Electrolyte liquidus temperature prediction method, medium and electronic device

Abstract

The application discloses an electrolyte initial crystallization temperature prediction method, a medium and an electronic device, wherein the method comprises: obtaining to-be-tested electrolyte data, the to-be-tested electrolyte data comprising to-be-tested electrolyte components and mass percentage contents of the to-be-tested electrolyte components; obtaining molar fractions of elements based on the law of conservation of mass, the stoichiometry principle and the elements corresponding to the to-be-tested electrolyte components in the to-be-tested electrolyte data; obtaining point features of each element based on first attributes of the elements, and obtaining edge features between each two elements based on second attributes of the elements and the molar fractions; the first attributes are used for representing inherent physical and chemical attributes of the elements, and the second attributes are used for at least representing interaction strengths determined based on concentration coupling relationships between the elements; and the point features and the edge features are input into a pre-trained neural network prediction model to obtain a predicted electrolyte initial crystallization temperature; and the application improves the prediction accuracy and generalization ability of the initial crystallization temperature.

Description

Technical Field

[0001] This application belongs to the interdisciplinary field of aluminum electrolysis industry and artificial intelligence, and in particular relates to a method for predicting the primary crystallization temperature of an electrolyte, a medium, and an electronic device. Background Technology

[0002] In aluminum electrolysis, the primary crystallization temperature of the electrolyte is a core parameter determining the thermal balance control strategy of the electrolytic cell. Accurately controlling the primary crystallization temperature is crucial for maintaining electrolyte superheat within the ideal range, optimizing current efficiency, and extending the electrolytic cell's lifespan. Furthermore, establishing a precise mapping relationship between electrolyte composition and primary crystallization temperature allows for the scientific optimization of the electrolyte formulation.

[0003] Current methods for predicting primary crystallization temperature mainly rely on manual sampling thermal analysis or empirical regression formulas. However, these formulas have fixed forms and struggle to capture the complex nonlinear coupling effects between components. Furthermore, prediction accuracy drops significantly when the electrolyte composition exceeds the preset applicable range. While introduced machine learning methods improve data fitting capabilities, they typically treat the content of each additive as an independent numerical feature input, failing to consider the actual interaction mechanisms between ions within the molten salt system. This results in models lacking physical interpretability and exhibiting weak generalization ability.

[0004] This application addresses the practical needs of refined control and electrolyte formulation optimization in the aluminum electrolysis industry by providing a primary crystallization temperature prediction method that combines physical interpretability, stoichiometric rigor, and data-driven learning capabilities. This method aims to solve the problems of low prediction accuracy and poor generalization ability of existing technologies in complex composition environments. Summary of the Invention

[0005] The embodiments of this application provide a method, medium, and electronic device for predicting the primary crystallization temperature of an electrolyte, which can solve the problems of low prediction accuracy and poor generalization ability of the prior art in complex composition environments.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0007] According to one aspect of the embodiments of this application, a method for predicting the primary crystallization temperature of an electrolyte is provided, the method comprising: Acquire the data of the electrolyte to be tested, wherein the data of the electrolyte to be tested includes the components of the electrolyte to be tested and the mass percentage content of the components of the electrolyte to be tested; Based on the law of conservation of mass, the principle of stoichiometry, and the elements corresponding to the components of the electrolyte to be tested in the electrolyte data, the mole fraction of the element is obtained; wherein, the mole fraction is the ratio between the number of moles of a single element in the electrolyte data and the total number of moles of all elements. Point features of each element are obtained based on a first attribute of the element, and edge features between each pair of elements are obtained based on a second attribute of the element and the mole fraction; wherein, the first attribute is used to characterize the inherent physicochemical properties of the element, and the second attribute is used at least to characterize the interaction strength determined based on the concentration coupling relationship between the elements; The point features and edge features are input into a pre-trained neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte.

[0008] In some embodiments of this application, the first attribute of the element includes atomic mass, ionic radius, electronegativity, and ionic charge number; obtaining point features of each element based on the first attribute includes: Standardize each of the first attributes of the element to obtain the point features of each element; The second attribute of the element includes the ionic radius, the electronegativity, and the ionic charge number; the edge features include a concentration coupling term, a concentration gradient term, a bond length geometry term, a chemical bond property term, and a charge product term; obtaining the edge features between each pair of elements based on the second attribute of the element and the mole fraction includes: The concentration coupling term of the edge feature is obtained by multiplying the mole fractions of each pair of elements. The concentration gradient term of the edge feature is obtained based on the absolute value of the difference between the mole fractions of each pair of elements. The bond length geometry of the edge feature is obtained based on the sum of the ionic radii of each pair of elements. Based on the difference in electronegativity between each pair of elements, the chemical bond property term of the edge feature is obtained; The charge product term of the edge feature is obtained by multiplying the ionic charge of each pair of elements.

[0009] In some embodiments of this application, the step of inputting the point features and the edge features into a pre-trained neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte includes: The point features are mapped to initial latent space features. Based on the initial latent space features, the edge features, and the attention mechanism, the aggregation weights of the element and its neighboring elements are obtained to obtain the output latent space features of the multi-layer graph attention layer of the neural network prediction model. The aggregation weights are used to characterize the influence of different ion pairs on the properties of the electrolyte in the molten salt system corresponding to the electrolyte. The output latent space features are weighted and aggregated based on the mole fraction to obtain a graph embedding vector; wherein the graph embedding vector is used to characterize the thermodynamic state of the molten salt system corresponding to the electrolyte; The graph embedding vector is input into the multilayer perceptron of the neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte.

[0010] In some embodiments of this application, mapping the point features to initial latent space features, and obtaining the aggregation weights of the element and its neighboring elements based on the initial latent space features, the edge features, and the attention mechanism to obtain the output latent space features of the multi-layer graph attention layer of the neural network prediction model, includes: Based on the mapping weights and mapping bias vectors of the mapping layer of the neural network prediction model, the point features are mapped to initial latent space features in the mapping layer. Based on the attention parameter vector, the edge features, the initial latent space features of the element and its neighbors, the correlation coefficient between the element and its neighbors in the current layer is obtained; wherein, the current layer is the current layer of the multi-layer graph attention layer; Based on the normalized exponential function and the correlation coefficient, the aggregation weight of the element and its neighboring elements in the current layer is obtained; Based on the aggregation weights and the latent space features of the neighboring elements in the current layer, the latent space features of the element in the next layer are obtained, until the output latent space features of the element in the last layer are obtained; wherein, the next layer is the layer below the current layer of the multi-layer graph attention layer, and the last layer is the last layer of the multi-layer graph attention layer.

[0011] In some embodiments of this application, the step of weighting and aggregating the output latent space features based on the mole fraction to obtain the graph embedding vector includes: Based on the mole fraction and the output latent space features, the graph embedding vector of the electrolyte data to be tested in the global aggregation layer of the neural network is obtained; Alternatively, based on the mole fraction and the assigned weights, the gate weight vector of the electrolyte data to be tested in the global aggregation layer is obtained; based on the gate weight vector and the output latent space features, the graph embedding vector of the electrolyte data to be tested is obtained.

[0012] In some embodiments of this application, the step of inputting the graph embedding vector into the multilayer perceptron of the neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte includes: Based on the hidden weights and hidden bias vectors of the hidden layers of the neural network prediction model, the output weights and output bias vectors of the multilayer perceptron of the neural network prediction model, and the graph embedding vector, the predicted primary crystallization temperature of the electrolyte is obtained.

[0013] In some embodiments of this application, the pre-trained neural network prediction model is trained in the following manner: Obtain a sample dataset, wherein the sample dataset includes multiple initial samples and the actual initial crystallization temperature corresponding to each initial sample, and each initial sample includes multiple electrolyte components and the mass percentage content of each electrolyte component; Based on the law of conservation of mass, the principle of stoichiometry, and the sample elements corresponding to the various electrolyte components in each initial sample, the mole fraction of the sample elements is obtained; The point features of each sample element are obtained based on the first attribute of the sample element, and the edge features between each pair of sample elements are obtained based on the second attribute of the sample element and the mole fraction. The point features and edge features corresponding to the sample elements are input into a pre-constructed neural network model to obtain the predicted initial crystallization temperature of the initial sample. Based on the actual primary crystallization temperature, the predicted primary crystallization temperature, and the pre-constructed loss function, the neural network model is trained to obtain a trained neural network prediction model.

[0014] In some embodiments of this application, training the neural network model based on the actual primary crystallization temperature, the predicted primary crystallization temperature, and a pre-constructed loss function to obtain a trained neural network prediction model includes: Substituting the predicted primary crystallization temperature and the actual primary crystallization temperature into the pre-constructed loss function yields the loss value; Based on the loss value and / or the pre-set training epochs, the loss function is backpropagated to complete the training of the neural network model and obtain the neural network prediction model.

[0015] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein at least one piece of program code is stored therein, the at least one piece of program code being loaded and executed by a processor to implement the method described above.

[0016] According to one aspect of the embodiments of this application, an electronic device is provided, the electronic device including one or more processors and one or more memories, the one or more memories storing at least one piece of program code, the at least one piece of program code being loaded and executed by the one or more processors to implement the method as described above.

[0017] Based on the above solution, the technical solution provided in this application has at least the following advantages and advancements: This application first obtains the components and their mass percentages of the electrolyte to be tested, and then converts them into mole fractions of elements (physical constraint weights) based on the law of conservation of mass and the principles of stoichiometry. This eliminates the dependence of traditional methods on the applicability of specific compound formulations or empirical formulas, and can more fundamentally reflect the chemical composition of the molten salt system. Furthermore, the inherent physicochemical properties of each element (first attribute) are constructed as point features, while the mole fraction of elements combined with their concentration coupling relationship (second attribute) is used to construct edge features between elements. This graph structure modeling, which simultaneously includes the element's own properties and the interactions between elements, allows the model to capture the complex nonlinear coupling effects between different ions in the electrolyte due to concentration changes, resulting in stronger physical interpretability. By inputting the above point features and edge features into a pre-trained neural network prediction model, the prediction of the initial crystallization temperature is achieved, thereby effectively improving the prediction accuracy and generalization ability in complex multi-element molten salt systems.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 Flowchart of the electrolyte primary crystallization temperature prediction method provided in this application Figure 1 ; Figure 2 Flowchart of the electrolyte primary crystallization temperature prediction method provided in this application Figure 2 ; Figure 3 Flowchart of the electrolyte primary crystallization temperature prediction method provided in this application Figure 3 ; Figure 4 A schematic diagram of the electrolyte primary crystallization temperature prediction device provided in this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0022] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0023] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0024] It should also be noted that the terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such uses of the terms can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described.

[0025] In existing aluminum electrolysis industrial practices, the traditional method of obtaining primary crystallization temperature mainly relies on manual sampling thermal analysis or empirical regression formulas. For example, an offline primary crystallization temperature prediction model based on an improved Rerstreken formula is used to estimate the temperature using the mass percentage of electrolyte components. However, these empirical formulas are usually fitted based on experimental data within a specific component range. Their fixed form makes it difficult to capture the complex nonlinear coupling effects between components, and the prediction accuracy decreases significantly when the electrolyte composition exceeds the formula's preset applicable range.

[0026] With the development of artificial intelligence technology, machine learning methods have been introduced into the field of primary crystallization temperature prediction, such as using Back Propagation Artificial Neural Network (BP-ANN) and Support Vector Machine (SVM) to build prediction models. Although these methods improve data fitting ability, they typically treat the cryolite ratio (CR) and the content of each additive as independent numerical feature inputs, making it difficult to consider the actual interaction mechanisms between ions within the molten salt system, such as chemical bonding, electrostatic interactions, and coordination structures. This results in a lack of physical interpretability in the models and weak generalization ability beyond the sample coverage.

[0027] Therefore, in order to meet the practical needs of refined control and electrolyte formulation optimization in the aluminum electrolysis industry, there is an urgent need to develop a new primary crystallization temperature prediction method that combines physical interpretability, stoichiometric rigor, and data-driven learning capabilities, so as to solve the problems of low prediction accuracy and poor generalization ability of existing technologies in complex composition environments.

[0028] Figure 1 Flowchart of the electrolyte primary crystallization temperature prediction method provided in this application Figure 1 ,like Figure 1 As shown, the method includes: S101. Obtain the data of the electrolyte to be tested, wherein the data of the electrolyte to be tested includes the electrolyte components to be tested and the mass percentage content of the electrolyte components to be tested.

[0029] The electrolyte data to be tested can be obtained by acquiring an electrolyte sample from an online sampling port at the aluminum electrolysis production site or from an electrolyte sample prepared in the laboratory, performing quantitative analysis of the chemical composition of the sample, and measuring the electrolyte data of the sample. This electrolyte data can correspond to the data of one sample.

[0030] The mass percentage content of the electrolyte component to be tested is the ratio between the mass of a single electrolyte component to be tested in each sample and the mass of all electrolyte components to be tested. For example, if the electrolyte to be tested includes 7 types of compounds, there are 7 electrolyte components to be tested. The mass of the 7 electrolyte components to be tested is the total mass of the electrolyte components to be tested. The mass percentage content of each electrolyte component to be tested is the ratio of its own mass to the total mass.

[0031] In this embodiment, the electrolyte components to be tested include various fluoride salts such as sodium fluoride (NaF), aluminum fluoride (AlF3), lithium fluoride (LiF), magnesium fluoride (MgF2), potassium fluoride (KF), calcium fluoride (CaF2), and aluminum oxide (Al2O3). If the total mass of the above compounds is 100 grams and the mass of NaF is 45 grams, then the mass percentage content of NaF is 45%.

[0032] S102. Based on the law of conservation of mass, the principle of stoichiometry, and the elements corresponding to the components of the electrolyte to be tested in the electrolyte data, obtain the mole fraction of the element; wherein, the mole fraction is the ratio between the number of moles of a single element in the electrolyte data and the total number of moles of all elements.

[0033] Based on the law of conservation of mass and the principles of chemometrics, this application constructs a mapping from the mass content of compounds to the mole fraction of elements, unifying all components into a vector of mole fractions of the various elements contained in the system.

[0034] The law of conservation of mass states that in a closed system, the total mass of all substances participating in a chemical reaction is equal to the total mass of the products. This means that the types and numbers of atoms remain unchanged before and after a chemical reaction; mass cannot be created or destroyed. Therefore, when the mass percentage of each compound is known, the total number of moles of each element can be calculated. The principle of stoichiometry refers to the ability to determine the fixed atomic ratios among the constituent elements based on the chemical formula of a compound. For example, in AlF3, the molar ratio of fluoride (Al) to aluminum (F) is 1:3.

[0035] In the above examples of electrolyte components to be tested, the elements corresponding to the electrolyte components to be tested include aluminum (Al), oxygen (O), fluorine (F), lithium (Li), magnesium (Mg), potassium (K), calcium (Ca) and sodium (Na).

[0036] For each compound: obtain the product of the atomic mass and the number of atoms of each constituent element, and sum the corresponding products of each constituent element to obtain the molar mass of the compound; divide the mass of the compound by its molar mass to obtain the number of moles of the compound; multiply the number of atoms of each element in the compound by the number of moles of the compound to obtain the molar contribution of each element to the compound. It is evident that the calculation does not directly use the percentage content, but rather the mass itself. However, this calculation is actually constrained by the total mass; therefore, the relative proportions of each compound are crucial for the calculation and serve as a normalized representation, unaffected by the actual mass of the sample. After obtaining the percentage content, an arbitrary baseline total mass can be assumed for subsequent calculations; for simplicity and intuitiveness, a total mass of 100 grams can be chosen.

[0037] For Al, its mole count is the sum of the number of Al moles obtained from AlF3 and Al2O3. For K, its mole count comes from the number of K moles obtained from KF. After obtaining the mole counts for all 8 elements, the mole count of a single element is divided by the total number of moles of all elements to obtain the mole fraction of a single element.

[0038] This application models the molten salt system as a graph structure. ,in For a collection of element nodes, This is a set of fully connected edges. A node feature matrix is ​​constructed, injecting the inherent physicochemical properties of the elements; an edge feature tensor is constructed, defining the interaction strength based on the concentration coupling relationship between elements, and a fully connected graph of elements is constructed.

[0039] The set of element nodes is That is, corresponding to the above 8 elements; obtain the first... mole fraction of elements The formula is as follows:

[0040] in, For the first The number of moles of each element ; For the first The number of moles of each element; The sum of the mole numbers of all elements; the vector of mole fractions of the elements can be represented as: , This is a transpose.

[0041] S103. Obtain point features of each element based on the first attribute of the element, and obtain edge features between each pair of elements based on the second attribute of the element and the mole fraction; wherein, the first attribute is used to characterize the inherent physicochemical properties of the element, and the second attribute is used at least to characterize the interaction strength determined based on the concentration coupling relationship between the elements.

[0042] For example, the first attribute of the element includes atomic mass, ionic radius, electronegativity, and ionic charge number; obtaining the point features of each element based on the first attribute of the element includes: standardizing each of the first attributes of the element to obtain the point features of each element; The second attribute of the element includes the ionic radius, the electronegativity, and the ionic charge number; the edge features include a concentration coupling term, a concentration gradient term, a bond length geometry term, a chemical bond property term, and a charge product term; obtaining the edge features between each pair of elements based on the second attribute of the element and the mole fraction includes: The concentration coupling term of the edge feature is obtained by multiplying the mole fractions of each pair of elements; the concentration gradient term of the edge feature is obtained by the absolute value of the difference between the mole fractions of each pair of elements; the bond length geometry term of the edge feature is obtained by summing the ionic radii of each pair of elements; the chemical bond property term of the edge feature is obtained by the difference in electronegativity of each pair of elements; and the charge product term of the edge feature is obtained by multiplying the ionic charges of each pair of elements.

[0043] Inherent physicochemical properties such as atomic mass, ionic radius, electronegativity, and ionic charge number. Concentration coupling relationships such as concentration coupling terms and concentration gradient terms.

[0044] The initial point feature matrix can be defined as: ;in, Represent a lines and A real matrix of columns; The number of nodes, where each node represents an element, can be 8; The number of the first attribute can be 4; for the second attribute... The first element (i.e., the first) (a number of nodes), whose initial feature vectors are... It can be represented as:

[0045] in, Standardize for zero mean (Z-Score); For the first atomic mass of an element; For the first The ionic radius of the element; For the first Electronegativity of elements; For the first The ionic charge number of a certain element.

[0046] Edge features are used to represent nodes and nodes Interactions between them, edge features It can be represented as:

[0047] in, This is a concentration coupling term; For concentration gradient terms; For bond length geometry terms; This refers to the chemical bond properties. This is a charge product term, which is signed and used to describe the strength of the electrostatic interaction. To take the absolute value; This is a multiplication operation. The edge features are five-dimensional decoupling features, designed to describe the interactions between ions from different physical dimensions.

[0048] S104. Input the point features and the edge features into a pre-trained neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte.

[0049] The pre-trained neural network prediction model in this application is obtained by training a pre-built neural network model. This application names the neural network model Physically Constrained Elemental Graph Interaction Network (PCEI-Net). The neural network model includes a feature mapping layer, a multi-layer graph attention layer, a global aggregation layer, and a multi-layer perceptron.

[0050] During training, the neural network model is trained using the point and edge features of the training samples as input and the actual initial crystallization temperature as the label, until the model training is complete, resulting in a trained neural network prediction model. Therefore, when using the model, the point and edge features obtained from the test sample can be input into the neural network prediction model to obtain the model's predicted electrolyte initial crystallization temperature for the test sample.

[0051] In the above embodiments, the components and their mass percentages of the electrolyte to be tested are first obtained, and then converted into mole fractions of elements based on the law of conservation of mass and the principle of stoichiometry. This eliminates the dependence of traditional methods on the applicability of specific compound formulations or empirical formulas, and can more fundamentally reflect the chemical composition of the molten salt system. Furthermore, the inherent physicochemical properties (first properties) of each element are constructed as point features, while the mole fraction of the elements combined with their concentration coupling relationship (second properties) are used to construct edge features between elements. This graph structure modeling, which simultaneously includes the element's own properties and the interactions between elements, allows the model to capture the complex nonlinear coupling effects between different ions in the electrolyte caused by concentration changes, resulting in stronger physical interpretability. By inputting the above point features and edge features into a pre-trained neural network prediction model, the prediction of the initial crystallization temperature is achieved, thereby effectively improving the prediction accuracy and generalization ability in complex multi-element molten salt systems.

[0052] Figure 2 Flowchart of the electrolyte primary crystallization temperature prediction method provided in this application Figure 2 ,like Figure 2 As shown, the method includes: S201. Map the point features to initial latent space features, and obtain the aggregation weights of the element and its neighboring elements based on the initial latent space features, the edge features, and the attention mechanism to obtain the output latent space features of the multi-layer graph attention layer of the neural network prediction model; wherein, the aggregation weights are used to characterize the influence of different ion pairs on the properties of the electrolyte in the molten salt system corresponding to the electrolyte.

[0053] For example, based on the mapping weights and mapping bias vectors of the mapping layer of the neural network prediction model, the point features are mapped to initial latent space features in the mapping layer; Based on the attention parameter vector, the edge features, the initial latent space features of the element and its neighbors, the correlation coefficient between the element and its neighbors in the current layer is obtained; wherein, the current layer is the current layer of the multi-layer graph attention layer; Based on the normalized exponential function and the correlation coefficient, the aggregation weight of the element and its neighboring elements in the current layer is obtained; Based on the aggregation weights and the latent space features of the neighboring elements in the current layer, the latent space features of the element in the next layer are obtained, until the output latent space features of the element in the last layer are obtained; wherein, the next layer is the layer below the current layer of the multi-layer graph attention layer, and the last layer is the last layer of the multi-layer graph attention layer.

[0054] In the feature mapping layer, the initial point feature matrix is ​​input. The physical properties are mapped to the first property learned by the graph network through a fully connected layer. Initial latent space features of the species That is, mapping point features to initial latent space features:

[0055] in, For the first The initial feature vector of the point features of each element; These are the mapping weights of the mapping layer; The mapping bias vector of the mapping layer; In the graph attention layer, elements are retrieved. With neighboring elements In the current layer Correlation coefficient (attention coefficient) ):

[0056] in, For the first The attention parameter vector of the layer; It is an activation function, officially called the leaky modified linear unit; and For the first The learnable weight matrix of the layer; is; For the first The first layer Point features of a type of element; For the first The first layer Point features of a type of element; This is a vector concatenation operation.

[0057] The attention coefficients are normalized using the Softmax function (an activation function and a normalized exponential function) to obtain the aggregate weights of the element and its neighbors in the current layer. :

[0058] in, natural constant The exponent with the base; For the first The set of neighboring elements of a given element; For the first The neighboring elements of this element; For the first Attention coefficients of all neighboring elements of a given element The sum of exponential functions; No. The element is in the next layer (the first layer) Latent space features of layers The update formula is:

[0059] in, For the activation function, either the ELU function (exponential linear activation function) or the ReLU function (corrected linear unit) can be used. For the first The learnable weight matrix of a layer can be compared with... Parameter sharing depends on the number of parameters required by the model.

[0060] To improve expressive power, a multi-head attention mechanism can also be used, which concatenates the outputs of K attention heads:

[0061] in, This indicates that the outputs of the first attention head to the Kth attention head are concatenated into a vector. (The final sentence, "Get the Kth attention head's output," appears to be an error and doesn't need a direct translation.) The latent space features of each element in the next layer are obtained until the output latent space features of the element in the last layer are obtained.

[0062] For each node pair, this application concatenates the features of the central node, the features of the neighboring nodes, and the corresponding five-dimensional decoupled edge features into vectors; it calculates the non-standardized correlation coefficient between nodes using a learnable weight matrix and a non-linear activation function, and generates the final attention weights through a normalization function; it adopts a multi-head attention mechanism to aggregate the outputs of multiple attention heads to capture multi-dimensional ion interaction information within the molten salt system.

[0063] S202. The output latent space features are weighted and aggregated based on the mole fraction to obtain a graph embedding vector; wherein the graph embedding vector is used to characterize the thermodynamic state of the molten salt system corresponding to the electrolyte.

[0064] For example, based on the mole fraction and the output latent space features, the graph embedding vector of the electrolyte data to be tested in the global aggregation layer of the neural network is obtained; Alternatively, based on the mole fraction and the assigned weights, the gate weight vector of the electrolyte data to be tested in the global aggregation layer is obtained; based on the gate weight vector and the output latent space features, the graph embedding vector of the electrolyte data to be tested is obtained.

[0065] In the global aggregation layer, graph embedding vectors The following is obtained using the adaptive concentration-weighted summation operator proposed in this application:

[0066] in, For elements in the last layer The output latent space features; For gated functions; For Hadamah accumulation; To distribute the weights evenly, where V The value can be the number of elements, such as 8; These are learnable scalar parameters, constrained to the range [0,1] by the Sigmoid function (an activation function), used to dynamically adjust the contributions of the physical constraint term and the data averaging term. The graph embedding vector can also be obtained using a simplified concentration hard constraint: .

[0067] This application introduces a learnable scalar parameter for dynamically adjusting between physical constraint weights based on element mole fraction and uniformly distributed weights based on the number of nodes. Through adaptive learning of the parameter, the model can strictly follow stoichiometric logic while correcting non-ideal behaviors caused by polarization and other effects in the molten salt system, thereby improving the model's accuracy under complex conditions.

[0068] S203. Input the graph embedding vector into the multilayer perceptron of the neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte.

[0069] For example, the predicted primary crystallization temperature of the electrolyte is obtained based on the hidden weights and hidden bias vectors of the hidden layers of the neural network prediction model, the output weights and output bias vectors of the multilayer perceptron of the neural network prediction model, and the graph embedding vector.

[0070] The predicted primary crystallization temperature of the electrolyte is obtained through a multilayer perceptron. The formula is as follows:

[0071] in, These are the output weights of the multilayer perceptron; The hidden weights of the hidden layer; This refers to the hidden bias vector of the hidden layer; This is the output bias vector of the multilayer perceptron.

[0072] This application embeds a graph into a vector input to a multilayer perceptron. Through layer-by-layer mapping of deep hidden layers, it mines the complex nonlinear coupling information between ions contained in the vector. Utilizing the nonlinear fitting capability of the perceptron, the high-dimensional graph features are mapped into a single physical scalar, and finally, the predicted value of the primary crystallization temperature corresponding to the current electrolyte composition is calculated and output.

[0073] In the above embodiments, by mapping the point features of each element to initial latent space features, and using the attention mechanism combined with edge features in the graph attention layer to dynamically calculate the aggregation weights between the element and its neighboring elements, the neural network can adaptively allocate attention coefficients according to the actual physicochemical interactions (such as charge multiplication, concentration coupling, etc.) of different ion pairs in the molten salt system. This design simulates the differences in the contribution of different ions to the electrolyte properties, and through the iterative update of multiple graph attention layers, the latent space features of each element are fully integrated with the structured information of neighboring elements, thereby breaking through the limitation of traditional models that simply treat each component as an independent numerical feature, and enhancing the modeling ability for complex nonlinear coupling interactions between ions. Next, a weighted aggregation operation based on mole fraction (introducing learnable scalar parameters) is adopted to combine physical concentration constraints with data-driven adaptive adjustment, which ensures that the graph embedding vector strictly follows the principles of stoichiometry, and allows the model to flexibly correct non-ideal behaviors (such as polarization effects) in the molten salt system, generating a graph embedding vector that more accurately characterizes the thermodynamic state of the electrolyte molten salt system. Finally, the graph is embedded into the vector input of the multilayer perceptron to predict the primary crystallization temperature. The entire process, from feature construction and graph message passing to global readout, is highly isomorphic to the physical composition of the electrolyte system, thereby significantly improving the accuracy, physical interpretability, and generalization ability for extreme components or small sample scenarios in complex molten salt environments with multiple elements.

[0074] Figure 3 Flowchart of the electrolyte primary crystallization temperature prediction method provided in this application Figure 3 ,like Figure 3 As shown, the method includes: S301. Obtain a sample dataset, wherein the sample dataset includes multiple initial samples and the actual initial crystallization temperature corresponding to each initial sample, and each initial sample includes multiple electrolyte components and the mass percentage content of each electrolyte component.

[0075] S302. Based on the law of conservation of mass, the principle of chemometrics, and the sample elements corresponding to the various electrolyte components in each initial sample, obtain the mole fraction of the sample elements.

[0076] S303. Obtain the point features of each sample element based on the first attribute of the sample element, and obtain the edge features between each pair of sample elements based on the second attribute of the sample element and the mole fraction.

[0077] S304. Input the point features and edge features corresponding to the sample elements into a pre-constructed neural network model to obtain the predicted primary crystal temperature of the initial sample.

[0078] S305. Based on the actual primary crystallization temperature, the predicted primary crystallization temperature, and the pre-constructed loss function, the neural network model is trained to obtain a trained neural network prediction model.

[0079] For example, training the neural network model based on the actual primary crystallization temperature, the predicted primary crystallization temperature, and a pre-constructed loss function to obtain a trained neural network prediction model includes: Substitute the predicted primary crystal temperature and the actual primary crystal temperature into a pre-constructed loss function to obtain a loss value; based on the loss value and / or a pre-set number of training rounds, backpropagate the loss function to complete the training of the neural network model and obtain a neural network prediction model.

[0080] In the training process of this application, the sample data acquired can be historical production or experimental data. The acquired sample dataset includes multiple electrolyte components, the mass percentage content of each electrolyte component, and the true initial crystallization temperature. The true initial crystallization temperature can be used as the label for the training samples. The multiple electrolyte components and the mass percentage content of each electrolyte component are not directly input into the neural network model to be trained. Instead, they are first processed into point features and edge features, as described in the prediction model usage process above. The point features and edge features are then input into the neural network model to be trained. The specific processing method for the temperature prediction process of the neural network model is also as described in the prediction model usage process above, and will not be repeated here. After the initial samples of the sample dataset are processed into point features and edge features, they can be used to construct the training dataset. The training dataset includes multiple training samples and the corresponding true initial crystallization temperature for each training sample. Each training sample includes point features and edge features.

[0081] During training, the neural network model is based on a pre-built loss function. Perform backpropagation to train the neural network model:

[0082] in, The total number of training samples; No. The true initial crystallization temperature of each training sample; No. Predicted primary crystallization temperature for each training sample; The regularization coefficient is used. For all trainable parameters of the model; This is the square of the L2 norm. If the loss value obtained from the loss function gradually converges to a minimum, or the model reaches the preset number of training rounds, a neural network prediction model is obtained, which is used to predict the initial crystallization temperature of the electrolyte.

[0083] This application uses the mean squared error between the predicted value and the true label value as the loss function, introduces a regularization term into the loss function, and continuously updates the network parameters through the gradient descent algorithm until the prediction error of the initial crystal temperature meets the requirements.

[0084] In summary, this application achieves a characteristic transformation from additive ratio to intrinsic elemental distribution by converting the mass content of each compound component into elemental mole fraction. This can more directly reflect the physicochemical properties of electrolytes and significantly improve the model's generalization ability and physical interpretability under complex compositions.

[0085] An adaptive physics-weighted mechanism is introduced. This mechanism achieves an adaptive trade-off between physical constraint terms and data averaging terms through learnable scalar parameters. This design enables the model to strictly adhere to stoichiometry principles while correcting for non-ideal behaviors in molten salt systems caused by polarization and other effects, significantly enhancing the model's generalization ability and robustness under small to medium sample data and extreme compositions.

[0086] A highly physically constrained five-dimensional decoupled edge feature was constructed. This application decomposes the interionic interaction potential energy into five independent edge features (including concentration, bond length, bond strength, and charge factor), and specifically introduces a signed charge product term as the core feature. This decoupling design enables the graph neural network to nonlinearly learn the composite relationship between the interactions of different ion pairs and the primary crystallization temperature through its attention mechanism, avoiding the physical distortion caused by neglecting the charge sign in traditional models. The model structure is highly isomorphic to the physical composition of the molten salt system.

[0087] This application can provide data support for the refined control of thermal balance in aluminum electrolytic cells, which helps to maintain ideal superheat, optimize current efficiency, and extend cell life. At the same time, by revealing the nonlinear mapping relationship between electrolyte composition and primary crystallization temperature, this method can be further extended to the precise optimization of electrolyte formulation, providing an important technical solution for the aluminum industry to achieve energy conservation, emission reduction, and digital transformation.

[0088] The following describes an apparatus embodiment of this application, which can be used to perform the methods described in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the methods described in the above embodiments of this application.

[0089] Figure 4 This is a schematic diagram of the electrolyte primary crystallization temperature prediction device provided in this application. The device 400 may include: an acquisition module 401, a first processing module 402, a second processing module 403, and a prediction module 404.

[0090] The acquisition module 401 is used to acquire the data of the electrolyte to be tested, wherein the data of the electrolyte to be tested includes the components of the electrolyte to be tested and the mass percentage content of the components of the electrolyte to be tested; The first processing module 402 is used to obtain the mole fraction of the element based on the law of conservation of mass, the principle of stoichiometry, and the element corresponding to the electrolyte component in the electrolyte data to be tested; wherein the mole fraction is the ratio between the number of moles of a single element in the electrolyte data to the total number of moles of all elements. The second processing module 403 is used to obtain point features of each element based on the first attribute of the element, and to obtain edge features between each pair of elements based on the second attribute of the element and the mole fraction; wherein, the first attribute is used to characterize the inherent physicochemical properties of the element, and the second attribute is used to characterize at least the interaction strength determined based on the concentration coupling relationship between the elements. The prediction module 404 is used to input the point features and the edge features into a pre-trained neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte.

[0091] It should be noted that for any content not described in detail in this embodiment of the device, please refer to the method described above.

[0092] Based on the same concept, embodiments of this application provide an electronic device. Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application. Figure 5 As shown, the electronic device includes a memory 504, a processor 502, and a computer program stored in the memory 504 and executable on the processor 502. When the processor 502 executes the computer program, it implements the method provided in any embodiment of this application.

[0093] Among them, Figure 5 In this document, a bus architecture (represented by bus 500) is used. Bus 500 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 502 and memory represented by memory 504. Bus 500 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 505 provides an interface between bus 500 and receiver 501 and transmitter 503. Receiver 501 and transmitter 503 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 502 is responsible for managing bus 500 and general processing, while memory 504 can be used to store data used by processor 502 during operation.

[0094] Based on the same concept, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method provided in any embodiment of this application.

[0095] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored as one or more instructions or codes on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope and spirit of this application and the appended claims. For example, due to the nature of software, the functions described above may be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit.

[0096] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0097] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0098] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0099] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0100] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0101] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for predicting the primary crystallization temperature of an electrolyte, characterized in that, The method includes: Acquire the data of the electrolyte to be tested, wherein the data of the electrolyte to be tested includes the components of the electrolyte to be tested and the mass percentage content of the components of the electrolyte to be tested; Based on the law of conservation of mass, the principle of stoichiometry, and the elements corresponding to the components of the electrolyte to be tested in the electrolyte data, the mole fraction of the element is obtained; wherein, the mole fraction is the ratio between the number of moles of a single element in the electrolyte data and the total number of moles of all elements. Point features of each element are obtained based on a first attribute of the element, and edge features between each pair of elements are obtained based on a second attribute of the element and the mole fraction; wherein, the first attribute is used to characterize the inherent physicochemical properties of the element, and the second attribute is used at least to characterize the interaction strength determined based on the concentration coupling relationship between the elements; The point features and edge features are input into a pre-trained neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte.

2. The method according to claim 1, characterized in that, The first attribute of the element includes atomic mass, ionic radius, electronegativity, and ionic charge number; obtaining the point features of each element based on the first attribute includes: Standardize each of the first attributes of the element to obtain the point features of each element; The second attribute of the element includes the ionic radius, the electronegativity, and the ionic charge number; the edge features include a concentration coupling term, a concentration gradient term, a bond length geometry term, a chemical bond property term, and a charge product term; obtaining the edge features between each pair of elements based on the second attribute of the element and the mole fraction includes: The concentration coupling term of the edge feature is obtained by multiplying the mole fractions of each pair of elements. The concentration gradient term of the edge feature is obtained based on the absolute value of the difference between the mole fractions of each pair of elements. The bond length geometry of the edge feature is obtained based on the sum of the ionic radii of each pair of elements. Based on the difference in electronegativity between each pair of elements, the chemical bond property term of the edge feature is obtained; The charge product term of the edge feature is obtained by multiplying the ionic charge of each pair of elements.

3. The method according to claim 1, characterized in that, The step of inputting the point features and the edge features into a pre-trained neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte includes: The point features are mapped to initial latent space features. Based on the initial latent space features, the edge features, and the attention mechanism, the aggregation weights of the element and its neighboring elements are obtained to obtain the output latent space features of the multi-layer graph attention layer of the neural network prediction model. The aggregation weights are used to characterize the influence of different ion pairs on the properties of the electrolyte in the molten salt system corresponding to the electrolyte. The output latent space features are weighted and aggregated based on the mole fraction to obtain a graph embedding vector; wherein the graph embedding vector is used to characterize the thermodynamic state of the molten salt system corresponding to the electrolyte; The graph embedding vector is input into the multilayer perceptron of the neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte.

4. The method according to claim 3, characterized in that, The step of mapping the point features to initial latent space features, and obtaining the aggregation weights of the element and its neighboring elements based on the initial latent space features, the edge features, and the attention mechanism to obtain the output latent space features of the multi-layer graph attention layer of the neural network prediction model includes: Based on the mapping weights and mapping bias vectors of the mapping layer of the neural network prediction model, the point features are mapped to initial latent space features in the mapping layer. Based on the attention parameter vector, the edge features, the initial latent space features of the element and its neighbors, the correlation coefficient between the element and its neighbors in the current layer is obtained; wherein, the current layer is the current layer of the multi-layer graph attention layer; Based on the normalized exponential function and the correlation coefficient, the aggregation weight of the element and its neighboring elements in the current layer is obtained; Based on the aggregation weights and the latent space features of the neighboring elements in the current layer, the latent space features of the element in the next layer are obtained, until the output latent space features of the element in the last layer are obtained; wherein, the next layer is the layer below the current layer of the multi-layer graph attention layer, and the last layer is the last layer of the multi-layer graph attention layer.

5. The method according to claim 3, characterized in that, The weighted aggregation of the output latent space features based on the mole fraction to obtain the graph embedding vector includes: Based on the mole fraction and the output latent space features, the graph embedding vector of the electrolyte data to be tested in the global aggregation layer of the neural network is obtained; Alternatively, based on the mole fraction and the assigned weights, the gate weight vector of the electrolyte data to be tested in the global aggregation layer is obtained; based on the gate weight vector and the output latent space features, the graph embedding vector of the electrolyte data to be tested is obtained.

6. The method according to claim 3, characterized in that, The step of inputting the graph embedding vector into the multilayer perceptron of the neural network prediction model to obtain the predicted primary crystallization temperature of the electrolyte includes: Based on the hidden weights and hidden bias vectors of the hidden layers of the neural network prediction model, the output weights and output bias vectors of the multilayer perceptron of the neural network prediction model, and the graph embedding vector, the predicted primary crystallization temperature of the electrolyte is obtained.

7. The method according to claim 1, characterized in that, The pre-trained neural network prediction model is trained in the following way: Obtain a sample dataset, wherein the sample dataset includes multiple initial samples and the actual initial crystallization temperature corresponding to each initial sample, and each initial sample includes multiple electrolyte components and the mass percentage content of each electrolyte component; Based on the law of conservation of mass, the principle of stoichiometry, and the sample elements corresponding to the various electrolyte components in each initial sample, the mole fraction of the sample elements is obtained; The point features of each sample element are obtained based on the first attribute of the sample element, and the edge features between each pair of sample elements are obtained based on the second attribute of the sample element and the mole fraction. The point features and edge features corresponding to the sample elements are input into a pre-constructed neural network model to obtain the predicted initial crystallization temperature of the initial sample. Based on the actual primary crystallization temperature, the predicted primary crystallization temperature, and the pre-constructed loss function, the neural network model is trained to obtain a trained neural network prediction model.

8. The method according to claim 7, characterized in that, The process of training the neural network model based on the actual primary crystallization temperature, the predicted primary crystallization temperature, and a pre-constructed loss function to obtain a trained neural network prediction model includes: Substituting the predicted primary crystallization temperature and the actual primary crystallization temperature into the pre-constructed loss function yields the loss value; Based on the loss value and / or the pre-set training epochs, the loss function is backpropagated to complete the training of the neural network model and obtain the neural network prediction model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one piece of program code, which is loaded and executed by a processor to implement the method as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, The electronic device includes one or more processors and one or more memories, wherein at least one piece of program code is stored in the one or more memories, and the at least one piece of program code is loaded and executed by the one or more processors to implement the method as claimed in any one of claims 1 to 8.

Images