectic formulation design models using machine learning algorithms presents significant challenges. For instance, machine learning models cannot directly understand chemical formulas like "LiF-BeF2". The lava system represented by this chemical formula needs to be "translated" into a quantitative digital language that the model can understand, in a way that machine learning can utilize.

[0100] Specifically, in this embodiment, a comprehensive and quantitative digital description of the entire molten salt system is established by describing the basic properties (such as structural features) of the individual ions (atoms) that constitute the molten salt and their tendency and manner of interaction (such as interaction features), thereby predicting its macroscopic properties (such as density, specific heat, melting point, or reverse prediction of molten salt composition and / or phase transition temperature parameters).

[0101] When describing the structural characteristics of ions, the primary consideration is to select parameters that reflect the ion's basic identity, mass, and spatial dimensions, as these are the foundation for constructing any material model and represent the static characteristics describing the structure. When describing the interactions between ions—that is, their dynamic and chemical properties—the primary consideration is to identify features that describe the source and strength of the interionic forces, as these are strongly correlated with the structure and properties of molten salts.

[0102] Based on the above analysis and considerations, and after screening and adjustments by the inventors, in this embodiment, the feature parameters collected as model input data involve the atomic and ionic parameters of the molten salt system, including:

[0103] Structural characteristic parameters: atomic number (Z), group (F), period (pe), block (Block), number of protons (p), number of electrons (s), number of neutrons (N), atomic mass (M), mass number (A), number of outer electrons (V), atomic radius (ra), covalent radius (rco), van der Waals radius (rf), metallic radius (rm), ionic radius (ri), atomic volume (Va), etc.

[0104] Interaction characteristic parameters: atomic electronegativity (E), atomic ionization potential (ip), n1 / 3ws value, Miedema electronegativity (EM), ionization energy (ie), atomic affinity (a), atomic polarizability (αa), ionic polarizability (αi), ionic electronegativity (Ei), etc.

[0105] S2. Data Processing

[0106] S21. Data Cleaning

[0107] The collected dataset may contain erroneous values. This embodiment uses the following method to identify and select the best values ​​to improve the quality of the dataset and ensure the accuracy of subsequent modeling.

[0108] 1) Screening Steps

[0109] Based on the reasonable range and patterns of each output and input data, erroneous data is corrected or eliminated. Specifically, each data point is analyzed and judged; if it meets the elimination criteria, it is eliminated; if it meets the correction criteria, it is corrected.

[0110] A. Removal Criteria: Data shall be removed if it meets any of the following conditions.

[0111] Scenario 1: The data values ​​violate physical and / or chemical laws.

[0112] Data values ​​that violate basic physical laws or chemical common sense should be discarded immediately. These types of errors are usually caused by recording mistakes or serious malfunctions in measuring equipment. For example: negative values ​​for properties such as density, viscosity, and heat capacity; the sum of the mole fraction or mass fraction of components being much greater than 1 (e.g., >1.05) or much less than 1 (e.g., <0.95), and it is impossible to determine which component was recorded incorrectly; the measurement temperature of a molten salt at normal pressure being much lower than its accepted melting point, etc.

[0113] Scenario 2: The relevant variables of the data are missing.

[0114] Understandably, some data points are correlated with other variables. For example, a data point (usually the dependent variable, such as a property) may lack key independent variables that have a decisive influence on it. Without these independent variables, the data point loses its contextual meaning and cannot be used for model training; therefore, it should be removed. For instance, the viscosity value of molten salt may be given, but the corresponding temperature may not be recorded; or the density of a ternary molten salt system may be given, but only the ratio of the two components may be recorded, lacking complete component information.

[0115] Scenario 3: The data source indicates that the data is unreliable.

[0116] In some cases, if the authors explicitly state in the literature or report from which the data is sourced that a particular data point or series is unreliable due to experimental contamination, equipment malfunction, methodological defects, or other reasons, and do not provide a correction method, such data is unreliable and should be discarded. For example, the literature may mention that "the data for this component may be too high due to crucible corrosion," or the authors may note that "the temperature controller malfunctioned during the measurement process, resulting in inaccurate temperature readings."

[0117] Scenario 4: Data that appears for the second or subsequent times.

[0118] If a dataset contains two or more identical entries (with the same composition, temperature, pressure, and property values), this may be due to multiple mergings of datasets from different sources. Therefore, to prevent the model from assigning excessive weights to these points during training and to avoid overfitting, duplicate entries should be removed, keeping only one.

[0119] Scenario 5: Extreme values ​​in statistics or physics, and data that have been manually verified as erroneous.

[0120] If a data point is statistically or physically an extreme outlier, its source data should usually be traced back to check for entry errors. After manual review, data deemed erroneous should be removed. For example, the viscosity of a certain component might suddenly be three orders of magnitude higher than surrounding data points.

[0121] Case 6: The sum of the components in a multi-component system deviates from 100% by more than 1%.

[0122] When the sum of the components in a multi-component system exceeds 100% by a significant margin, it indicates that the data contains a large error and should not be used.

[0123] B. Correction criteria and correction methods: When data falls into the following categories, correction shall be performed according to the correction method for that category.

[0124] If data errors are clear, predictable, and can be corrected using reliable rules, then correction should be chosen to improve database quality. Specifically, the following rules can be followed.

[0125] I. Unit Correction: When the unit of a data value is inconsistent with the standard unit of that type of data, it is converted to the corresponding value in the standard unit.

[0126] This is the most common and easiest mistake to correct: different documents may use different units or symbols. For example, some temperature units are Celsius, while the model needs to be unified to Kelvin. Understandably, unifying units establishes a standard unit system (e.g., temperature in K, pressure in Pa, density in g / cm³). 3 Then, write a script to unify all data under this system, such as the unified formula for temperature: T(K)=T(℃)+273.15.

[0127] II. Systematic Error Correction: When the data source shows that the data from that source has a systematic bias, the data from that source is corrected by deducting the offset.

[0128] For example, if data from a particular paper or laboratory is consistently too high or too low, comparison with a large amount of high-precision data can reveal that the data from a certain source consistently deviates from a fixed percentage or value—that is, the offset is known. In this case, batch corrections can be performed based on the known offset. For example, subtracting 5K from all temperature data from that source.

[0129] III. Component Normalization Correction: When the sum of the components in a multi-component system deviates from 100% by ≤1%, normalization is performed according to the proportion of each component. After correction, the sum of the proportions of each component equals 100%.

[0130] Due to rounding or reporting accuracy issues, the sum of the components in a multi-component system is not exactly equal to 1 (or 100%), but the deviation is very small (e.g., between 0.99 and 1.01). For example, a ternary system with reported components A = 0.40, B = 0.30, and C = 0.31 has a sum of 1.01. Proportional normalization can be performed by dividing each component value by their sum, so that the corrected sum of the components will exactly equal 1.

[0131] The formula is: xA'=xA / (xA+xB+xC)

[0132] Where x is the original proportion value of each component (such as xA, xB, xC, etc.), x' is the normalized and corrected proportion value of the component (such as xA', xB', xC', etc.), and the denominator (xA+xB+xC) is the sum of the original proportion values ​​of all components.

[0133] 2) Optimal selection step

[0134] When faced with multiple sources providing different values, the reliability and consistency of the data can be ensured by statistically analyzing the frequency of each source and prioritizing the values ​​obtained from the majority of sources.

[0135] S22. Data normalization processing.

[0136] Normalization transforms data to the range [0,1] to eliminate the influence of different units and numerical ranges, ensuring that all features are at the same scale during model training. Specifically, the max-min normalization method is used, with the following formula:

[0137]

[0138] in:

[0139] x represents the parameter value of one of the features in the feature parameters;

[0140] xmin represents the minimum value of this feature;

[0141] xmax represents the maximum value of this feature;

[0142] xnorm represents the normalized feature parameter values.

[0143] S3. Model Training

[0144] The dataset was divided into training, validation, and test sets according to conventional methods in the field for model training, hyperparameter tuning, and model evaluation. Specifically, a test set was reserved for model evaluation at an 8:2 ratio of Non-Test to Test set. Non-Test sets were then subjected to Leave-One-Out Cross-Validation (LOOCV), randomly divided into 10 groups, with one set used as the validation set and the other nine as the training set for cross-validation.

[0145] S31. Model Training

[0146] This embodiment employs various model algorithms, including SVM (Support Vector Machine), Ridge (Ridge Regression), Negative Feedback ANN (Neural Network with Negative Feedback), KNN (K-Nearest Neighbors), and DTR (Decision Tree Regression), to train the training set data. The specific algorithm formulas follow the conventional algorithm formulas in this field. If adjustments are needed, they can be made using general methods.

[0147] For example, the sigmoid activation function used is σ(x) = 1 / (1 + e^(-x)), and the momentum is... θi = θi-1 - vi, where vi is the current velocity and γ is the momentum parameter, which is a positive number less than 1.

[0148] The parameters are then adjusted by optimizing the loss function (mean squared error, MSE function, is used in this embodiment) to obtain the trained model. Similarly, this process can be performed using the commonly used mean squared error, MSE function in this field.

[0149] S32. Model Evaluation and Optimization

[0150] The training model is used to tune its hyperparameters using a validation set, and the model is evaluated using a test set. Specifically, the coefficient of determination R is used as follows: 2 Conduct an assessment:

[0151]

[0152] Where yi is the true value. For predicted values, is the mean of the true values, and n is the sample size.

[0153] The model was optimized by tuning hyperparameters, and R was ultimately selected from these different algorithms. 2 Models greater than 0.8

[0154] S4. Feature Engineering

[0155] The main goal of feature engineering is to enhance the model's ability to capture key information from the data through feature selection and transformation, thereby improving the model's predictive performance. By extracting features that are more strongly correlated with the target variable, overfitting can be reduced and the model's generalization ability can be improved.

[0156] S41. Feature Selection

[0157] In this embodiment, to further improve the model's prediction performance, PCA (Principal Component Analysis) is first used for dimensionality reduction and feature extraction. PCA is a linear dimensionality reduction technique that maximizes the variance of the data on the new coordinate axes by projecting the data onto a new coordinate system. The formula for PCA is as follows:

[0158] Xnew=X·V

[0159] in:

[0160] X represents the original data matrix.

[0161] V represents a matrix composed of principal components, each of which is a linear combination of the original data.

[0162] Xnew represents the data matrix after dimensionality reduction.

[0163] The specific methods and conditions for the principal component analysis described above can be followed in accordance with conventional practices in the art. In this embodiment, features that meet the following conditions are selected as highly correlated feature parameters after analysis:

[0164] (1) Principal components (or factors) with eigenvalues ​​> 1: to ensure their effective explanatory power;

[0165] (2) The top k features with a cumulative variance contribution rate of ≥70% to 80% indicate that the variables as a whole are strongly correlated (which can be summarized by a few components);

[0166] (3) Combining the characteristic of loading coefficient >|0.7|, it is possible to locate which specific variables are highly correlated on common components and identify the group of highly correlated variables.

[0167] S42. Feature Transformation

[0168] Furthermore, based on their own practical experience, the inventors considered that many physical and chemical laws are not determined by absolute values ​​but by relative values. Therefore, they designed the concept of "combined features". By combining and transforming different features, such as taking the ratio of two feature parameters as a new feature to describe the ionic structure of mutual interaction, they created innovative features that can more effectively describe the strength and type of interaction between ions.

[0169] For example, using the ratio of ionic radius (ri) to electronegativity (Ei), ri / Ei, as a new characteristic, this combined feature couples size and electronic effects together, and can represent a small-radius, high-charge cation (such as Be). 2+Al3+ possesses a very strong polarization ability, which strongly attracts and distorts the electron clouds of surrounding anions. Using the ratio Z / ri (atomic coefficient (Z) to ionic radius (ri) as a new characteristic, this study can reflect the variation of ionic radii of ions in different charge states of the same element and provide information on the efficiency of ion filling in the crystal lattice.

[0170] In other words, the model in this embodiment does not need to learn the complex law that "the combination of small size and strong electron attraction will have a huge impact" from scratch. Instead, the inventors directly provide this physical insight to the model through the above-mentioned feature combination. By integrating these features, we effectively enhance the model's ability to distinguish different types of ions and predict certain properties, accelerate model convergence, and improve model accuracy and generalization ability.

[0171] Specifically, in this embodiment, the ratio of ionic radii, the ratio of ionic radius (ri) to electronegativity (Ei), and the ratio of atomic coefficient (Z) to ionic radius (ri) were selected as new combined feature parameters after conversion for subsequent model optimization. Among them, the ratio of ionic radii is the ratio of cation radius (r+) to anion radius (r-).

[0172] S5. Model Optimization

[0173] The feature parameters obtained from the feature engineering above, which are more correlated with the target variable, and the new combined feature parameters are used as input data to optimize the parameters of each model. Grid Search and Random Search are used for hyperparameter optimization.

[0174] Among them, grid search is used to systematically traverse the hyperparameter space to find the optimal combination of hyperparameters. Random search is used to randomly sample in the hyperparameter space to find the optimal combination of hyperparameters.

[0175] For example, for artificial neural networks, model performance can be optimized by adjusting hyperparameters such as the learning rate and the number of neurons.

[0176] For the model parameters of better artificial neural network algorithms (such as R...) 2 Optimize parameters (greater than 0.8) such as solver, activation function, number of neural layers, number of nodes, error function, learning rate, and momentum. Parameter optimization can be achieved using methods like grid search and random search to improve the model's predictive performance.

[0177] In this process, the feature parameters input into the algorithm come from the feature engineering in the first step, and the output feature parameters are evaluated and adjusted based on the model's prediction results. Validation criteria for the optimization results typically use metrics such as mean squared error (MSE) and coefficient of determination (R²) to ensure the model's generalization ability and accuracy.

[0178] Specifically, model parameters can be optimized in conventional ways. For example, optimization includes adjusting and selecting a more suitable solver, using non-linear activation functions (such as the Sigmoid activation function) to increase the model's expressive power, and changing the model's complexity and learning ability by adjusting the number of layers and nodes in each layer of the neural network. The error function is adjusted to mean squared error, and the learning rate is adjusted to control the model's convergence speed at a suitable level, balancing efficiency and the acquisition of the optimal solution.

[0179] Optimizing the above parameters can significantly improve the model's predictive performance. The validation criterion for the optimization results is typically cross-validation, which involves calculating MSE and R² on the validation set. 2 The generalization ability and prediction accuracy of the model are evaluated using metrics such as accuracy.

[0180] Accuracy is calculated as follows: points with an error within 20% are defined as acceptable "accurate points", and the number of accurate points divided by the total number of test points is the accuracy.

[0181] Ultimately, the optimal combination of model parameters that performs best on the validation set, i.e., the combination of parameters corresponding to the model with the highest accuracy, is selected to ensure the reliability and effectiveness of the model in practical applications.

[0182] S6. Model Iteration

[0183] Repeat steps S3-S5 above to continuously optimize the prediction results, and finally establish the optimal prediction model.

[0184] Meanwhile, during the continuous optimization process, outlier data values ​​in the dataset will be cleaned repeatedly based on the model's prediction results to continuously optimize data quality and improve the model's prediction accuracy and stability.

[0185] Finally, optimization is stopped when the accuracy rate is greater than 90%, thus obtaining the molten salt eutectic formulation design model.

[0186] Example 2

[0187] A machine learning-based method for designing molten salt eutectic formulations includes the following steps:

[0188] The characteristic parameters of each component salt in the multi-component molten salt system to be predicted are input into the molten salt eutectic formulation design model established in Example 1. The model is used to calculate and output the predicted minimum eutectic point ratio and minimum eutectic point melting temperature of the multi-component molten salt system.

[0189] Example 3

[0190] Prediction of the minimum eutectic point ratio and minimum eutectic point melting temperature of the binary molten salt system NaNO3-Ca(NO3)2.

[0191] This embodiment uses the model obtained by the model building method in Embodiment 1 for prediction. Specifically, the model training and prediction method used is a negative feedback artificial neural network algorithm. This algorithm is used to train and predict the atoms and ions (Na+, Na+, Na+, and Ca+) in the single-component salts NaNO3 and Ca(NO3)2 of the binary molten salt system NaNO3-Ca(NO3)2. + N 5+ O 2- Ca 2+ The characteristic parameters ra (atomic radius), E (atomic electronegativity), αa (atomic polarizability), M (atomic mass), Cn (coordination number), E (lattice energy), and T (melting point of single salt) of NaNO3, KNO3, and KCl are used as highly correlated characteristic parameters. The combined characteristic parameters z / r (ratio of charge to ionic radius), z+ / z- (ratio of anion to cation charge), r+ / r- (ratio of anion to cation radius), and 1 / z (reciprocal of charge) are used as input parameters and input into the model. The model is used to calculate the predicted output data, the minimum eutectic point ratio and the minimum eutectic point melting temperature.

[0192] The model predicts the lowest eutectic point ratio to be 0.53-0.47. DSC measurements show only one melting peak at this ratio. Figure 2 The figure shows the eutectic composition, indicating that the prediction is accurate.

[0193] The model predicted a minimum eutectic melting temperature of 211℃, while the measured melting point was 220℃, with an error of 4.4%, less than 5%. This demonstrates very high accuracy.

[0194] Example 4

[0195] Prediction of the minimum eutectic point ratio and minimum eutectic point melting temperature of the ternary molten salt system NaNO3-KNO3-KCl.

[0196] This embodiment uses the model obtained by the model building method in Embodiment 1 for prediction. Specifically, the model training and prediction method used is a negative feedback artificial neural network algorithm. This algorithm is used to train and predict the atoms and ions (Na+, K+, and K+) in the single-component salts NaNO3, KNO3, and KCl of the ternary molten salt system NaNO3-KNO3-KCl. + N 5+O 2- K + Cl - The characteristic parameters ra (atomic radius), E (atomic electronegativity), αa (atomic polarizability), M (atomic mass), Cn (coordination number), and E (lattice energy) are used as highly correlated characteristic parameters. The combined characteristic parameters z / r (ratio of charge number to ionic radius), z+ / z- (ratio of anion and cation charge number), r+ / r- (ratio of anion and cation radius), 1 / z (reciprocal of charge number), coordination number Cn, and lattice energy E are used as input parameters. These parameters are then input into the model, and the model is used to calculate the predicted output data, including the minimum eutectic point ratio and the minimum eutectic point melting temperature.

[0197] The model predicts the lowest eutectic point ratio to be 0.46-0.40-0.14. DSC measurements show only one melting peak at this ratio. Figure 3 The figure shows the eutectic composition, indicating that the prediction is accurate.

[0198] The model predicted a minimum eutectic melting temperature of 210℃, while the measured melting point was 211℃, with an error of 0.47%, which is less than 0.05%. This demonstrates very high accuracy.

[0199] Example 5

[0200] The model obtained according to the method in Example 1 above was used to predict the properties of binary and ternary molten salt combinations with different cations and anions. The results are shown in the table below.

[0201] Table 1. Comparison of Model Predictions and Measured Values

[0202]

[0203]

[0204] The above results show that the model can maintain the prediction error of formulation and properties of binary and ternary molten salt combinations of different cations and anions within 10%, indicating that the model has high generalization performance.

Claims

1. A method for establishing a model of a eutectic formulation design of a molten salt, characterized by, Comprising the following steps: S1. Establishing a database: Collecting characteristic parameters of each atom and ion in a single-component salt, the characteristic parameters including structural characteristic parameters and / or interaction characteristic parameters of atoms and ions; collecting molten salt combination compositions and / or phase transition temperature data of known multi-component molten salt systems composed of the single-component salt, to form a data set; S2. Data processing: Normalizing the characteristic parameters, the normalization being converting data into dimensionless characteristic parameter values in the range of 0-1; S3. Model training Dividing the data set after normalization into a training set, a validation set and a test set; Obtaining data of the training set, taking the characteristic parameters as input data, taking the molten salt combination compositions and / or phase transition temperature data as output data, establishing a model, training the model with several machine learning algorithms, and adjusting model parameters with an optimized loss function to obtain several trained models; Obtaining the data of the verification set and the data of the test set, performing hyperparameter optimization on the training model with the data of the verification set, and evaluating the model performance with the data of the test set to obtain a determination coefficient R 2 , screening the determination coefficient R 2 The model reaching the threshold value is a primary screening model; S4. Feature engineering Extracting high-correlation characteristic parameters from the preliminary screening model by principal component analysis, the high-correlation characteristic parameters being characteristic parameters with high correlation with the output data, and obtaining combination characteristic parameters reflecting the strength and type of ion interaction in a multi-component molten salt system through conversion of the characteristic parameters, the high-correlation characteristic parameters and the combination characteristic parameters being input data for subsequent model optimization; S5. Model optimization Optimizing each model parameter in the preliminary screening model obtained through the above screening with the validation set, and evaluating the optimized model by substituting data of the test set, to screen a model with good prediction accuracy, i.e. the molten salt eutectic formula design model.

2. The method of establishing a fused salt eutectic formulation design model of claim 1, wherein, The S1 satisfies at least one of the following conditions: (1) The melting point of the single-component salt is ≤1050℃, the boiling point is ≥380℃, and the decomposition temperature is ≥380℃; (2) The single-component salt includes at least one of NaNO3, KNO3, Ca(NO3)2, Na2CO3, K2CO3, NaCl, KCl, Na2SO4 and K2SO4; (3) The structural characteristic parameters include at least one of atomic number Z, group F, period pe, block Block, proton number p, electron number s, neutron number N, atomic mass M, mass number A, outer electron number V, atomic radius ra, covalent radius rco, van der Waals radius rf, metal radius rm, ion radius ri, atomic volume Va and coordination number Cn; (4) The interaction characteristic parameters include at least one of atomic electronegativity E, atomic ionization potential ip, n1 / 3ws value, Miedema electronegativity EM, ionization energy ie, atomic affinity a, atomic polarizability αa, ionic polarizability αi, ionic electronegativity Ei, and lattice energy E; (6) The molten salt combination composition and / or phase transition temperature data include at least one of eutectic point ratio, eutectic point melting temperature, eutectic point density and eutectic point specific heat. (7) the eutectic molten salt formula is a binary molten salt system composed of NaNO3 and Ca(NO3)2, the composition and / or phase transition temperature data of the molten salt combination are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa, and atomic mass M, lattice energy E, coordination number Cn, and single-component salt melting point T; (8) the eutectic molten salt formula is a ternary molten salt system composed of NaNO3, KNO3 and KCl, the composition and / or phase transition temperature data of the molten salt combination are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa, atomic mass M, lattice energy E, and coordination number Cn.

3. The method of establishing a fused salt eutectic formulation design model of claim 1, wherein, The S2 step satisfies at least one of the following conditions: (1) In the S2 step, before data normalization processing, a data cleaning step is further performed, the data cleaning step includes a screening step and / or a selection step, the screening step is: according to the reasonable range and law of each output data and input data, correcting or rejecting the error data; the selection step is: when different parameter values of the same data are collected from different sources, the frequency of each parameter value is counted, and the parameter value with the highest frequency is taken as the data value of the data; Preferably, the screening step includes the following steps: analyzing and judging each data, when it meets the rejection standard, rejecting the data; when it meets the correction standard, correcting the data; The rejection standard is: when the data meets any of the following conditions, the data is rejected: Condition one: the data value violates the physical and / or chemical law; Condition two: the associated variables of the data are missing; Condition three: the data source material shows that the data is unreliable; Condition four: the data appears twice or more than twice; Condition five: extreme value in statistics or physics, and the data verified by manual is an error value; Condition six: the deviation of the sum of the components of the multi-component system from 100% is >1%; The correction standard and correction method are: when the data meets the following categories, the correction method is corrected according to the category: I, unit correction: when the data value unit is inconsistent with the standard unit of the data, the standard unit value is converted; II, system error correction: when the data source material shows that the data of this source has systematic deviation, the data of this source is corrected by deducting the deviation; III, component normalization correction: when the deviation of the sum of the components of the multi-component system from 100% is ≤1%, the components are normalized according to the proportion of each component, and the sum of the proportions of the corrected components is equal to 100%; (2) The normalization processing is converted according to the following formula: Wherein: x represents the parameter value of a characteristic parameter; x min represents the minimum value of the feature; x max represents the maximum value of the feature; x norm represents the normalized feature parameter value.

4. The method of establishing a fused salt eutectic formulation design model of claim 1, wherein, The S3 step satisfies at least one of the following conditions: (1) the machine learning algorithm includes: negative feedback artificial neural network algorithm, support vector machine algorithm, ridge regression algorithm, nearest neighbor analysis algorithm, decision tree algorithm or random deep forest algorithm; (2) the optimization loss function selection: at least one of the average absolute error, Huber loss and mean square error; (3) the determination coefficient R 2 reaching the threshold value means R 2 ≥ 0.8; (4) the eutectic melt salt formula is a binary melt salt system composed of NaNO3 and Ca(NO3)2, the melt salt combination composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature, the characteristic parameters include atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M, and the machine learning algorithm is a negative feedback artificial neural network algorithm; (8) the eutectic melt salt formula is a ternary melt salt system composed of NaNO3, KNO3 and KCl, the melt salt combination composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature, the characteristic parameters include atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M, and the machine learning algorithm is a negative feedback artificial neural network algorithm.

5. The method of establishing a fused salt eutectic formulation design model of claim 1, wherein, The S4 step satisfies at least one of the following conditions: (1) the principal component analysis is calculated using the following formula: Xnew=X·V wherein: X represents an initial data matrix; V represents a matrix composed of principal components, each principal component being a linear combination of the initial data; Xnew represents a data matrix after dimension reduction; (2) the high correlation characteristic parameters meet the following conditions: I: eigenvalue > 1; II: the cumulative variance contribution rate of all high correlation characteristic parameters is ≥ 70% to 80%; III: the combined loading coefficient is > |0.7|; (3) the combination characteristic parameters include at least one of ion radius ratio, ion radius to electronegativity ratio, and atomic coefficient to ion radius ratio.

6. The method of establishing a fused salt eutectic formulation design model of claim 1, wherein, The S5 step satisfies at least one of the following conditions: (1) the optimization method is at least one of grid search and random search; preferably, it is grid search; (2) the grid search is to systematically traverse the hyperparameter space to screen the best hyperparameter combination; (3) the random search is to randomly sample the hyperparameter space to screen the best hyperparameter combination; (4) the model parameters include at least one of solver, activation function, number of neural layers, number of nodes, error function, learning rate, and momentum.

7. The method of establishing a model of a fused salt eutectic formulation design according to any one of claims 1 to 6, wherein Further comprising an S6 model iteration step, in which the S3-S5 steps are repeated to dynamically optimize the prediction effect and establish a prediction model with an accuracy rate greater than a threshold value; Preferably, the threshold value of the accuracy rate is 90%.

8. A model for designing eutectic formulations of molten salts, characterized in that, Established by the melt salt eutectic formula design model establishment method of any one of claims 1-7.

9. The fused salt eutectic formulation design model of claim 8, wherein, Satisfies at least one of the following conditions: (1) the single-component salt includes at least one of NaNO3, KNO3, Ca(NO3)2, Na2CO3, K2CO3, NaCl, KCl, Na2SO4 and K2SO4; (2) the eutectic melt salt formula is a binary melt salt system composed of NaNO3 and Ca(NO3)2, the melt salt combination composition and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include atomic radius (ra), atomic electronegativity (E), atomic polarizability (αa) and atomic mass (M). (3) the eutectic molten salt formula is a ternary molten salt system composed of NaNO3, KNO3 and KCl, the molten salt combination component and / or phase transition temperature data are the lowest eutectic point ratio and / or the lowest eutectic point melting temperature; the characteristic parameters include: atomic radius ra, atomic electronegativity E, atomic polarizability αa and atomic mass M (4) the machine learning algorithm is a negative feedback artificial neural network; (5) the determination coefficient R 2 ≥ 0.8; (6) the accuracy is ≥ 90%.

10. A machine learning based molten salt eutectic formulation design method, characterized by, The method comprises the following steps: inputting the characteristic parameters of each single component salt in the to-be-predicted multicomponent molten salt system into the molten salt eutectic formula design model of any one of claims 8-9, calculating with the model, and outputting the predicted molten salt combination component and / or phase transition temperature data of the multicomponent molten salt system.

Images