A method and system for predicting the physical properties of ionic liquids

Abstract

This invention belongs to the field of ionic liquid property prediction, and discloses a method and system for predicting ionic liquid properties. The method includes: obtaining a global structural feature vector of the ionic liquid that does not change with environmental factors from ionic liquid data; inputting the feature vector into a dual-branch neural network for decoupled calculation, with the main branch outputting the macroscopic property prediction value of the ionic liquid; outputting thermodynamic reference constants that constrain the properties to be predicted from the auxiliary branch; determining a joint loss function by dynamically weighting and fusing the residual loss term and the thermodynamic physical constraint loss term through empirical data fitting; training the dual-branch neural network based on the joint loss function; inputting the features of the ionic liquid to be tested and environmental factors into the trained dual-branch neural network, with the main branch outputting the final property prediction result of the ionic liquid that conforms to the thermodynamic evolution law. This invention solves the problems of insufficient accuracy of thermodynamic models and low efficiency of computational chemistry methods in existing ionic liquid property prediction technologies.

Description

Technical Field

[0001] This invention belongs to the field of ionic liquid property prediction, and specifically relates to a method and system for predicting the properties of ionic liquids. Background Technology

[0002] Currently, predicting key physicochemical properties of ionic liquids (such as gas absorption and viscosity) has become an important means to accelerate the large-scale industrial application of ionic liquids. Among existing traditional prediction methods, thermodynamic models (such as the equation of state (PSRK, Predictive Soave–Redlich–Kwong) and the conductor-like screening model (COSMO-RS)) often only achieve semi-quantitative accuracy due to excessive parameter simplification; while computational chemistry methods such as density functional theory (DFT) and molecular dynamics (MD) simulations are too computationally expensive to meet the needs of large-scale screening.

[0003] Therefore, Graph Neural Network (GNN) models, because they do not require the design of cumbersome molecular descriptors and can directly convert the anions and cations of ionic liquids into molecular graph structures for end-to-end feature learning and property prediction, are currently the most cutting-edge and mainstream technology in this field.

[0004] However, existing pure data-driven prediction frameworks, such as standard GNNs, lack underlying thermodynamic and physical constraints. Their architecture, which relies solely on mathematical fitting and is detached from physical laws, hinders their industrialization and in-depth application when predicting the complex macroscopic thermodynamic properties of ionic liquids. Specifically: 1. An irrational strong coupling exists between molecular microstructure features and macroscopic environmental features, leading to distorted feature extraction and low training efficiency: Existing GNN models typically employ a fusion strategy when processing cross-dimensional data. This involves directly concatenating the microscopic static structural features extracted from the molecular diagram with macroscopic dynamic environmental variables such as temperature and pressure before inputting the mixture into a fully connected network for fitting. This approach fails to adhere to the objective physical law that "some intrinsic physical constants of ionic liquids are independent of external environmental conditions," resulting in feature entanglement problems. Nonlinear fluctuations in environmental factors interfere with the feature learning of the global molecular structure during backpropagation, causing feature extraction distortion, significantly increasing the difficulty of model optimization, and reducing convergence speed and final prediction performance.

[0005] 2. Predictive results violate fundamental laws of physical consistency, leading to failure under extreme conditions: The physicochemical properties of ionic liquids are significantly affected by external environmental factors such as temperature and pressure, and follow specific thermodynamic laws. However, traditional GNN models do not incorporate physical constraints during training, making their predictions prone to exhibiting anomalous trends that violate common sense. For example, in regions lacking high-pressure data, the model may predict an erroneous trend of decreasing absorption with increasing pressure. This lack of physical consistency severely limits the model's extrapolation capabilities under extreme temperature and pressure conditions where experimental data is scarce.

[0006] 3. High dependence on high-quality experimental data, weak predictive ability with small sample data: The learning ability of GNNs relies heavily on large-scale, high-quality experimental data. However, ionic liquids are diverse, experimental determination is costly and time-consuming, and publicly available datasets can only cover a very small chemical space. For small sample scenarios such as high-viscosity systems and novel functionalized ionic liquids, pure data-driven GNNs are prone to overfitting, failing to learn the common thermodynamic evolution laws of ionic liquids, resulting in a significant decrease in prediction accuracy. This makes it difficult to meet the actual needs of industry for high-throughput virtual screening of unknown ionic liquids.

[0007] In summary, existing ionic liquid property prediction technologies suffer from insufficient accuracy in thermodynamic models and low efficiency in computational chemistry methods. While mainstream GNNs balance accuracy and efficiency, they are unable to achieve accurate, stable, and extrapolable predictions of ionic liquid physicochemical properties due to issues such as lack of physical constraints, strong coupling between microscopic and macroscopic features, lack of physical consistency in predictions, and poor generalization ability with small samples. Consequently, they fail to meet the needs of industrial applications. Summary of the Invention

[0008] To address the above problems, this invention provides a method for predicting the properties of ionic liquids, comprising the following steps: The ionic liquid structures in the ionic liquid dataset are transformed into molecular graphs. GNN is used to extract features and aggregate information from the molecular graphs to obtain global structural feature vectors of ionic liquids that do not change with environmental factors. The dual-branch neural network is constructed by including a main branch and an auxiliary branch. The global structural feature vector is input into the dual-branch neural network for decoupled calculation. The main branch concatenates and maps the ionic liquid structural feature vector with environmental dependent variables, and outputs the macroscopic property prediction value of the ionic liquid. The auxiliary branch extracts and outputs the thermodynamic reference constant constraining the property to be predicted based on the feature vector and relevant environmental dependent variables. The joint loss function is determined by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term by fitting empirical data. The thermodynamic and physical constraint loss term forces the macroscopic property prediction value of the main branch to be consistent with the thermodynamic reference constant of the auxiliary branch in accordance with the physical law. Backpropagation and parameter optimization are performed on the dual-branch neural network based on the joint loss function to complete the model training. The characteristics and environmental factors of the ionic liquid to be tested are input into the trained dual-branch neural network, and the main branch outputs the final property prediction results of the ionic liquid that conform to the thermodynamic evolution law.

[0009] Furthermore, the ionic liquid structures in the ionic liquid dataset are transformed into molecular graphs. GNNs are then used to extract features and aggregate information from these molecular graphs to obtain global structural feature vectors of the ionic liquids that are independent of environmental factors. This process includes the following steps: The molecular composition information of ionic liquids is obtained and a molecular graph is constructed. Atomic properties are extracted as initial node features and chemical bond properties are extracted as initial edge features. An adjacency matrix representing topological connection information is constructed based on the connection relationship between atoms. The initial node features, initial edge features and adjacency matrix are input into GNN. Based on the neighbor relationship defined by the adjacency matrix, GNN performs multiple rounds of local node feature interaction and update through information transmission mechanism to obtain deep node feature vectors. The feature vectors of all deep nodes are aggregated by global pooling through the readout function, mapped to a fixed-dimensional graph-level feature representation, and used as the global structural feature vector of the ionic liquid that does not change with environmental factors.

[0010] Furthermore, the auxiliary branch extracts and outputs the thermodynamic reference constants that constrain the properties to be predicted based on the eigenvectors and relevant environmental dependent variables, including the following steps: When predicting phase equilibrium properties, the auxiliary branch concatenates the global structural feature vector of the ionic liquid with the temperature environment variable to construct a joint input feature matrix. Based on the joint input feature matrix, the activation function, the weight matrix and bias vector that the auxiliary branch can learn autonomously, the enthalpy of dissolution is determined. Based on the van der Hoff equation and Henry's law, the enthalpy of dissolution is mapped to temperature to obtain the logarithmic value of the Henry's constant.

[0011] Furthermore, the auxiliary branch extracts and outputs the thermodynamic reference constants that constrain the properties to be predicted based on the feature vector and relevant environmental dependent variables, and also includes the following steps: When predicting rheological and transport properties, the auxiliary branch cuts off the input of environmental dependent variables in the feature extraction stage. The auxiliary branch is internally bifurcated into three parallel fully connected sub-layers. The three fully connected sub-layers generate viscosity limit parameters, apparent activation energy related parameters, and ideal glass transition temperature based on the global structural feature vector of the ionic liquid, respectively. Based on the VFT equation, external temperature, viscosity limit parameter, apparent activation energy related parameter and ideal glass transition temperature, which characterize the external environment, are mapped to the theoretical benchmark value of viscosity logarithm.

[0012] Furthermore, the main branch concatenates and maps the ionic liquid's structural feature vector with environmental dependent variables to output the predicted macroscopic properties of the ionic liquid, including the following steps: The global structural feature vector of the ionic liquid is directly concatenated with the normalized environmental factors to form a joint feature vector, which is then input into a multilayer perceptron. After nonlinear transformation, the macroscopic property prediction values ​​of the ionic liquid are directly output.

[0013] Furthermore, for the prediction of phase equilibrium physical properties, the thermodynamic physical constraint loss terms include Henry's law loss, physical monotonicity penalty loss of CO2 uptake with respect to pressure, and physical monotonicity penalty loss of CO2 uptake with respect to temperature.

[0014] Furthermore, for rheological and transport-related physical properties, the thermodynamic physical constraint loss terms include the VFT equation consistency constraint loss, the physical monotonicity penalty loss of viscosity with respect to pressure, and the physical monotonicity penalty loss of viscosity with respect to temperature.

[0015] Further, determining the Henry's Law loss includes the following steps: It is determined that the gas absorption at constant temperature is proportional to the equilibrium partial pressure. A penalty residual is calculated between the predicted absorption of the main branch and the logarithmic value of the Henry's constant of the auxiliary branch, based on the Henry's law loss calculation.

[0016] Furthermore, the consistency constraint loss of the VFT equations is determined, including: In viscosity prediction, the predicted logarithmic viscosity output by the forced main branch must converge with the viscosity limit parameter, apparent activation energy related parameter, ideal glass transition temperature, and theoretical reference value of viscosity logarithm generated by the auxiliary branch.

[0017] Furthermore, the physical monotonicity penalty loss of CO2 uptake with respect to pressure and the physical monotonicity penalty loss of CO2 uptake with respect to temperature are determined, including: By penalizing any non-physical parameter iteration path that violates the principle that CO2 absorption increases with increasing pressure or decreases with increasing temperature, we obtain the physical monotonicity penalty loss of CO2 absorption with respect to pressure and the physical monotonicity penalty loss of CO2 absorption with respect to temperature, respectively.

[0018] Furthermore, determining the physical monotonicity penalty loss of viscosity with respect to pressure and the physical monotonicity penalty loss of viscosity with respect to temperature includes the following steps: To ensure that the global solution for fluid viscosity prediction increases with increasing pressure and decreases with increasing temperature, physical monotonicity penalty losses for viscosity with respect to pressure and physical monotonicity penalty losses for viscosity with respect to temperature are constructed respectively.

[0019] Furthermore, the joint loss function is determined by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term through empirical data fitting, including the following steps: The joint loss function is obtained by summing the product of the empirical data fitting residual loss term, the total thermodynamic and physical constraint loss term, and the weighted coefficient hyperparameters.

[0020] Further, the loss term for the empirical data fitting residuals is determined, including the following steps: For phase equilibrium property prediction, the mean square error of the CO2 mole fraction directly predicted by the main branch and the experimentally measured baseline value is calculated to obtain the fitting residual loss term of the empirical data of CO2 absorption. For the prediction of rheological and transport properties, the logarithmic value of the ionic liquid viscosity predicted by the main control branch is extracted and supervisedly calibrated with the actual logarithmic value characterized by experiments to obtain the residual loss term of viscosity empirical data fitting.

[0021] Further, the total thermodynamic and physical constraint loss term is determined, including the following steps: For phase equilibrium property prediction, the Henry's law loss, the physical monotonicity penalty loss of CO2 uptake with respect to pressure, and the physical monotonicity penalty loss of CO2 uptake with respect to temperature are linearly superimposed to obtain the total thermodynamic physical constraint loss term for CO2 uptake prediction. For predicting rheological and transport-related physical properties, the consistency constraint loss of the VFT equation, the physical monotonicity penalty loss of viscosity with respect to pressure, and the physical monotonicity penalty loss of viscosity with respect to temperature are linearly spatially superimposed to obtain the total thermodynamic physical constraint loss term for viscosity prediction.

[0022] The present invention also provides an ionic liquid property prediction system, comprising: The data processing module is used to convert the ionic liquid structure in the ionic liquid dataset into a molecular graph, and to use GNN to extract features and aggregate information from the molecular graph to obtain the global structural feature vector of the ionic liquid that does not change with environmental factors. The model building module is used to construct a dual-branch neural network, including a main branch and an auxiliary branch. The dual-branch neural network is used to decouple the calculation of the input global structural feature vector. The main branch concatenates and maps the ionic liquid structural feature vector with environmental dependent variables to output the macroscopic property prediction value of the ionic liquid. The auxiliary branch extracts and outputs the thermodynamic reference constants that constrain the property to be predicted based on the feature vector and related environmental dependent variables. The loss function construction module is used to determine the joint loss function by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term through empirical data fitting. The thermodynamic and physical constraint loss term forces the macroscopic property prediction value of the main branch to be consistent with the thermodynamic reference constant of the auxiliary branch in accordance with the physical laws. The prediction module is used to perform backpropagation and parameter optimization on the dual-branch neural network based on the joint loss function to complete the model training. The characteristics and environmental factors of the ionic liquid to be tested are input into the trained dual-branch neural network, and the main branch outputs the final property prediction results of the ionic liquid that conform to the thermodynamic evolution law.

[0023] The beneficial effects of this invention are: 1. This invention deeply integrates a general thermodynamic equation describing the variation of specific thermodynamic properties with environmental factors (such as temperature and pressure) into a dual-branch neural network, constructs a prediction model with physical interpretability and extrapolation robustness, and finally controls the weight update of the network through a joint loss function containing physical law constraint penalty terms, thereby achieving property prediction with high physical consistency. It is not only applicable to the properties of carbon-captured ionic liquids, but also has general applicability and can be extended to the accurate prediction of a variety of key thermodynamic properties such as conductivity, surface tension, and density.

[0024] 2. The main branch of this invention combines global molecular features with environmental dependent variables to output macroscopic physical property predictions; the auxiliary branch acts as an automatic physical property extractor, autonomously learning from high-dimensional features and outputting the intrinsic thermodynamic constants that dominate the fluid properties as theoretical benchmarks for physical constraints.

[0025] 3. The dual-branch neural network of the present invention adds an empirical data fitting residual loss term to the joint loss function of the model training, which forces the macroscopic prediction value output by the main branch to be equal to or close to the theoretical benchmark value calculated by substituting the constant generated by the auxiliary branch into the classical thermodynamic equation (such as Henry's law, van der Hoff equation, and modified VFT rheological limit equation), thereby constructing a physical defense line.

[0026] 4. The dual-branch neural network of the present invention applies a forced loss penalty to the iterative direction that violates the first / second law of thermodynamic monotonicity (e.g., the phenomenon that the absorption amount decreases with increasing pressure, or the viscosity increases with increasing temperature, etc.) to ensure the physical reliability under the global extrapolation environment.

[0027] 5. The auxiliary branch of the dual-branch neural network of the present invention can adaptively route the input based on the physical differences of the properties to be solved: when solving the phase equilibrium state constant that depends on the evolution of the environment (such as the derivation of Henry's constant), the auxiliary branch input not only includes molecular features, but also additionally splices environmental dependent variables (temperature T); when solving the inherent limiting law benchmark constant of the substance itself (such as ideal glass transition temperature, viscosity limit parameter), the auxiliary branch input truncates the environmental dependent variables and forces it to be reconstructed only from global molecular features.

[0028] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart illustrating a method for predicting the physical properties of ionic liquids according to an embodiment of the present invention is shown. Figure 2 A schematic diagram of an ionic liquid property prediction system according to an embodiment of the present invention is shown. Figure 3 Histograms comparing the basic prediction performance of a dual-branch neural network according to an embodiment of the present invention with that of an existing pure data-driven GNN under different underlying architectures are shown. Figure 4 A scatter plot showing the fitting of the predicted values ​​and the actual experimental values ​​of the dual-branch neural network on an unknown test set of ionic liquids according to an embodiment of the present invention is presented. Figure 5 The diagram shows a comparison of the prediction trajectories of the dual-branch neural network and the ordinary GNN model under different extreme thermodynamic conditions for the same specific ionic liquid system. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It should 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 data can be interchanged where appropriate for the embodiments of this application described herein.

[0033] This invention provides a method and system for predicting the physical properties of ionic liquids. Addressing the inherent limitations of existing data-driven GNN models in predicting complex thermodynamic properties of ionic liquids—such as the uninterpretable nature of the model as a "pure mathematical black box," the tendency to violate fundamental physical laws under extreme conditions, and the poor generalization ability due to excessive reliance on massive amounts of data—the invention solves the following problems with existing GNN models: 1. The existing GNN model suffers from the problem of "strong coupling" between molecular structure features and external environmental features, leading to distorted feature representation and low learning efficiency.

[0034] 2. Existing GNN models have problems with prediction results that deviate from actual physical laws and lack thermodynamic consistency.

[0035] 3. Existing GNN models rely excessively on massive experimental data, resulting in low prediction accuracy on sparse, small sample datasets.

[0036] like Figure 1 As shown, a method for predicting the physical properties of ionic liquids includes the following steps: S1. Transform the ionic liquid structures in the ionic liquid dataset into molecular graphs. Use GNN (Graph Neural Network) to extract features and aggregate information from the molecular graphs to obtain global structural feature vectors of ionic liquids that do not change with environmental factors. .

[0037] The process of converting ionic liquids in the ionic liquid dataset into molecular graphs includes the following steps: Generating SMILES strings from the molecular composition information of all ionic liquids in the dataset; using the obtained SMILES strings as input, and converting the ionic liquids into molecular graphs using the RDKit package. Molecular diagram Based on the atomic feature matrix, chemical bond feature matrix, and adjacency matrix, atomic attributes are extracted as initial node features, and chemical bond attributes are extracted as initial edge features. An adjacency matrix representing topological connectivity information is then constructed based on the inter-atomic connection relationships, as detailed below:

[0038] in, The feature matrix of atoms (nodes) This represents the total number of atoms in the ionic liquid. Let be the dimension of the atomic feature space. Atoms and eigenvectors; Represents the characteristic matrix of chemical bonds (edges), where It is twice the total number of chemical bonds (because the molecular diagram is an undirected graph). As a dimension of chemical bond characteristics, Atoms and The eigenvectors of the chemical bonds between them (bidirectional). An edge index matrix, also known as an adjacency matrix, stored in coordinate (COO) format, encodes the connections between atoms, thereby capturing the topology of the molecular graph.

[0039] Specifically, GNNs are used to extract and aggregate features from the molecular graph topology to obtain global structural feature vectors of ionic liquids that are independent of environmental factors (temperature, pressure, etc.). The process includes the following steps: Using a Generative Neural Network (GNN) as the backbone encoder, initial node features, initial edge features, and an adjacency matrix are input into the GNN. Based on the neighbor relationships defined by the adjacency matrix, the GNN performs multiple rounds of local node feature interaction and updates through a message passing mechanism to obtain deep node feature vectors. Finally, all deep node feature vectors are aggregated using global pooling (GP) through a readout function, mapped to a fixed-dimensional graph-level feature representation, and used as the global structural feature vector of the ionic liquid that does not change with environmental factors. .

[0040] For example, GNNs include architectures such as MPNN (Message Passing Neural Network), GIN (Graph Isomorphism Network), PNA (Principal Neighbourhood Aggregation Network), or AttentiveFP (Attentive FingerPrint Network).

[0041] S2. Construct a dual-branch neural network including a main branch and an auxiliary branch. Input the global structural feature vector into the dual-branch neural network for decoupled calculation. The main branch concatenates and maps the ionic liquid structural feature vector with environmental dependent variables, and outputs the macroscopic property prediction value of the ionic liquid. The auxiliary branch extracts and outputs the thermodynamic reference constants that constrain the property to be predicted based on the feature vector and related environmental dependent variables.

[0042] Existing machine learning models directly input molecular structural features and environmental dependent variables (temperature, pressure) into an extremely complex fully connected network to fit target values, causing the model to become an uninterpretable "pure mathematical black box." In previously unseen high temperature or high pressure ranges, it is prone to making garbled predictions that violate physical common sense.

[0043] This invention proposes a dual-branch neural network with decoupled physical properties, which receives the global structural feature vector extracted by a GNN encoder. Based on the principles of physical chemistry, the information flow is decoupled into two parallel computational branches, including the thermodynamic intrinsic constant estimation branch (auxiliary branch) and the macroscopic environmental response prediction branch (main branch).

[0044] The auxiliary branch of this invention autonomously generates key physical parameters that dominate specific thermodynamic equations based on rigorous physical mechanisms, serving as a theoretical benchmark. To ensure physical authenticity, the input features of the auxiliary branch are configured differentially and adaptively according to the type of thermodynamic equation to be embedded.

[0045] The auxiliary branch autonomously extracts and outputs the thermodynamic reference constants that constrain the properties to be predicted based on the feature vector, including the following steps: When predicting phase equilibrium properties, such as the CO2 absorption of ionic liquids, the auxiliary branch concatenates the global structural feature vector of the ionic liquid with the temperature to construct a joint input feature matrix. Based on the joint input feature matrix, the activation function, and the weight matrix and bias vector that the auxiliary branch can learn autonomously, the enthalpy of dissolution is determined.

[0046] Based on the van der Hoff equation and Henry's law, the enthalpy of dissolution is mapped to temperature to obtain the logarithmic value of the Henry constant, a theoretical phase equilibrium reference parameter that fully conforms to the temperature-dependent evolution logic.

[0047] When predicting properties that depend on environmental conditions, such as gas-liquid phase equilibrium, in order to make the model move away from being a "pure mathematical black box," the auxiliary branch of this invention performs deep "hard encoding" of the parameterized mapping (Wx+b) of the dual-branch neural network with the classical thermodynamic analytical expression.

[0048] For example, taking CO2 uptake prediction as an example, the van der Hoff equation and Henry's law are embedded, and the specific implementation logic is as follows: First, the global structural feature vector of the ionic liquid extracted by the GNN encoder is... The ambient temperature T is concatenated along the feature dimension to construct a joint input feature matrix. Subsequently, the system's assisted fully connected layer performs the following linear transformation network operation to autonomously generate a latent variable with a clear physical meaning—the enthalpy of dissolution (…). ), as detailed below: (1) in, This represents the weight matrix that can be learned autonomously in the auxiliary branch. This represents the bias vector that can be learned autonomously in the auxiliary branch; This represents a specific activation function, whose core function is to act as a hard boundary, ensuring that the predicted enthalpy of dissolution is strictly limited to a physically reasonable range.

[0049] After automatically extracting the latent enthalpy of dissolution, the model no longer directly predicts the amount of absorption, but strictly follows the physical laws of the van der Hoff equations in standard macroscopic thermodynamics. Using the generated... In conjunction with the ambient temperature T, perform mapping calculations at the level of mathematical equations: (2) Where R is the ideal gas constant (8.314 J / mol·K). These are the model fitting coefficients (empirical constants) used to adjust the enthalpy of dissolution term relative to the Henry's constant. Influence weight, The model fitting constant (empirical intercept term) is physically equivalent to the integral constant C in the equation. Through the aforementioned physical fusion computational framework, the auxiliary branch precisely reconstructs the output of the discrete neuron into the logarithmic value of the theoretical phase equilibrium benchmark parameter—the Henry's constant—which fully conforms to the temperature-dependent evolution logic. This establishes the physical foundation for the main branch road.

[0050] When predicting rheological and transport properties, such as the viscosity of ionic liquids, the auxiliary branch cuts off the input of environmental dependent variables (temperature, pressure) in the feature extraction stage. The auxiliary branch is internally bifurcated into three parallel fully connected sub-layers. The three fully connected sub-layers generate viscosity limit parameters, apparent activation energy related parameters, and ideal glass transition temperature based on the global structural feature vector of the ionic liquid.

[0051] Based on the VFT equation (Vogel-Fulcher-Tammann equation), external temperature, viscosity limit parameter, apparent activation energy related parameter and ideal glass transition temperature, which characterize the external environment, are mapped to the theoretical reference value of viscosity logarithm.

[0052] When predicting rheological and transport characteristics such as fluid viscosity, since the constants that reflect the intrinsic rheological properties of the fluid must be physically independent of changes in external temperature, the auxiliary branch will adaptively cut off the input of environmental dependent variables (temperature, pressure) in the feature extraction stage and perform "hard encoding" through specific matrix operations and physical boundary activation functions at the underlying level.

[0053] For example, taking viscosity prediction as an example, the VFT equation is embedded, and the specific calculation logic is as follows: The adaptively adjusted auxiliary branch is determined solely by the invariant global structural feature vector of the ionic liquid. Single-drive. The auxiliary branch internally branches into three parallel fully connected sub-layers, which perform Wx+b linear mapping and nonlinear activation transformation with specific physical constraints on the three unknown state constants in the VFT equation, where W is the weight matrix, x is the input unknown state constant, and b is the bias constant.

[0054] Viscosity limiting parameter ( ): Represents the limiting viscous state of the system at extremely high temperatures, obtained using a direct linear mapping: (3) In the formula, Represents the linear mapping weight matrix. This represents the bias vector of the linear mapping.

[0055] Parameters related to apparent activation energy ( ): This represents the pseudo-activation energy required for a molecule to overcome the energy barrier of structural rearrangement, and it must be consistently positive at the physical boundary. Therefore, the network specifically applies a Softplus activation function to this output. (4) In the formula, This represents the weight matrix of the linear layer. This represents the bias vector of the linear layer.

[0056] Ideal glass transition temperature ( ): This represents the theoretical baseline temperature at which long-range molecular motion is completely frozen. To prevent the model from generating unrealistic and out-of-range temperatures during iterations, the network is forced to use the Sigmoid activation function to confine its linear output values ​​within strict physical normalization constraints. (5) In the formula, This represents the weight matrix of the linear layer. This represents the bias vector of the linear layer.

[0057] The above-mentioned physically reasonable spatial parameters were automatically extracted and constrained. After that, the system retains it as an internal constant module. The computation graph then forcibly introduces the external temperature, which characterizes the properties of the external environment. Combined with the above three parameter combinations, perform modified VFT equation-level mathematical mappings: (6) In the formula, This represents the theoretical baseline value for the logarithmic viscosity.

[0058] The main branch concatenates and maps the structural feature vector of the ionic liquid with the environmental dependent variable to output the macroscopic property prediction value of the ionic liquid, including the following steps: The global structural feature vector of the ionic liquid output by the GNN encoder is directly concatenated with the normalized environmental factors (such as temperature and pressure) to form a joint feature vector, which is then input into a multi-layer perceptron (MLP). After nonlinear transformation, the macroscopic property prediction values ​​of the ionic liquid are directly output.

[0059] For example: In a CO2 capture task, output the amount of CO2 absorbed ( In the viscosity prediction task, the logarithmic value of the viscosity of the ionic liquid is output. ).

[0060] S3. The joint loss function is determined by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term through empirical data fitting. The thermodynamic and physical constraint loss term forces the macroscopic property prediction value of the main branch to maintain physical consistency with the thermodynamic reference constant of the auxiliary branch.

[0061] In this embodiment of the invention, the macroscopic property prediction value of the main branch is forced to be consistent with the thermodynamic reference constant of the auxiliary branch by means of thermodynamic physical constraint loss term. This mechanism forcibly decouples and rebinds the output of the main branch from the theoretical reference of the auxiliary branch.

[0062] For example, for phase equilibrium property prediction (taking CO2 absorption prediction as an example, embedding Henry's phase equilibrium constraints), the thermodynamic physical constraint loss term includes Henry's law loss, physical monotonicity penalty loss of CO2 absorption with respect to pressure, and physical monotonicity penalty loss of CO2 absorption with respect to temperature; in CO2 absorption, the gas absorption at isothermal temperature is determined according to Henry's law theorem. Proportional to equilibrium partial pressure ( ),Right now In the formula, Represent the Henry's constant. Construct the Henry's law loss ( ), calculate the predicted absorption of the main branch ( The logarithmic value of the Henry's constant of the auxiliary branch ( The penalty residuals between () are as follows: (7) For processes involving gas absorption and exothermicity, determine the physical monotonicity penalty loss of CO2 absorption with respect to pressure and the physical monotonicity penalty loss of CO2 absorption with respect to temperature, including: Penalize any non-physical parameter iterative path that violates the principle that CO2 uptake increases with increasing pressure (gradient) or decreases with increasing temperature (gradient), and obtain the physical monotonicity penalty loss of CO2 uptake with respect to pressure, respectively. Physical monotonicity penalty loss of CO2 uptake with respect to temperature : (8) (9) In the formula, Represents the mean function, This represents the activation function of the rectified linear unit, used to impose a positive penalty on partial derivatives that violate physical laws.

[0063] For example, for rheological and transport-related physical properties (taking viscosity prediction as an example, embedding VFT rheological boundary constraints), the thermodynamic physical constraint loss term includes VFT equation consistency constraint loss, viscosity physical monotonicity penalty loss with respect to pressure, and viscosity physical monotonicity penalty loss with respect to temperature; in viscosity prediction, the logarithmic value of the predicted viscosity output by the forced main branch is used. It must be considered in conjunction with the viscosity limiting parameters, apparent activation energy parameters, ideal glass transition temperature, and theoretical logarithmic viscosity reference values ​​generated by the auxiliary branch. Maintaining high convergence ensures that the overall macroscopic predictions of the neural network will never deviate from the physical laws of the VFT, as follows: (10) In the formula, This represents the consistency constraint loss of the VFT equation. This represents the Mean Squared Error function.

[0064] For example, regarding the prediction of rheological and transport-related physical properties (taking viscosity prediction as an example), considering rheological dynamics, a rigid barrier is imposed to ensure that the global solution space for fluid viscosity prediction strictly maintains an increase in pressure and a decrease in temperature, respectively, by constructing a physical monotonicity penalty loss for viscosity with respect to pressure. Physical monotonicity penalty loss of viscosity with respect to temperature The details are as follows: (11) (12) The joint loss function is determined by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term through empirical data fitting, including the following steps: The empirical data is fitted with the residual loss term, and then compared with the total thermodynamic and physical constraint loss term and the weighted coefficient hyperparameter. Summing the products of , we obtain the joint loss function, as follows: (13) In the formula, Denotes the joint loss function. This represents the residual loss term from fitting empirical data. This represents the total thermodynamic and physical constraint loss term.

[0065] To ensure the network converges robustly to strict physical constraints from a purely data-driven approach, this invention employs specific empirical data loss methods for different physical prediction scenarios. ) and thermodynamic loss ( Dynamic fusion construction of ).

[0066] For example, for phase equilibrium property prediction (taking CO2 uptake prediction as an example), the residual loss term of the empirical data fitting of CO2 uptake ( The data is constructed by comparing the actual output of the main branch with experimental data points, and the CO2 absorption directly predicted by the main branch is extracted. ) and the experimentally measured baseline value ( The mean square error is calculated as follows: (14) The Henry's Law loss and the physical monotonicity penalty loss of CO2 absorption with respect to pressure are considered. And the physical monotonicity penalty loss of CO2 uptake with respect to temperature By performing linear spatial superposition, the total thermodynamic and physical constraint loss term for the predicted CO2 uptake is obtained as follows: (15) In the formula, This represents the total thermodynamic and physical constraint loss term in the prediction of CO2 absorption.

[0067] For example, in predicting rheological and transport properties (taking viscosity prediction as an example), the residual loss term is fitted using empirical viscosity data. By benchmarking rheological experimental data, the logarithmic value of the ionic liquid viscosity predicted by the main control circuit is extracted. , compared with the true logarithmic value represented by the experiment Supervised calibration is performed as follows: (16) The VFT equation consistency constraint loss and the viscosity physical monotonicity penalty loss with respect to pressure are applied. and the physical monotonicity penalty loss of viscosity with respect to temperature By performing linear spatial superposition, the total thermodynamic and physical constraint loss term for viscosity prediction is obtained. The details are as follows: (17) The embodiments of this invention are like putting an unbreakable mathematical-physical exoskeleton on the "pure mathematical black box" of AI algorithms, by... and Coupling completely overturns the limitations of previous GNN models that only tuned parameters based on predicted MSE. It dynamically establishes a dual balance between empirical data fitting and adherence to basic thermodynamic principles, ensuring that the extremely complex neural network parameter mapping network always converges steadily in a spatial domain with rigorous physical meaning.

[0068] S4. Backpropagation and parameter optimization are performed on the dual-branch neural network based on the joint loss function to complete the model training; the characteristics of the ionic liquid to be tested and environmental factors (temperature and pressure) are input into the trained dual-branch neural network, and the main branch outputs the final property prediction results of the ionic liquid that conform to the thermodynamic evolution law.

[0069] This invention determines the joint loss function by dynamically weighting and fusing the residual loss term fitted with empirical data with the thermodynamic physical constraint loss term. This breaks the core control center of the "pure mathematical black box" and eliminates "non-physical prediction phenomena". Throughout the entire training lifecycle of neural network backpropagation and weight optimization, this invention overturns the conventional criterion of relying solely on the "MSE error between the predicted value and the experimental true value" and innovatively introduces a physical penalty term constructed from the thermodynamic theory system, forcing the model to obey physical laws from the top level.

[0070] Based on the above methods for predicting the properties of ionic liquids, such as Figure 2 As shown, this embodiment of the invention also provides an ionic liquid property prediction system, including a data processing module, a model building module, a loss function building module, and a prediction module.

[0071] The data processing module is used to convert the ionic liquid structure in the ionic liquid dataset into a molecular graph, and to use GNN to extract features and aggregate information from the molecular graph to obtain a global structural feature vector of the ionic liquid that does not change with environmental factors.

[0072] The model building module is used to construct a dual-branch neural network, including a main branch and an auxiliary branch. The dual-branch neural network is used to decouple the calculation of the input topological feature vector. The main branch concatenates and maps the ionic liquid structural feature vector with the environmental dependent variable, and outputs the macroscopic property prediction value of the ionic liquid. The auxiliary branch extracts and outputs the thermodynamic reference constants that constrain the property to be predicted based on the feature vector and related environmental dependent variables.

[0073] The loss function construction module is used to determine the joint loss function by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term through empirical data fitting. The thermodynamic and physical constraint loss term forces the macroscopic property prediction values ​​of the main branch to maintain physical consistency with the thermodynamic reference constants of the auxiliary branch.

[0074] The prediction module is used to perform backpropagation and parameter optimization on the dual-branch neural network based on the joint loss function to complete the model training. The characteristics and environmental factors of the ionic liquid to be tested are input into the trained dual-branch neural network, and the main branch outputs the final property prediction results of the ionic liquid that conform to the thermodynamic evolution law.

[0075] To rigorously verify the technological advancement and practical industrial value of the ionic liquid property prediction method and system of this invention, a comparative experiment was conducted to systematically evaluate the performance of the standard pure data-driven graph neural network (hereinafter referred to as standard GNN) and the dual-branch neural network (hereinafter referred to as TCGNN) with thermodynamic physical constraints proposed in this invention.

[0076] 1. Verification of basic prediction accuracy and analysis of beneficial effects.

[0077] Feasibility verification implementation: The experiment selected five mainstream GNN architectures (including MPNN, GAT, GIN, PNA, and Attentive FP) as bases. Using the same ionic liquid thermodynamics dataset and a rigorous data partitioning method, both the standard GNN and the TCGNN of this embodiment were trained to evaluate the model's real-world generalization ability for novel ionic liquids. Experimental Results and Beneficial Effects Analysis: Comprehensive and superior performance indicators, such as... Figure 3 As shown, histograms (a) and (b) represent the coefficients of determination (R²) for predicting CO2 absorption properties, respectively. 2 Comparing the root mean square error (RMSE) with histograms (c) and (d), the R values ​​predicted for viscosity properties are shown in the histograms (c and d). 2 Compared with RMSE.

[0078] In histograms (a), (b), (c), and (d), the blue bars represent the current mainstream GNN architectures (including MPNN, GAT, GIN, PNA, and Attentive FP), and the red bars represent the TCGNN of this invention.

[0079] Histogram (a) shows that, for predicting CO2 absorption properties, TCGNN has the highest R-value across all architectures. 2 All of them are higher than MPNN, GAT, GIN, PNA and Attentive FP. Histogram (b) shows that for predicting CO2 absorption properties, TCGNN's RMSE is lower than MPNN, GAT, GIN, PNA and Attentive FP on all architectures.

[0080] Histogram (c) shows that, for viscosity property prediction, TCGNN has the highest R-value across all architectures. 2 All of them are higher than MPNN, GAT, GIN, PNA and Attentive FP. Histogram (d) shows that for viscosity property prediction, TCGNN has a lower RMSE than MPNN, GAT, GIN, PNA and Attentive FP on all architectures.

[0081] Depend on Figure 3The test set comparison results show that the dual-branch neural network with physical constraints proposed in this embodiment of the invention achieves a comprehensive leap in prediction accuracy compared to almost all GNN-based systems. Taking the PNA architecture as an example, after incorporating physical constraints, its performance in CO2 capacity prediction tasks jumps to an extremely high level, while RMSE decreases significantly; the improvement in accuracy is even more significant in complex rheological mappings such as viscosity prediction.

[0082] Extremely close scatter distribution, such as Figure 4 As shown, scatter plot (a) shows the fitting performance of CO2 absorption, and scatter plot (b) shows the fitting performance of the logarithmic value of fluid viscosity. Taking the PNA architecture as an example, the scatter plots formed by the predicted values ​​(vertical axis) generated by the embodiments of the present invention and the actual experimental test values ​​(horizontal axis) converge very closely to both sides of the diagonal (y=x). Whether for predicting CO2 absorption constrained by the environment or for predicting viscosity, the present invention achieves high accuracy.

[0083] The beneficial effects are as follows: it proves that embedding the macroscopic thermodynamic framework as a "penalty term" into the GNN model not only does not destroy the model's original ability to fit massive amounts of data, but also plays an extremely effective role in forced regularization. It forces the underlying message passing mechanism to stop blindly "memorizing" the data distribution (suppressing overfitting) and instead truly explore the topological features that produce causal laws with macroscopic properties, thus thoroughly improving the algorithm's generalization prediction accuracy in the chemical space of new materials.

[0084] 2. Physical consistency verification and analysis of the beneficial effects of extrapolation robustness.

[0085] Feasibility verification implementation: To examine the model's behavioral boundaries after leaving the comfort zone (i.e., the data-dense region of the training set), the experiment extracted the same specific representative ionic liquid and continuously plotted its theoretical CO2 phase equilibrium absorption curves under different conditions using both the standard GNN and the TCGNN of this embodiment. Specific observations included: ① continuous increase in system temperature under normal pressure; ② continuous extension of system pressure towards the ultra-high pressure region (>40000 kPa, a sparse region without training data) under normal temperature conditions. The results were compared to see if the two predicted curves conformed to the logic of real matter evolution. Figure 5 Presented in the middle.

[0086] Experimental results and analysis of beneficial effects, such as Figure 5 As shown, trajectory comparison diagram (a) compares the predicted CO2 absorption curves under normal pressure with temperature changes, and trajectory comparison diagram (b) compares the predicted CO2 absorption curves under normal temperature with increased pressure.

[0087] pass Figure 5The microscopic trajectory tracking of a single specific molecule reveals that the standard GNN, acting as a "pure mathematical black box," is highly susceptible to catastrophic "physical collapse" under extreme conditions. For example, at high temperatures, its predicted CO2 absorption exhibits an "abnormal rise" contrary to common sense; and when faced with extreme pressure ranges exceeding the training coverage, the standard GNN's prediction curve drops precipitously. In stark contrast, the TCGNN proposed in this embodiment, which incorporates physical constraints, demonstrates a monotonous and smooth trend in its prediction curve under any extreme temperature and pressure, strictly adhering to and conforming to the prior physical laws governing CO2 gas absorption.

[0088] The beneficial effects are as follows: It profoundly confirms the successful closed loop of the "automatic partial derivative differential penalty" and "cross-branch equation fusion" mechanism in the embodiments of the present invention. From the lowest level of mathematical reasoning, it completely blocks the weight iteration path of the network deviating from the physical law, and endows the model of the present invention with extremely strong extrapolation robustness. This makes the technical solution of the present invention not only limited to serving as an interpolation fitting tool in the laboratory, but also able to safely resist the risk of data loss caused by extreme working conditions such as high temperature and deep pressure in real industry.

[0089] The prediction method and system of this invention, as well as their known and potential application areas: CCUS (Carbon Capture, Utilization, and Storage) and Green Separation Engineering: This is the most direct known application. It can be used for high-throughput virtual screening of millions of unknown ionic liquids and deep eutectic solvents (DES) when absorbing carbon-containing gas particles in industrial flue gas, to discover "perfect solvents" with extremely high adsorption capacity and easy flow, thus accelerating the implementation of dual-carbon goals.

[0090] Development of High-Energy Battery Electrolytes and Energy Storage Materials: Because the dual-branch neural network in this invention can embed rheological and transport equations, it has great potential for application in the design of novel wide-temperature-range electrolyte systems for lithium / sodium-ion batteries. By predicting its ionic conductivity and viscosity under extremely cold / hot conditions, long-life, highly safe energy storage electrolyte materials can be screened.

[0091] Chemical Industry Software and Digital Twin Systems: The lightweight and highly accurate dual-branch neural network of this invention makes it a promising replacement for traditional thermodynamic fluid property packages, which are extremely bulky and time-consuming (or require repeated trial and error to calibrate parameters), and can become the underlying advanced property estimation engine for next-generation chemical process simulation software (such as Aspen Plus and process digital twin control systems).

[0092] The prediction method and system product implementation and application of this invention are as follows: The "AI + Materials Discovery" SaaS cloud platform product deploys the algorithm architecture built in this invention on a cloud server, developing it into a SaaS service platform for research institutes and new materials companies. Users only need to input the desired combination of anions and cations (SMILES string) and the target temperature / pressure conditions, and the platform can generate a physically reliable thermodynamic property evaluation report in seconds.

[0093] Embedded core algorithm module licensing (API interface service): Provides calling services to existing materials information platforms or industrial IoT control systems in the form of an API black-box interface. During continuous fluid production, industries can use the monitored system T / P data and specific fluid molecular structures to call the algorithm interface of this invention in real time, instantly obtaining dynamic phase equilibrium values ​​and fluid state changes that are difficult to measure directly within the equipment, for process control and early warning.

[0094] A reverse-engineered "AI-generated" novel solvent design closed-loop tool: In future evolution, the "TCGNN property evaluator" of this invention will serve as the "reward function" for generative adversarial networks (GANs) or reinforcement learning (RL). Novel molecules will be randomly assembled by the generative model, and the physical feasibility and quality of their properties will be rigorously verified by this model, forming a fully automated pipeline for "AI-invented novel molecules."

[0095] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the physical properties of ionic liquids, characterized in that, Includes the following steps: The ionic liquid structures in the ionic liquid dataset are transformed into molecular graphs. A Generative Neural Network (GNN) is used to extract features and aggregate information from the molecular graphs to obtain global structural feature vectors of the ionic liquids that are independent of environmental factors. The process includes the following steps: acquiring molecular composition information of the ionic liquids and constructing a molecular graph; extracting atomic properties as initial node features and chemical bond properties as initial edge features; and constructing an adjacency matrix representing topological connections based on inter-atomic connections. The initial node features, initial edge features, and adjacency matrix are input into the GNN. Based on the neighbor relationships defined by the adjacency matrix, the GNN performs multiple rounds of local node feature interaction and updates through an information transfer mechanism to obtain deep node feature vectors. All deep node feature vectors are then globally pooled and aggregated using a readout function, mapped to a fixed-dimensional graph-level feature representation, which is used as the global structural feature vector of the ionic liquids that is independent of environmental factors. The dual-branch neural network is constructed, consisting of a main branch and an auxiliary branch. The global structural feature vector of the ionic liquid is input into the dual-branch neural network for decoupled computation. The main branch concatenates and maps the global structural feature vector of the ionic liquid with environmental dependent variables, outputting the predicted macroscopic properties of the ionic liquid. The auxiliary branch extracts and outputs the thermodynamic reference constants constraining the predicted properties based on the feature vector and relevant environmental dependent variables, including the following steps: When predicting phase equilibrium properties, the auxiliary branch concatenates the global structural feature vector of the ionic liquid with the temperature environment variable to construct a joint input feature matrix. Based on the joint input feature matrix, activation function, weight matrix and bias vector that the auxiliary branch can learn autonomously, the enthalpy of dissolution is determined. Based on the van der Hoff equation and Henry's law, the enthalpy of dissolution is mapped to temperature to obtain the logarithmic value of Henry's constant. When predicting rheological and transport-related physical properties, the auxiliary branch cuts off the input of environmental dependent variables in the feature extraction stage. The auxiliary branch is internally bifurcated into three parallel fully connected sublayers. The three fully connected sublayers generate viscosity limit parameters, apparent activation energy related parameters, and ideal glass transition temperature based on the global structural feature vector of the ionic liquid, respectively. Based on the VFT equation, the external temperature, viscosity limit parameters, apparent activation energy related parameters, and ideal glass transition temperature, which characterize the external environmental properties, are mapped to the theoretical benchmark value of the viscosity logarithm. The joint loss function is determined by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term by fitting empirical data. The thermodynamic and physical constraint loss term forces the macroscopic property prediction value of the main branch to be consistent with the thermodynamic reference constant of the auxiliary branch in accordance with the physical law. Backpropagation and parameter optimization are performed on the dual-branch neural network based on the joint loss function to complete the model training. The characteristics and environmental factors of the ionic liquid to be tested are input into the trained dual-branch neural network, and the main branch outputs the final property prediction results of the ionic liquid that conform to the thermodynamic evolution law.

2. The method for predicting the physical properties of ionic liquids according to claim 1, characterized in that, The main branch concatenates and maps the global structural feature vector of the ionic liquid with environmental dependent variables to output the macroscopic property prediction values ​​of the ionic liquid, including the following steps: The global structural feature vector of the ionic liquid is directly concatenated with the normalized environmental factors to form a joint feature vector, which is then input into a multilayer perceptron. After nonlinear transformation, the macroscopic property prediction values ​​of the ionic liquid are directly output.

3. The method for predicting the physical properties of ionic liquids according to claim 1 or 2, characterized in that, For phase equilibrium property prediction, the thermodynamic physical constraint loss terms include Henry's law loss, physical monotonicity penalty loss of CO2 uptake with respect to pressure, and physical monotonicity penalty loss of CO2 uptake with respect to temperature.

4. The method for predicting the physical properties of ionic liquids according to claim 1 or 2, characterized in that, For rheological and transport-related physical properties, the thermodynamic physical constraint loss terms include the VFT equation consistency constraint loss, the physical monotonicity penalty loss of viscosity with respect to pressure, and the physical monotonicity penalty loss of viscosity with respect to temperature.

5. The method for predicting the physical properties of ionic liquids according to claim 3, characterized in that, Determining the Henry's Law loss involves the following steps: It is determined that the gas absorption at constant temperature is proportional to the equilibrium partial pressure. A penalty residual is calculated between the predicted absorption of the main branch and the logarithmic value of the Henry's constant of the auxiliary branch, based on the Henry's law loss calculation.

6. The method for predicting the physical properties of ionic liquids according to claim 4, characterized in that, Determine the consistency constraint loss of the VFT equations, including: In viscosity prediction, the predicted logarithmic viscosity output by the forced main branch must converge with the viscosity limit parameter, apparent activation energy related parameter, ideal glass transition temperature, and theoretical reference value of viscosity logarithm generated by the auxiliary branch.

7. The method for predicting the physical properties of ionic liquids according to claim 3, characterized in that, Determine the physical monotonicity penalty loss of CO2 uptake with respect to pressure and the physical monotonicity penalty loss of CO2 uptake with respect to temperature, including: By penalizing any non-physical parameter iteration path that violates the principle that CO2 absorption increases with increasing pressure or decreases with increasing temperature, we obtain the physical monotonicity penalty loss of CO2 absorption with respect to pressure and the physical monotonicity penalty loss of CO2 absorption with respect to temperature, respectively.

8. The method for predicting the physical properties of ionic liquids according to claim 4, characterized in that, Determining the physical monotonicity penalty loss of viscosity with respect to pressure and the physical monotonicity penalty loss of viscosity with respect to temperature includes the following steps: To ensure that the global solution for fluid viscosity prediction increases with increasing pressure and decreases with increasing temperature, physical monotonicity penalty losses for viscosity with respect to pressure and physical monotonicity penalty losses for viscosity with respect to temperature are constructed respectively.

9. The method for predicting the physical properties of ionic liquids according to claim 1, characterized in that, The joint loss function is determined by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term using empirical data fitting, including the following steps: The joint loss function is obtained by summing the product of the empirical data fitting residual loss term, the total thermodynamic and physical constraint loss term, and the weighted coefficient hyperparameters.

10. The method for predicting the physical properties of ionic liquids according to claim 9, characterized in that, Determining the residual loss term for fitting empirical data includes the following steps: For phase equilibrium property prediction, the mean square error of the CO2 mole fraction directly predicted by the main branch and the experimentally measured baseline value is calculated to obtain the fitting residual loss term of the empirical data of CO2 absorption. For the prediction of rheological and transport properties, the logarithmic value of the ionic liquid viscosity predicted by the main control branch is extracted and supervisedly calibrated with the actual logarithmic value characterized by experiments to obtain the residual loss term of viscosity empirical data fitting.

11. The method for predicting the physical properties of ionic liquids according to claim 9, characterized in that, Determining the total thermodynamic and physical constraint loss term includes the following steps: For phase equilibrium property prediction, the Henry's law loss, the physical monotonicity penalty loss of CO2 uptake with respect to pressure, and the physical monotonicity penalty loss of CO2 uptake with respect to temperature are linearly superimposed to obtain the total thermodynamic physical constraint loss term for CO2 uptake prediction. For predicting rheological and transport-related physical properties, the consistency constraint loss of the VFT equation, the physical monotonicity penalty loss of viscosity with respect to pressure, and the physical monotonicity penalty loss of viscosity with respect to temperature are linearly spatially superimposed to obtain the total thermodynamic physical constraint loss term for viscosity prediction.

12. A system for predicting the physical properties of ionic liquids, characterized in that, include: The data processing module is used to transform the ionic liquid structure in the ionic liquid dataset into a molecular graph. It then uses a Generative Neural Network (GNN) to extract features and aggregate information from the molecular graph to obtain a global structural feature vector of the ionic liquid that is independent of environmental factors. The process includes the following steps: acquiring the molecular composition information of the ionic liquid and constructing a molecular graph; extracting atomic properties as initial node features and chemical bond properties as initial edge features; and constructing an adjacency matrix representing topological connections based on the connections between atoms. The initial node features, initial edge features, and adjacency matrix are input into the GNN. Based on the neighbor relationships defined by the adjacency matrix, the GNN performs multiple rounds of local node feature interaction and updates through an information transfer mechanism to obtain deep node feature vectors. Finally, a global pooling aggregation is performed on all deep node feature vectors using a readout function, mapping them to a fixed-dimensional graph-level feature representation, which is then used as the global structural feature vector of the ionic liquid that is independent of environmental factors. The model building module is used to construct a dual-branch neural network, including a main branch and an auxiliary branch. This dual-branch neural network is used to decouple the calculation of the input global structural feature vector of the ionic liquid. The main branch concatenates and maps the global structural feature vector of the ionic liquid with environmental dependent variables, outputting the predicted macroscopic properties of the ionic liquid. The auxiliary branch extracts and outputs the thermodynamic reference constants that constrain the predicted properties based on the feature vector and relevant environmental dependent variables, including the following steps: When predicting phase equilibrium properties, the auxiliary branch concatenates the global structural feature vector of the ionic liquid with the temperature environment variable to construct a joint input feature matrix. Based on the joint input feature matrix, activation function, weight matrix and bias vector that the auxiliary branch can learn autonomously, the enthalpy of dissolution is determined. Based on the van der Hoff equation and Henry's law, the enthalpy of dissolution is mapped to temperature to obtain the logarithmic value of Henry's constant. When predicting rheological and transport-related physical properties, the auxiliary branch cuts off the input of environmental dependent variables in the feature extraction stage. The auxiliary branch is internally bifurcated into three parallel fully connected sublayers. The three fully connected sublayers generate viscosity limit parameters, apparent activation energy related parameters, and ideal glass transition temperature based on the global structural feature vector of the ionic liquid, respectively. Based on the VFT equation, the external temperature, viscosity limit parameters, apparent activation energy related parameters, and ideal glass transition temperature, which characterize the external environmental properties, are mapped to the theoretical benchmark value of the viscosity logarithm. The loss function construction module is used to determine the joint loss function by dynamically weighting and fusing the residual loss term and the thermodynamic and physical constraint loss term through empirical data fitting. The thermodynamic and physical constraint loss term forces the macroscopic property prediction value of the main branch to maintain physical consistency with the thermodynamic reference constant of the auxiliary branch. The prediction module is used to perform backpropagation and parameter optimization on the dual-branch neural network based on the joint loss function to complete the model training. The characteristics and environmental factors of the ionic liquid to be tested are input into the trained dual-branch neural network, and the main branch outputs the final property prediction results of the ionic liquid that conform to the thermodynamic evolution law.

Images