Physical information constrained petrochemical material hybrid modeling physical property calculation system and method thereof

Abstract

The invention relates to the field of petrochemical material physical property calculation, in particular to a physical information constrained petrochemical material hybrid modeling physical property calculation system and method, and the system comprises a data processing module which achieves the standardized cleaning and fusion of multi-source heterogeneous data; the multi-scale characteristic module realizes material characteristic characterization through quantum-mesoscopic-macroscopic cross-scale modeling; the physical constraint module is embedded into a thermodynamic constitutive equation and a phase equilibrium criterion to ensure the self-consistency of the model; the component interaction module constructs a non-ideal mixing effect prediction model based on a deep potential energy field theory, and establishes a component interaction knowledge graph; the uncertainty quantification module adopts a Bayesian deep learning method to assess and predict a confidence interval, a system integrates a physical mechanism and a data driving method, physical property prediction errors are reduced by 15%-20%, meanwhile, a molecular structure-physical property associated visual analysis tool is provided, process optimization and new product development decision are supported, and the reliability of the system is improved. And the core requirements of the petrochemical industry on high-precision and interpretable physical property prediction are met.

Description

Technical Field

[0001] The present invention relates to the field of petrochemical material property calculation, and specifically to a physical information-constrained petrochemical material mixture modeling property calculation system and method thereof, which can accurately predict various physical and chemical properties of petrochemical material mixtures based on multi-scale feature representation and physical constraints. Background Art

[0002] The prediction of petrochemical material properties is a key link in the petrochemical industry and is of great significance for process design, product development, and quality control. The traditional property prediction methods mainly include three categories: experimental measurement method, empirical correlation method, and theoretical calculation method. Although the experimental measurement method has high accuracy, it is time-consuming and costly, and it is difficult to meet the needs of rapid research and development; the empirical correlation method relies on a large amount of historical data, and the prediction ability is limited by the scope and quality of historical data; the theoretical calculation method can predict material properties starting from the molecular structure, but the calculation complexity is high, and it is difficult to apply to complex mixture systems.

[0003] In recent years, with the development of machine learning technology, data-driven property prediction methods have gradually emerged, such as QSPR (Quantitative Structure-Property Relationship) method, artificial neural network, etc. These methods can learn complex non-linear relationships from a large amount of data and achieve relatively rapid property prediction. However, the existing data-driven methods have the following problems: First, most of them are single-component models and are difficult to accurately handle the component interaction effects in mixtures; second, the prediction results often violate basic physical laws, such as the laws of thermodynamics; third, the feature representation is limited to a single scale and is difficult to comprehensively capture the relationship between molecular structure and macroscopic material properties; fourth, the prediction ability for data-sparse regions is weak; fifth, the uncertainty assessment is insufficient and it is difficult to support reliable decision-making.

[0004] Therefore, there is an urgent need for a petrochemical material mixture property prediction system that can integrate physical laws and multi-scale features to improve prediction accuracy and reliability on the premise of ensuring physical rationality. Summary of the Invention

[0005] The purpose of the invention is to provide a physical information-constrained petrochemical material mixture modeling property calculation system and method thereof. The system integrates physical law constraints and multi-scale feature representation, can accurately predict various physical and chemical properties of petrochemical material mixtures, and at the same time provides reliable uncertainty assessment.

[0006] The present invention proposes a physical information-constrained petrochemical material mixture modeling property calculation system, including: Data processing and management module, for: collecting multi-source data of petrochemical materials; cleaning and quality controlling the multi-source data; integrating the multi-source data into standard format data; Multi-scale feature characterization module, communicatively connected to the data processing and management module, for: receiving the standard format data; constructing multi-level petrochemical material feature characterization from quantum scale to macroscopic scale; realizing the fusion of features at different scales; Physical constraint modeling module, communicatively connected to the multi-scale feature characterization module, for: receiving the multi-level petrochemical material feature characterization; embedding thermodynamic laws and physical law constraints into the model structure; ensuring that the physical property prediction results meet the requirements of thermodynamic consistency; Component interaction modeling module, communicatively connected to the multi-scale feature characterization module and the physical constraint modeling module, for: accurately capturing the complex interactions between different components in the petrochemical material mixture; predicting the non-ideal mixing effect between components; constructing a component interaction knowledge base; Uncertainty quantification module, communicatively connected to the physical constraint modeling module and the component interaction modeling module, for: evaluating the reliability of physical property prediction results; quantifying the sources and magnitudes of prediction uncertainties; generating confidence intervals for prediction results; Physical property prediction and analysis module, communicatively connected to the physical constraint modeling module, the component interaction modeling module and the uncertainty quantification module, for: predicting various physical and chemical properties of petrochemical materials based on the outputs of the above modules; analyzing the relationship between molecular structure and physical properties; providing visualization and interpretation functions.

[0007] Physical information-constrained petrochemical material mixing modeling physical property calculation method, including: collecting multi-source data of petrochemical materials, and cleaning, quality controlling and standardizing the multi-source data to obtain standard format data; based on the standard format data, constructing multi-level petrochemical material feature characterization from quantum scale to macroscopic scale, and realizing the fusion of features at different scales; embedding thermodynamic laws and physical law constraints into the model structure to ensure that the physical property prediction results meet the requirements of thermodynamic consistency; accurately capturing the complex interactions between different components in the petrochemical material mixture, predicting the non-ideal mixing effect between components, and constructing a component interaction knowledge base; evaluating the reliability of physical property prediction results, quantifying the sources and magnitudes of prediction uncertainties, and generating confidence intervals for prediction results; based on the above steps, predicting various physical and chemical properties of petrochemical materials, analyzing the relationship between molecular structure and physical properties, and providing visualization and interpretation functions.

[0008] The beneficial effects of the present invention include: 1. Through multi-scale feature fusion, comprehensively characterize the structure-property relationship of petrochemical materials from quantum scale to macroscopic scale, and improve the prediction accuracy. The physical property prediction accuracy is increased by more than 30% compared with traditional methods, which is particularly obvious for complex mixture systems.

[0009] 2. By embedding the laws of thermodynamics and physical constraints into the model structure, ensure that the prediction results conform to the basic physical laws and avoid non - physical solutions. The thermodynamic consistency error of the system prediction results is reduced to less than 1 / 5 of the traditional method.

[0010] 3. Through a specially designed component interaction modeling module, accurately capture the complex interactions and synergistic effects between different components in the mixture, and solve the limitations of the single - component method. For highly non - ideal mixture systems, the prediction error of the activity coefficient is reduced by more than 40%.

[0011] 4. It can simultaneously predict more than 15 key physical properties, including thermodynamic properties, transport properties, combustion properties, and environmental safety properties, meeting the comprehensive requirements of the petrochemical industry for various physical property parameters.

[0012] 5. Through systematic uncertainty quantification techniques, provide reliable prediction uncertainty assessment to support risk - based decision - making. The confidence interval coverage rate of the prediction results reaches more than 95%, which is highly consistent with the actual statistical distribution.

[0013] 6. Through transfer learning and knowledge distillation techniques, effectively solve the data sparsity problem, achieve efficient model generalization, and reduce the dependence on a large amount of data. When there is only 10% of the training data, the model performance degradation does not exceed 15%.

[0014] 7. Provide clear physical interpretations and intuitive visualization functions to assist scientific understanding and process optimization, and enhance user trust.

[0015] 8. Seamlessly dock with existing petrochemical industrial process simulation systems and process design platforms for easy practical application and promotion. The system can be integrated with mainstream simulation software such as Aspen Plus and HYSYS, and support standard data format exchange. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is the overall architecture diagram of the system of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] The present invention will be further described in detail below with reference to the drawings and specific embodiments. Those skilled in the art should understand that these embodiments are only used to illustrate the present invention and should not be construed as limiting the present invention.

[0018] Refer to Figure 1 , the physical - information - constrained petrochemical material mixing modeling physical property calculation system provided by the present invention includes: a data processing and management module 1, a multi - scale feature characterization module 2, a physical constraint modeling module 3, a component interaction modeling module 4, an uncertainty quantification module 5, and a physical property prediction and analysis module 6. In addition, the system may further include a model training and optimization module 7 and a system integration and application module 8.

[0019] The data processing and management module 1 is used to collect multi-source data of petrochemical materials, clean, quality-control, and standardize these data, and finally form data in a standard format. The multi-scale feature characterization module 2 receives this data in the standard format, constructs a multi-level petrochemical material feature characterization from the quantum scale to the macroscopic scale, and realizes the fusion of features at different scales. The physical constraint modeling module 3 receives the multi-level petrochemical material feature characterization, embeds the constraints of thermodynamic laws and physical rules into the model structure to ensure that the physical property prediction results meet the requirements of thermodynamic consistency. The component interaction modeling module 4 accurately captures the complex interactions between different components in the petrochemical material mixture, predicts the non-ideal mixing effects between components, and constructs a component interaction knowledge base. The uncertainty quantification module 5 evaluates the reliability of the physical property prediction results, quantifies the sources and magnitudes of the prediction uncertainties, and generates the confidence intervals of the prediction results. The physical property prediction and analysis module 6 predicts various physical and chemical properties of petrochemical materials based on the outputs of the above modules, analyzes the relationship between molecular structure and physical properties, and provides visualization and interpretation functions.

[0020] Preferably, the system further includes a model training and optimization module 7, which is used to implement multi-task joint learning, optimize the implementation of physical constraints, optimize the model architecture and hyperparameters, and evaluate the model performance and select the optimal model. In addition, the system may further include a system integration and application module 8, which is used to provide a user interaction interface, realize the integration with industrial systems, execute data exchange and standardization, and support system deployment and expansion, and adapt to different application scenarios.

[0021] The data processing and management module 1 includes a multi-source data acquisition unit, a data cleaning and quality control unit, a data integration and standardization unit, and a knowledge graph construction unit.

[0022] The multi-source data acquisition unit is used to obtain petrochemical material data from multiple sources. In an embodiment of the present invention, the multi-source data acquisition unit can collect data from three main data sources: experimental databases (such as NIST ThermoDataEngine, DIPPR database, etc.), theoretical calculation results (such as quantum chemical calculations, molecular dynamics simulations, etc.), and industrial process data (such as factory production data, quality control data, etc.). Preferably, the multi-source data acquisition unit supports multiple data formats, including CSV, Excel, JSON, and XML, etc., and provides an API interface for automated data acquisition.

[0023] The data cleaning and quality control unit is communicatively connected to the multi-source data acquisition unit, receives the collected raw data, and cleans and controls its quality. This unit first detects and processes outliers, using statistical methods (such as the Z-score, IQR method) and physical laws (such as the laws of thermodynamics) to identify abnormal data. For example, for liquid density data, when the temperature increases, the density usually decreases. If there is a data point where the density increases instead when the temperature increases, it may be an outlier. Secondly, this unit verifies the thermodynamic consistency of the data, checking whether the data satisfies the basic thermodynamic equations such as the Maxwell relation. Finally, a quality score is given to the data, evaluating the data quality from three dimensions of integrity, accuracy, and consistency, and assigning weights according to factors such as the reliability of the data source, the data acquisition method, and the data age. For example, the standard reference data certified by NIST may obtain a quality score of 0.9 - 1.0, while the literature data without peer review may only have a score of 0.5 - 0.7.

[0024] The data integration and standardization unit is communicatively connected to the data cleaning and quality control unit, receives the cleaned data, and integrates it into a standard format. This unit first unifies the physical property data format, defining a standard structure that includes substance identifiers, condition parameters (temperature, pressure, etc.), and physical property values. Secondly, it standardizes the molecular representation, uniformly converting molecular structures in different formats into SMILES or InChI formats. For example, converting a molecular structure diagram, molecular name, or CAS number into a standard SMILES representation. Thirdly, it converts the physical property units, unifying the physical property data in different unit systems into the SI unit or other selected standard unit systems. For example, converting the temperature from Fahrenheit or Kelvin to Celsius, and converting the pressure from psi or bar to Pa.

[0025] The knowledge graph construction unit is communicatively connected to the data integration and standardization unit, and constructs a petrochemical material knowledge graph based on the standardized data. This unit first extracts substance entities and property relationships, identifying chemical substances, physical property parameters, and their relationships from text and structured data. Secondly, it constructs substance-property-condition triples, forming the basic units of the knowledge graph. For example, (n-heptane, density, 20°C / 1atm) = 0.684 g / cm³. Finally, it realizes knowledge reasoning and completion, using known relationships to infer unknown knowledge and fill in data gaps. For example, given the densities at two temperatures, the density at an intermediate temperature can be inferred by linear interpolation.

[0026] In one embodiment of the present invention, the data processing and management module 1 adopts a distributed data storage architecture, uses MongoDB to store unstructured data, PostgreSQL to store structured data, and Neo4j to store knowledge graphs. Inside the module, a message queue (such as Kafka) is used for data transmission to ensure the efficiency and reliability of the data processing flow. In addition, this module implements a version control mechanism to record all data changes and supports data backtracking and comparison.

[0027] The multi-scale feature characterization module 2 includes a quantum chemical feature characterization unit, a molecular structure feature characterization unit, a mesoscopic scale feature characterization unit, a macroscopic thermodynamic feature characterization unit, and a multi-scale feature fusion unit.

[0028] The quantum chemical feature characterization unit is used to construct molecular features from the quantum scale. In one embodiment of the present invention, this unit extracts molecular electronic structure features based on the density functional theory (DFT) calculation method, such as the energy levels of the frontier molecular orbitals (HOMO and LUMO), energy gap, and electron density distribution. At the same time, multi-pole moment parameters such as molecular dipole moment and quadrupole moment are calculated to characterize the polarity and charge distribution of the molecule. In addition, intermolecular interactions are also characterized, and different types of intermolecular interaction energies are calculated, such as electrostatic interaction energy, dispersion force, and hydrogen bond interaction energy.

[0029] Preferably, the quantum chemical feature characterization unit adopts the B3LYP / 6-31G(d) calculation level to perform full quantum chemical calculations on small molecules and functional groups, and semi-empirical methods or molecular fragment methods are used for approximate calculations on large molecules. Taking ethanol molecule as an example, the HOMO energy level calculated by this unit is -6.52 eV, the LUMO energy level is -0.85 eV, the energy gap is 5.67 eV, the dipole moment is 1.69 D, the number of hydrogen bond donors is 1, and the number of hydrogen bond acceptors is 1. These quantum features reflect the properties of the molecule at the electronic level and are of great significance for predicting intermolecular interactions and specific physical properties (such as polarity-related properties).

[0030] The molecular structure feature characterization unit is communicatively connected to the quantum chemical feature characterization unit and is used to construct features from the molecular scale. This unit first extracts molecular topological structure features, represents the molecule as an atom-bond connection graph, and calculates topological indices such as Wiener index and Balaban index. Secondly, molecular geometric structure features are constructed to characterize the arrangement of atoms in three-dimensional space, and molecular shape descriptors such as sphericity and ellipticity, as well as surface area and volume, are calculated. Thirdly, functional groups are identified and characterized, such as hydroxyl group, carboxyl group, ester group, etc., and the occurrence times and positions of various functional groups are counted. Finally, molecular fingerprints are generated, such as extended connectivity fingerprint (ECFP), MACCS key fingerprint, etc., for molecular similarity comparison.

[0031] In one embodiment of the present invention, for n - hexane molecules, the features extracted by the molecular structure feature characterization unit include: the number of carbon atoms is 6, the number of hydrogen atoms is 14, the number of single bonds is 19, the number of rings is 0, the number of rotatable bonds is 5, the aromaticity is 0, the topological polar surface area is 0 Ų, the Wiener index is 36, etc. These structural features intuitively reflect the molecular backbone structure and connection mode, and play an important role in predicting physical properties such as boiling point and melting point.

[0032] The mesoscopic scale feature characterization unit is communicatively connected to the molecular structure feature characterization unit and is used to construct features from the mesoscopic scale. This unit first characterizes the molecular aggregate features, describes the packing and arrangement of molecules at the mesoscopic scale, and predicts the self - assembly tendency and interfacial properties. Secondly, it extracts the phase behavior features, predicts the possibility of a mixture forming multiple phases, and characterizes the critical parameters and phase diagram features. Thirdly, it constructs the statistical thermodynamics features, describes the molecular spatial distribution law, velocity distribution, and correlation function features.

[0033] Preferably, the mesoscopic scale feature characterization unit uses molecular dynamics simulation and Monte Carlo simulation methods to generate the molecular aggregate structure, and then extracts statistical mechanics features such as the radial distribution function and structure factor. For example, for the n - hexane - benzene mixture, this unit can simulate the phase behavior of the two components at different temperatures and concentrations, calculate the mixing enthalpy and excess volume, and predict the possible phase separation conditions. These mesoscopic features are of great significance for predicting phase equilibrium properties, transport properties, etc.

[0034] The macroscopic thermodynamics feature characterization unit is communicatively connected to the mesoscopic scale feature characterization unit and is used to construct features from the macroscopic scale. This unit first constructs the state equation parameter features, including critical parameters (temperature, pressure, volume), acentric factor, polarity factor, etc., which are parameter sets for describing the PVT relationship. Secondly, it extracts the thermodynamic property features, such as phase change parameters (vaporization heat, fusion heat, etc.), entropy capacity parameters, and free energy parameters. Thirdly, it characterizes the transport property features, such as viscosity parameters, thermal conductivity features, and diffusion coefficient features. Finally, it constructs the mixing rule parameters, such as binary interaction parameters, activity coefficient model parameters, etc.

[0035] In one embodiment of the present invention, the state equation parameter features constructed by the macroscopic thermodynamics feature characterization unit for n - hexane include: critical temperature Tc = 507.6 K, critical pressure Pc = 3.025 MPa, critical volume Vc = 368 cm³ / mol, and acentric factor ω = 0.301. These macroscopic features have direct significance for state equation calculation and physical property prediction.

[0036] The multi-scale feature fusion unit is communicatively connected to each of the above-mentioned feature representation units and is used to achieve effective fusion of features at different scales. This unit first conducts feature correlation analysis, identifies the relationships between features at different scales, eliminates redundant information, and retains complementary information. Secondly, it performs feature selection and dimensionality reduction, screening key features using filter-based, wrapper-based, or embedded feature selection methods, and reducing the feature dimension through methods such as principal component analysis (PCA) and t-SNE. Thirdly, it conducts feature transformation and combination, normalizes, standardizes, and non-linearly transforms the features to construct high-order features. Finally, it realizes feature adaptive weight assignment, dynamically adjusting the weights of features at different scales according to the prediction task and physical property type.

[0037] Preferably, the multi-scale feature fusion unit adopts a feature fusion method based on the attention mechanism, assigning different weights to features at different scales. For example, for predicting density, the macro-thermodynamic features may obtain a weight of 0.5, the molecular structure features obtain a weight of 0.3, and the quantum chemical features obtain a weight of 0.2; while for predicting properties related to polarity (such as dielectric constant), the weight of the quantum chemical features may increase significantly. This adaptive weight assignment mechanism enables the system to flexibly adjust the importance of features according to the physical nature of different physical properties, improving the prediction accuracy.

[0038] Through the above four levels of feature representation and fusion, the multi-scale feature representation module 2 can comprehensively capture the structural and property features of petrochemical materials from the microscale to the macroscale, providing rich and multi-level feature representations for subsequent physical property prediction.

[0039] The physical constraint modeling module 3 includes a thermodynamic constraint embedding unit, an equation of state constraint unit, a transport process constraint unit, and a physical constraint optimization and integration unit.

[0040] The thermodynamic constraint embedding unit is used to embed the basic laws of thermodynamics into the model structure. This unit first implements the energy conservation constraint of the first law of thermodynamics, ensuring that the change in internal energy of the system is equal to the algebraic sum of heat and work, which can be expressed as: , where: is the change in internal energy, with the unit of J; is the heat absorbed by the system, with the unit of J; is the work done by the system to the outside, with the unit of J. A positive value indicates that the system absorbs heat or does work to the outside, and a negative value indicates that the system releases heat or the outside does work on the system. Secondly, this unit implements the entropy increase principle constraint of the second law of thermodynamics, ensuring that the entropy of an isolated system always increases, that is: , where: is the total entropy change, with the unit of J / K, including the sum of the system entropy change and the environment entropy change. This constraint ensures that the direction of any spontaneous process conforms to the principle of entropy increase and is the basic constraint for the evolution of physical systems. Again, the unit embeds the phase equilibrium condition constraint to ensure that the temperature, pressure, and chemical potential of each phase in a multiphase system are equal: , , , where: is the temperature of the phase, with the unit of K; is the pressure of the phase, with the unit of Pa; is the chemical potential of component in the phase, with the unit of J / mol. This set of constraints is a necessary condition for a multiphase system to reach thermodynamic equilibrium. Finally, the unit verifies the thermodynamic consistency relations, such as Maxwell relations, etc., to ensure that the derivative relations of thermodynamic state functions are satisfied: , , where: is the temperature, with the unit of K; is the volume, with the unit of ; is the pressure, with the unit of Pa; is the entropy, with the unit of J / K; represents the partial derivative of with respect to under the condition that the variable remains unchanged. These relations are the direct result of the complete differential property satisfied by thermodynamic state functions. Preferably, the thermodynamic constraint embedding unit uses the Lagrange multiplier method to incorporate these constraints into the loss function, that is, modifying the original loss function to: , where: is the original loss function, representing the error between the predicted value and the true value; is the Lagrange multiplier, a non - negative scalar parameter that weighs the importance of the original loss and the constraints; is the constraint condition, representing the degree of violation of the physical constraint, which should ideally be zero; represents the summation over all constraint conditions . For example, for the phase equilibrium constraint, one can set , minimize the square of the chemical potential difference. The equation of state constraint unit is communicatively connected to the thermodynamic constraint embedding unit for embedding the equation of state constraint of the substance. This unit first embeds the cubic equation of state constraint, such as the Peng-Robinson equation: , where: is the pressure, with the unit of Pa; is the gas constant, with a value of 8.314 J / (mol-K); is the temperature, with the unit of K; is the molar volume, with the unit of m3 / mol; and are parameters related to intermolecular forces and molecular size, with the units of Pa-m and mol is the temperature correction function, dimensionless, and is usually expressed as: , where: is the reduced temperature, dimensionless; is the critical temperature, with the unit of K; is the parameter related to the acentric factor , dimensionless, and can be calculated by .

[0041] Secondly, this unit embeds the SAFT series of advanced equation of state constraints, such as the PC-SAFT equation, which can more accurately describe chain-like molecules and hydrogen bond systems. Thirdly, the mixing rule constraint is implemented to define the calculation method of the equation of state parameters for multi-component systems, such as the quadratic mixing rule: , , where: and are the equation of state parameters of the mixture, with the same units as and ; and are the mole fractions of components and , dimensionless; , , are pure component parameters; is the binary interaction parameter, dimensionless, which characterizes the degree of deviation of the interaction between components and from the geometric mean; and represent the summation over all components and Perform summation.

[0042] Finally, the unit performs corrections to the special condition state equations, such as critical region correction, high-pressure correction, and extremely low-temperature correction, etc., to improve the prediction accuracy under special conditions. For example, the critical region correction can adopt the renormalization group theory to correct the critical exponent and avoid the inaccuracy of the classical state equation near the critical point.

[0043] The transport process constraint unit is communicatively connected to the thermodynamic constraint embedding unit and is used to embed the physical constraints describing the mass transfer phenomenon. This unit first embeds the viscosity model constraints, such as the low-density gas viscosity theory constraint (based on the Chapman-Enskog theory) and the liquid viscosity empirical relationship constraint (such as the Andrade equation). Secondly, it embeds the heat conduction constraint to ensure that the heat conduction prediction conforms to Fourier's law. Thirdly, it embeds the diffusion coefficient constraint to implement the Fick diffusion law and the Maxwell-Stefan diffusion model constraint. Finally, it embeds the surface property constraints, such as the surface tension theory constraint and the surface tension temperature dependence constraint (Eötvös rule).

[0044] In an embodiment of the present invention, the transport process constraint unit uses a modified Andrade equation to constrain the liquid viscosity: , Where: is the liquid viscosity, with the unit of Pa·s; is the temperature, with the unit of K; , and are substance-related parameters, where is a dimensionless constant, has the unit of K, is a dimensionless constant. For n-hexane, the parameter values are approximately . This equation reflects the exponential relationship between the liquid viscosity and temperature and is an important constraint for transport properties.

[0045] The physical constraint optimization and integration unit is communicatively connected to the above-mentioned constraint units and is used to optimize and integrate various physical constraints. This unit first optimizes the weights of each physical constraint and adjusts the importance of the constraints according to different physical properties and prediction tasks. Secondly, it combines the hard and soft constraint mechanisms, sets the physical laws that must be strictly satisfied as hard constraints, and sets the empirical rules as soft constraints. Thirdly, it coordinates the multi-scale constraint relationships to ensure the consistency of physical constraints at different scales. Finally, it evaluates the constraint embedding effect and monitors the changes in the constraint satisfaction degree and prediction performance.

[0046] Preferably, the physical constraint optimization integration unit adopts an adaptive constraint weight adjustment strategy to dynamically adjust the weights according to the degree of constraint violation. For example, when the model seriously violates the law of conservation of energy, the weight of the corresponding constraint will automatically increase; when the constraint has been well satisfied, the weight can be appropriately reduced to avoid over-constraint affecting the flexibility of the model. In addition, this unit adopts a phased constraint implementation strategy, first training a model that satisfies the basic physical constraints, and then gradually adding more complex constraints to avoid optimization difficulties.

[0047] Through the collaborative work of the above units, the physical constraint modeling module 3 can ensure that the physical property prediction results conform to the basic physical laws and thermodynamic laws, avoid physically unreasonable prediction results, and improve the reliability and interpretability of the model.

[0048] The component interaction modeling module 4 includes a binary component interaction unit, a multi-component interaction unit, a component interaction characterization network unit, and a non-ideal mixing effect unit.

[0049] The binary component interaction unit is used to accurately capture the interaction between two components. This unit first calculates the binary interaction parameter and predicts the interaction strength between components based on methods such as molecular similarity, interaction energy, and solubility parameter. The binary interaction parameter can be estimated by the following equation: , where: is the binary interaction parameter, dimensionless, representing the degree of deviation of the interaction between components and from the geometric mean; and are the attractive parameters of the pure components, with the unit of Pa·m , related to the critical parameters. When , it means that the interaction between components follows the geometric mean rule; means that the interaction between components is weaker than the geometric mean; means that the interaction between components is stronger than the geometric mean. Secondly, this unit predicts the binary mixture excess properties, such as excess enthalpy , excess volume and excess Gibbs free energy and so on. These excess properties reflect the degree of deviation of the mixture from the ideal behavior. For example, the excess enthalpy can be expressed as: , where: is the excess enthalpy, with the unit of J / mol; and are the mole fractions of the two components, dimensionless, and satisfy ; is a fitting parameter with the unit of J / mol; is the polynomial order, usually taken as 2 - 4; denotes the summation from to This equation is a form of the Redlich - Kister expansion and is commonly used to characterize the excess properties of binary mixtures.

[0050] Again, this unit predicts the binary phase diagram, including liquid - liquid phase equilibrium, vapor - liquid phase equilibrium, and azeotrope prediction, etc. Finally, it simulates special binary interaction effects, such as hydrogen - bond effect, charge - transfer effect, acid - base interaction, and complexation, etc.

[0051] In one embodiment of the present invention, for the ethanol - water system, the excess enthalpy predicted by the binary component interaction unit is negative at 25 °C, indicating an exothermic mixing process, which is consistent with the fact that hydrogen bonds are formed between the two components; while for the hexane - benzene system, the predicted excess enthalpy is positive, indicating an endothermic mixing process, which is due to the weakened intermolecular interaction caused by the difference in the molecular structures of the two molecules.

[0052] The multi - component interaction unit is communicatively connected to the binary component interaction unit and is used to handle the complex interactions among three or more components. This unit first identifies the higher - order interaction terms, determines whether there are interactions specific to three or more components, and how to evaluate the importance of these higher - order terms. The higher - order interaction terms can be expressed as follows: , where: is the excess Gibbs free energy with the unit of J / mol; , and are the mole fractions of components , and , dimensionless; is the binary interaction energy parameter with the unit of J / mol, characterizing the interaction between components and ; is the ternary interaction energy parameter with the unit of J / mol, characterizing the synergistic effect among components , and ; , and denote the summation over all components. The first term on the right - hand side of the equation represents all possible binary interactions, the second term represents all possible ternary interactions, and so on.

[0053] Secondly, the unit optimizes the multi-component mixing rules, generalizes them from binary parameters to multi-component systems, and considers the influence of concentration on the mixing rules. Thirdly, it predicts the multi-component phase behavior, identifies multi-phase regions, critical endpoints, and azeotropic multi-components. Finally, it simulates the multi-component dynamic interaction and predicts the influence of changes in component concentration, temperature, and pressure on the interaction.

[0054] Preferably, the multi-component interaction unit adopts a hierarchical method to handle multi-component interactions: first, it considers all possible binary interactions, and then evaluates whether higher-order interaction terms need to be introduced. For most systems, binary interactions can provide a sufficiently accurate description; however, for some special systems (such as mixtures of strongly polar components), introducing ternary interaction terms can significantly improve the prediction accuracy.

[0055] The component interaction characterization network unit is communicatively connected to the multi-component interaction unit and is used to capture the complex interactions between components through a neural network. This unit first constructs a component graph representation learning framework, represents the mixture as a network of component nodes, extracts the local interactions between components through graph convolution, and focuses on important component interactions through a graph attention mechanism. Secondly, it implements a concentration-aware interaction network, adjusts the interaction intensity according to the component concentration, and handles the high-concentration and low-concentration limit cases. Thirdly, it constructs a condition-sensitive interaction model to capture the influence of conditions such as temperature and pressure on component interactions. Finally, it captures long-range interactions and breaks through the limitations of local interactions through an attention mechanism and global information aggregation.

[0056] In one embodiment of the present invention, the component interaction characterization network unit uses a graph attention network (GAT) to construct a component interaction model, where the attention coefficient is calculated as follows: , where: is the attention coefficient, dimensionless, representing the importance of the influence of component j on component i; and are the component node feature vectors, with a dimension of , d is the feature dimension; W is the weight matrix, with a dimension of , is the transformed feature dimension; a is the attention vector, with a dimension of ; represents the vector concatenation operation; is the neighbor set of component i; exp represents the exponential function; LeakyReLU represents the leaky rectified linear unit activation function, defined as LeakyReLU(x)=max(0.01x,x); Denotes the sum over all neighbors k of component i. In this way, the model can automatically learn the contribution magnitudes of different components to the interaction. The non-ideal mixing effect unit A4 is communicatively connected to the component interaction characterization network unit A3 for accurately describing the mixing effect deviating from the ideal behavior. This unit first implements activity coefficient models such as the NRTL, UNIQUAC, and Wilson equations, etc., to predict the non-ideal behavior of components in the mixture. For example, the expression for calculating the activity coefficient by the NRTL equation is: , where: is the activity coefficient of component , dimensionless; is the mole fraction of component , dimensionless; is the energy parameter, dimensionless, usually expressed as , where is the interaction energy parameter between component and , with the unit of J / mol; is the NRTL model parameter, dimensionless, usually expressed as , where is the non-randomness parameter, dimensionless, usually between 0.1 - 0.5; , and denote the sum over all components. This equation is based on the local composition theory and takes into account the differences between the local environment around the molecules and the overall environment.

[0057] Secondly, this unit comprehensively models the excess properties, captures the correlations between different excess properties, and predicts the effects of temperature and pressure on the excess properties. Thirdly, it deals with special non-ideal systems such as strong electrolyte systems, polymer solutions, and supercritical fluid mixtures, etc. Finally, it predicts the phase separation phenomenon, realizes the thermodynamic stability criterion, calculates the boundaries of the two-liquid phase region and the critical solution temperature.

[0058] Preferably, the non-ideal mixing effect unit automatically selects the most suitable activity coefficient model according to the mixture characteristics. For example, for a completely miscible system, the Wilson equation usually gives good results; for a partially miscible system, the NRTL and UNIQUAC models are more suitable; for a polymer solution, the Flory-Huggins model may be needed. In addition, this unit also implements the hybrid model method, combining the advantages of different activity models to improve the applicable range and accuracy of the prediction.

[0059] Through the collaborative work of the above units, the component interaction modeling module 4 can accurately capture the complex interactions in the petrochemical material mixture, predict the non-ideal mixing effect, and provide key support for physical property prediction.

[0060] The uncertainty quantification module 5 includes a prediction uncertainty source analysis unit, a Bayesian uncertainty quantification unit, an uncertainty propagation and aggregation unit, and an uncertainty calibration and evaluation unit.

[0061] The prediction uncertainty source analysis unit is used to identify and analyze the sources of uncertainty. This unit first evaluates data uncertainty, including measurement errors of experimental data, uncertainties caused by data sparsity, and data noise, etc. The measurement uncertainty of experimental data can be obtained from the documentation of the data source or estimated based on the standard deviation of repeated measurements. Secondly, it analyzes model uncertainty, including uncertainties caused by model structure selection, parameter estimation uncertainties, and the risks of overfitting or underfitting brought by model complexity. Thirdly, it quantifies physical approximation uncertainty, evaluating the uncertainties introduced by factors such as state equation approximation, simplification of mixing rules, and quantum computing precision limitations. Finally, it evaluates prediction condition extrapolation uncertainty, analyzing the reliability when the prediction conditions (such as temperature, pressure, composition) exceed the range of training data.

[0062] In an embodiment of the present invention, for liquid density prediction, the prediction uncertainty source analysis unit finds that the data uncertainty at normal temperature and pressure is about ±0.5%, mainly due to experimental measurement errors; while under high temperature and high pressure conditions, the uncertainty may increase to more than ±2%, which is mainly caused by data sparsity and model extrapolation.

[0063] The Bayesian uncertainty quantification unit is communicatively connected to the prediction uncertainty source analysis unit and is used to achieve uncertainty quantification through the Bayesian method. This unit first implements a Bayesian neural network, replaces the deterministic weights with probability distributions, and characterizes parameter uncertainty through the posterior distribution. In the Bayesian neural network, the posterior distribution of the prediction y can be expressed as: , where: is the given input and the training data under which the posterior probability distribution of the output ; is the input feature vector; is the training data set; is the model parameter vector; is the parameter posterior distribution, representing the probability distribution of the parameters given the training data; is the prediction distribution given the parameters, usually a Gaussian distribution; Denotes the integration over all possible parameter spaces. This integration is usually not analytically computable and needs to be approximated by numerical methods such as variational inference or Monte Carlo sampling. Secondly, this unit implements the Monte Carlo Dropout technique, by keeping Dropout on during the prediction phase and making multiple random predictions to obtain the prediction distribution. The prediction mean and variance can be estimated as follows: , , where: is the prediction mean; Var is the prediction variance; is the model prediction for the th Dropout sample, i.e., the result of the forward propagation using the network with randomly masked neurons; is the intrinsic variance of the model prediction, i.e., the assumed variance of the observation noise; is the number of samples, usually set to 50 - 100; denotes taking the average of the sample results.

[0064] Thirdly, this unit constructs a Bayesian ensemble learning, combining the predictions of multiple Bayesian models to further improve the reliability of uncertainty estimation. Finally, a deep Gaussian process is implemented, constructing a kernel function through a deep network to handle Gaussian process regression for large-scale datasets.

[0065] Preferably, the Bayesian uncertainty quantification unit adopts different uncertainty quantification strategies for different types of physical properties. For example, for physical properties with rich data (such as density at normal temperature and pressure), a simple Monte Carlo Dropout method can be used; while for physical properties with sparse data or high nonlinearity (such as properties near the critical point), a complete Bayesian neural network or deep Gaussian process method may be required.

[0066] The uncertainty propagation and aggregation unit is communicatively connected to the Bayesian uncertainty quantification unit and is used to track the propagation of uncertainty during the prediction process. This unit first analyzes the uncertainty propagation chain, tracks how the uncertainty of input variables propagates to the prediction results, evaluates the uncertainty of intermediate features, and analyzes the accumulation of uncertainty in multi-step predictions. Secondly, it aggregates multi-source uncertainties, comprehensively considering various uncertainty sources such as data, models, and physical approximations, and identifies the main uncertainty contributions. Thirdly, it analyzes the conditional dependence uncertainty, studying the influence of conditions such as temperature, pressure, and composition on the prediction uncertainty. Finally, it handles extreme condition uncertainties, coping with high uncertainties in special cases such as near the critical point, phase transition regions, and trace components.

[0067] In one embodiment of the present invention, the uncertainty propagation and aggregation unit discovers that for the viscosity prediction of hydrocarbon mixtures, when the temperature approaches the critical point, the uncertainty increases significantly, and the prediction relative error may increase from ±5% under normal conditions to more than ±20%. This is mainly because the physical behavior in the critical region becomes highly non-linear and sensitive, and the generalization ability of the model is challenged in this region.

[0068] The uncertainty calibration and evaluation unit is communicatively connected to the uncertainty propagation and aggregation unit to ensure the accuracy of uncertainty estimation. This unit first implements uncertainty calibration methods, such as frequency calibration and temperature scaling, to ensure that the prediction interval coverage rate is consistent with the confidence level. The goal of frequency calibration is to make: , where: represents probability; is the true value; is the predicted mean; is the predicted standard deviation; is the calibration factor, which is the quantile of the standard normal distribution; is the significance level, usually set to 0.05, corresponding to a 95% confidence interval. The goal of calibration is to ensure that the actual coverage rate (i.e., the proportion of true values falling within the prediction interval) is equal to the nominal coverage rate (i.e., 1 - .

[0069] Secondly, this unit calculates reliability evaluation indicators, such as prediction interval coverage rate, negative log-likelihood, and calibration error, etc., to comprehensively evaluate the quality of uncertainty estimation. Thirdly, uncertainty visualization is realized, and the uncertainty distribution is intuitively displayed through prediction distribution diagrams, uncertainty heat maps, and confidence interval band diagrams, etc. Finally, abnormal predictions are detected, high-uncertainty predictions are marked, abnormal distribution shapes are found, and uncertainty mutations are analyzed.

[0070] Preferably, the uncertainty calibration and evaluation unit adopts a hierarchical calibration strategy to calibrate different condition intervals separately to improve the accuracy of calibration. For example, the temperature range can be divided into three intervals: low temperature, medium temperature, and high temperature, and the uncertainty estimation is calibrated separately for each interval. In addition, this unit also implements an adaptive re-calibration mechanism. As new data accumulates, the calibration parameters are updated regularly to ensure that the uncertainty estimation always remains accurate.

[0071] Through the collaborative work of the above units, the uncertainty quantification module 5 can provide a reliable prediction uncertainty assessment, quantify the credibility of the prediction results, and support risk-based decision-making.

[0072] The physical property prediction and analysis module 6 includes a thermodynamic property prediction unit, a transport property prediction unit, a combustion property prediction unit, and a structure-property correlation analysis unit.

[0073] The thermodynamic property prediction unit is used to predict various thermodynamic properties. This unit first predicts state properties, including basic thermodynamic properties such as density, specific heat capacity, enthalpy, and entropy. For example, the liquid density prediction can be achieved through the modified Rackett equation: , where: is the liquid density, with the unit of kg / m 3 is the critical density, with the unit of kg / m 3 is the critical compression factor, dimensionless, defined as , where , and are the critical pressure, critical molar volume, and critical temperature respectively; is the reduced temperature, dimensionless, defined as , where is the system temperature; the superscript 2 / 7 represents the power exponent. This equation is based on the critical point scaling principle and can accurately predict the liquid density from low temperature to near the critical point. Secondly, this unit predicts phase equilibrium properties, including vapor pressure, phase equilibrium composition, and critical parameters, etc. For example, the vapor pressure of a pure component can be predicted through the Antoine equation: , where: is the saturated vapor pressure, with the unit of Pa; is the temperature, with the unit of K; , and are substance-related constants, where is a dimensionless constant, has the unit of K, has the unit of K. This equation is an empirical relationship and can accurately predict the vapor pressure of pure components within a limited temperature range.

[0074] Again, this unit calculates thermodynamic derivative properties, such as expansion coefficient, compression coefficient, and Joule-Thomson coefficient, etc. Finally, it predicts mixture properties, such as mixture enthalpy, excess volume, and activity coefficient, etc.

[0075] In an embodiment of the present invention, for the liquid density prediction of C1-C10 normal alkanes in the temperature range of 283.15 - 373.15K and the pressure range of 0.1 - 10MPa, the average relative error is less than 0.5%, which is better than the traditional equation of state method (usually 1% - 3%). For complex mixtures, such as crude oil, this unit can predict the density based on the component composition, and the prediction error is controlled within 1%.

[0076] The transport property prediction unit is communicatively connected to the thermodynamic property prediction unit and is used to predict various transport properties. This unit first predicts viscosity, including gas viscosity, liquid viscosity, high-pressure viscosity, etc. For example, the prediction of liquid viscosity can be achieved through the free volume theory model: , where: is the liquid viscosity, with the unit of Pa·s; is the reference viscosity, with the unit of Pa·s, usually taking the viscosity value at the limit low density, is the substance-related constant, with the unit of m3 / mol; is the free volume, with the unit of m3 / mol, related to temperature and density, and can be expressed as , where is the actual molar volume, is the molar volume at close packing. This model is based on the free volume theory of molecular motion and can describe the variation of liquid viscosity with temperature and pressure.

[0077] Secondly, this unit predicts the heat conduction property and estimates the thermal conductivity of gases and liquids. Thirdly, it calculates the diffusion coefficient and predicts the self-diffusion coefficient and the mutual-diffusion coefficient. Finally, it predicts the surface properties, such as surface tension and interfacial tension, etc. Preferably, the transport property prediction unit uses a multi-scale fusion method to predict viscosity, combining molecular dynamics simulation and empirical correlations to improve the prediction accuracy. For hydrocarbon mixtures, the average relative error of viscosity prediction by this unit in the temperature range of 293.15 - 473.15 K is 3% - 5%, significantly better than the traditional corresponding states method (usually 10% - 15%). The combustion property prediction unit is communicatively connected to the thermodynamic property prediction unit and is used to predict the properties related to combustion. This unit first predicts the fuel evaluation indices, including octane number, cetane number, methane number, etc. For example, the research octane number (RON) of gasoline can be predicted from the component composition: , where: RON is the research octane number of the mixture, dimensionless; and are the volume fractions of components and , dimensionless; is the octane number of pure component , dimensionless; is the binary interaction term of components and , dimensionless, characterizing the non-linear effect of the mixing of two components on the octane number; and Denotes the summation over all components. This equation takes into account the linear contributions of the components and the binary interaction effects, and can accurately predict the octane number of the blended fuel.

[0078] Secondly, this unit predicts fire safety properties such as flash point, ignition point, and explosion limits. Thirdly, it calculates the combustion thermodynamic properties such as heat of combustion and adiabatic flame temperature. Finally, it predicts the combustion kinetic parameters such as ignition delay time and laminar flame speed.

[0079] In an embodiment of the present invention, the average absolute error of the Research Octane Number (RON) prediction of gasoline components by the combustion property prediction unit is 1.5 - 2.0 units, and the average absolute error of the cetane number prediction of diesel is 1.0 - 1.5 units. This accuracy meets the requirements of most industrial applications.

[0080] The structure-property correlation analysis unit is communicatively connected to the above-mentioned prediction units and is used to analyze the relationship between molecular structure and physical properties. This unit first identifies the key structural features and finds the key structural fragments and functional groups that affect the physical properties. Secondly, it predicts the structure modification effects and analyzes the substituent effects, homolog variation trends, and position isomer effects, etc. Thirdly, it analyzes the component interaction mechanism, identifies the synergistic and antagonistic effects, and constructs the interaction network between components. Finally, it analyzes the driving factors of physical property changes, quantifies the influence of each factor on the physical properties, and plots the response curve of the physical properties to the conditions.

[0081] Preferably, the structure-property correlation analysis unit uses interpretable artificial intelligence techniques such as SHAP value analysis to quantify the contributions of different molecular features to the physical property prediction. For example, for the boiling point prediction of alkanes, this unit finds that the number of carbon atoms is the most important factor (contributing about 75%), the degree of molecular branching is the second most important factor (contributing about 15%), and the presence of a cyclic structure is the third most important factor (contributing about 5%). This kind of analysis not only improves the interpretability of the prediction but also provides valuable guidance for molecular design.

[0082] Through the collaborative work of the above units, the physical property prediction and analysis module 6 can accurately predict various physical and chemical properties of petrochemical materials, analyze the relationship between molecular structure and physical properties, and provide a scientific basis for process optimization and product development.

[0083] In a preferred embodiment of the present invention, the system further includes a model training and optimization module 7, which includes a multi-task joint learning unit, a physical constraint optimization unit, a model architecture optimization unit, a hyperparameter optimization unit, and a model evaluation and selection unit.

[0084] The multi-task joint learning unit is used to achieve the joint prediction of multiple physical properties. First, the unit designs the physical property correlation structure, establishes the dependency relationships between physical properties, and designs the shared feature layer and physical property specific layers. For example, physical properties such as density, viscosity, and surface tension can share the low-level feature representations because they are all closely related to intermolecular forces. Second, the unit designs the multi-task loss function, assigns weights according to the importance and prediction difficulty of physical properties, and takes into account data uncertainty and consistency between physical properties. The multi-task loss function can be expressed as: , where: is the total multi-task loss function; is the loss function of task , usually the mean square error or mean absolute error; is the weight of task , dimensionless, reflecting the importance of the task; is the consistency constraint between task and , characterizing the physical relationship that the prediction results of the two tasks should satisfy; is the constraint weight, dimensionless; denotes the summation over all tasks ; denotes the summation over all task pairs .

[0085] Third, the unit implements the gradient balance strategy, normalizes the gradients of different tasks, detects and resolves gradient conflicts, and dynamically adjusts the task priorities. Finally, it implements the transfer and incremental learning strategy, using the learned physical property knowledge to assist in the learning of new physical properties and avoiding catastrophic forgetting.

[0086] Preferably, the multi-task joint learning unit adopts the soft parameter sharing method. The models of different tasks have their own parameters, but there are regularization constraints between the parameters to promote knowledge sharing between tasks. For closely related physical properties (such as density and viscosity), the degree of parameter sharing is high; while for less related physical properties (such as octane number and surface tension), the degree of parameter sharing is low. This flexible sharing mechanism enables the model to make full use of the correlations between tasks while maintaining sufficient flexibility.

[0087] The physical constraint optimization unit is used to optimize the implementation of physical constraints in the model. This unit first designs a constraint weighted loss function to balance the data-driven loss and the physical constraint loss, and adjusts the penalty intensity according to the degree of violation. Secondly, it implements the Lagrange multiplier method to handle equality constraints and inequality constraints, and designs an efficient multiplier update strategy. Thirdly, it designs physically-guided regularization to constrain the model using prior knowledge to ensure physical invariance and gradually increase the constraint intensity. Finally, it constructs a constraint satisfaction feedback mechanism to quantify the constraint satisfaction situation, dynamically adjust the constraint priority, and identify regions where constraints are difficult to satisfy.

[0088] In one embodiment of the present invention, for the energy conservation constraint of the first law of thermodynamics, the physical constraint optimization unit adopts a soft constraint method and adds the degree of constraint violation to the loss function: , where: is the modified total loss function; is the data-driven loss, usually the mean square error between the predicted value and the true value; is the constraint weight, dimensionless, controlling the importance of the constraint; MSE is the mean square error function, defined as MSE , where is the number of samples; 、 and are the change in internal energy, heat, and work respectively. The constraint weight is initially set to 0.1 and gradually increases to 1.0 as the training progresses to avoid over-constraining at the beginning of training, which may lead to optimization difficulties.

[0089] The model architecture optimization unit is used to optimize the architecture design of the neural network. This unit first optimizes the graph neural network, improves the graph convolutional layer, graph attention mechanism, and graph pooling strategy, and optimizes the component interaction representation. Secondly, it designs a multi-scale feature fusion network, constructs a cross-scale feature aggregation mechanism, optimizes the scale-specific sub-network, and realizes adaptive feature weighting. Thirdly, it constructs a physical embedding neural network architecture, designs a physical equation layer, a state variable transformation layer, and a physical consistency check layer. Finally, it implements adaptive architecture search to automatically discover the optimal network structure, balance the model complexity and performance, and optimize the architecture for different prediction tasks.

[0090] Preferably, the model architecture optimization unit adopts the neural architecture search (NAS) technology to automatically find the optimal network structure in a predefined search space. The search space includes hyperparameters such as the number of layers, layer width, activation function, connection method, and attention mechanism. Through an efficiency-optimized search algorithm (such as a gradient-based method or an evolutionary algorithm), this unit can find a network architecture that balances performance and efficiency with reasonable computing resources.

[0091] The hyperparameter optimization unit is used to optimize the hyperparameter settings of the model. This unit first implements automatic hyperparameter search, finding the optimal hyperparameters through grid search, random search, Bayesian optimization, or evolutionary algorithms. Secondly, it optimizes the learning rate strategy, designs a learning rate scheduling mechanism, implements an adaptive learning rate method, and adjusts the learning rate for constraint violations. Thirdly, it optimizes the regularization parameters, adjusting the L1 / L2 regularization strength, Dropout rate, and early stopping strategy parameters. Finally, it optimizes the model ensemble parameters, determines the number of ensemble models, controls model diversity, optimizes the ensemble weights, and selects the best ensemble method.

[0092] In one embodiment of the present invention, the hyperparameter optimization unit uses the Bayesian optimization method to search for the optimal hyperparameters, defining the search space as: the learning rate is in the range of [1e -5 , 1e -2 , the batch size is in the range of [16, 256], the L2 regularization coefficient is in the range of [1e -6 , 1e -3 , and the Dropout rate is in the range of [0.1, 0.5]. By optimizing the performance of the validation set, this unit found the optimal hyperparameter combination: the learning rate is 3e -4 , the batch size is 64, the L2 regularization coefficient is 5e -5 , and the Dropout rate is 0.2.

[0093] The model evaluation and selection unit is used to evaluate the model performance and select the optimal model. This unit first implements comprehensive multi-index evaluation, calculating prediction accuracy metrics (such as MAE, RMSE, R²), physical consistency evaluation metrics, and uncertainty quality metrics. Secondly, it designs a cross-validation strategy, implementing k-fold cross-validation, holdout validation, time series cross-validation, and stratified cross-validation. Thirdly, it conducts model comparison and selection, designs a model ranking mechanism, finds the Pareto optimal solution, conducts a statistical significance test, and introduces a model complexity penalty. Finally, it evaluates the extrapolation ability, testing the conditional extrapolation ability, component extrapolation ability, extreme condition prediction ability, and long-term prediction stability.

[0094] Preferably, the model evaluation and selection unit adopts a multi-objective evaluation framework, considering three objectives: prediction accuracy, physical consistency, and computational efficiency. For prediction accuracy, it uses the root mean square error (RMSE) and the coefficient of determination (R²); for physical consistency, it uses the degree of violation of thermodynamic constraints; for computational efficiency, it uses the inference time and the number of model parameters. Through Pareto front analysis, this unit can find a balanced model that performs well in all three objectives, rather than a model that only performs excellently in a single metric.

[0095] Through the collaborative work of the above-mentioned units, the model training and optimization module 7 can achieve multi-task joint learning, optimize the implementation of physical constraints, optimize the model architecture and hyperparameters, and evaluate the model performance and select the optimal model, thereby improving the prediction accuracy and reliability of the system.

[0096] In a preferred embodiment of the present invention, the system further includes a system integration and application module 8, which includes a user interface unit, an industrial integration unit, a data exchange and standardization unit, a deployment and expansion unit, and an application scenario adaptation unit.

[0097] The user interface unit is used to provide a user-friendly interaction interface. First, it designs a Web application interface, adopts a responsive layout, constructs a physical property prediction form, realizes result visualization, and manages user permissions. Second, it provides a command-line interface, supports batch commands and scripted operations, designs advanced query syntax, and customizes the output format. Third, it develops a programming API, designs a RESTful API, provides a Python client library and an SDK development package, and realizes a Web Socket real-time interface. Finally, it constructs a mobile application interface, optimizes the mobile device experience, supports offline prediction, and realizes data synchronization and push notifications.

[0098] Preferably, the user interface unit adopts a front-end and back-end separation architecture. The front end uses the React framework to build a responsive Web interface, and the back end provides RESTful API services. The user interface provides three main function modes: (1) single-point prediction mode, where the user inputs specific substances and conditions, and the system returns the prediction results and uncertainties; (2) batch prediction mode, where the user uploads a file containing multiple substances and conditions, and the system performs batch prediction and returns the results; (3) parameter optimization mode, where the user sets the target physical property range, and the system recommends the best component ratio that meets the conditions.

[0099] The industrial integration unit is used to achieve integration with industrial systems. First, it realizes the integration of process simulation software, develops interfaces for Aspen Plus, ProII / Petro-SIM, and HYSYS, and constructs a gPROMS dynamic model interface. Second, it provides an industrial control system interface, realizes an OPC UA server / client, integrates DCS systems and SCADA systems, and docks with MES systems. Third, it integrates the laboratory information system, designs a LIMS interface, supports automatic import of experimental data, generates analysis reports, and tracks sample information. Finally, it docks with the enterprise resource planning system, realizes SAP / Oracle integration, connects the product data management and supply chain management systems, and integrates the quality management system.

[0100] In one embodiment of the present invention, the industrial integration unit has developed an Aspen Plus physical property package plug-in, which allows users to directly call the physical property prediction function of this system in Aspen Plus. The plug-in is implemented through Aspen's physical property extension interface and supports the calculation of more than 15 physical property parameters, including density, vapor pressure, enthalpy, entropy, heat capacity, viscosity, etc. Integration tests have shown that after using this plug-in, the convergence of Aspen Plus in dealing with complex mixtures has been significantly improved, especially for highly non-ideal mixtures and supercritical conditions.

[0101] The data exchange and standardization unit is used to achieve the standardized exchange of data. Firstly, this unit supports multiple data exchange formats, including JSON / XML, CSV / Excel, HDF5 big data format, and CDF scientific data format. Secondly, it implements industry standard interfaces, supporting the CAPE-OPEN standard, OPCUA information model, ASTM data standard, and ISA-95 hierarchy. Thirdly, it provides data conversion and mapping functions, realizing unit conversion, data mode mapping, encoding conversion, and coordinate system conversion. Finally, it ensures secure data transmission through data encryption, security authentication, access logs, and data integrity verification.

[0102] Preferably, the data exchange and standardization unit has implemented a thermodynamic interface based on the CAPE-OPEN standard, enabling the system to be seamlessly integrated with any process simulation software that supports this standard. The CAPE-OPEN interface provides standardized thermodynamic calculation services, including physical property calculation, phase equilibrium calculation, and thermodynamic derivative calculation, etc. In addition, this unit also supports common physical property data exchange formats, such as DIPPRXML format and ThermoML format, facilitating the exchange of data with other physical property databases and systems.

[0103] The deployment and extension unit is used to handle the deployment and extension of the system. Firstly, this unit supports multi-environment deployment, providing cloud deployment solutions, edge computing deployment solutions, local server deployment solutions, and hybrid deployment strategies. Secondly, it implements containerization and microservices architecture, encapsulating Docker containers, using Kubernetes for orchestration, designing microservice decomposition, and building service meshes. Thirdly, it optimizes performance and extension, implements load balancing, supports horizontal scaling, optimizes caching, and realizes distributed computing. Finally, it manages version control and upgrades, tracks system versions, realizes smooth upgrades, designs rollback plans, and manages configurations.

[0104] In an embodiment of the present invention, the deployment and expansion unit adopts a microservices architecture, decomposing the system into multiple independent services such as data processing service, feature extraction service, model inference service, and result analysis service. Each service is encapsulated in a Docker container and orchestrated and managed through Kubernetes. This architecture enables the system to have high scalability and elasticity, capable of automatically scaling service instances according to the load and quickly recovering in case of service failures. Tests show that under high concurrency (100 requests per second), the system response time can still be maintained within 200 milliseconds, meeting the requirements of real-time applications.

[0105] The application scenario adaptation unit is used to optimize for different application scenarios. This unit first adapts to the oil refining application, develops crude oil evaluation tools, predicts fraction properties, optimizes the blending ratio of mixed crude oil, and predicts product quality. Secondly, it adapts to the chemical process application, optimizes the separation process, simulates the reaction process, designs the heat exchange network, and optimizes the process conditions. Thirdly, it adapts to new material development, provides solvent screening tools, predicts polymer properties, designs composite materials, and optimizes material performance. Finally, it supports environmental and safety assessments, evaluates environmental impacts, analyzes safety risks, predicts emission characteristics, and supports emergency responses.

[0106] Preferably, the application scenario adaptation unit has developed a crude oil evaluation and formulation optimization function for the oil refining application. This function allows users to input the basic properties (such as API gravity, sulfur content, etc.) and available amounts of multiple crude oils, and the system automatically recommends the optimal blending ratio, enabling the properties of the mixed crude oil to meet the requirements of the refinery while minimizing costs. Tests show that using this function can reduce the crude oil procurement cost by 2% - 3% while ensuring that the product quality meets the specification requirements. In addition, this unit also provides a formulation optimization function for the petrochemical product development scenario, which can recommend the optimal component ratio according to the target physical properties (such as octane number, viscosity, etc.), accelerating the product development process.

[0107] Through the collaborative work of the above units, the system integration and application module 8 can provide a user-friendly interaction interface, realize the integration with industrial systems, execute data exchange and standardization, support system deployment and expansion, and adapt to different application scenarios, improving the practicability and usability of the system.

[0108] The present invention also provides a physical property calculation method for petrochemical material mixing modeling with physical information constraints, including the following steps: Step S1: Collect multi-source data of petrochemical materials, and perform cleaning, quality control, and standardization processing on the multi-source data to obtain data in a standard format; Specifically, this step first collects petrochemical material data from multiple sources, including experimental databases, theoretical calculation results, and industrial process data. Then, the collected raw data is cleaned and quality-controlled, including detecting and processing outliers, verifying the thermodynamic consistency of the data, and performing quality scoring on the data. Next, the cleaned data is integrated into a standard format, including unifying the physical property data format, standardizing the molecular representation, and converting the physical property units. Finally, a petrochemical material knowledge graph is constructed based on the standardized data, including extracting substance entities and property relationships, constructing substance-property-condition triples, and realizing knowledge reasoning and completion.

[0109] Step S2: Based on the data in the standard format, construct a multi-level petrochemical material feature characterization from the quantum scale to the macroscopic scale, and realize the fusion of features at different scales; Specifically, this step first constructs molecular features from the quantum scale, including extracting molecular electronic structure features, calculating molecular dipole moments and multipole moments, and characterizing intermolecular interactions. Then, features are constructed from the molecular scale, including extracting molecular topological structure features, constructing molecular geometric structure features, identifying and characterizing functional groups, and generating molecular fingerprints. Next, features are constructed from the mesoscopic scale, including characterizing molecular aggregate features, extracting phase behavior features, and constructing statistical thermodynamics features. Again, features are constructed from the macroscopic scale, including constructing equation of state parameter features, extracting thermodynamic property features, characterizing transport property features, and constructing mixing rule parameters. Finally, the effective fusion of features at different scales is realized, including feature correlation analysis, feature selection and dimensionality reduction, feature transformation and combination, and feature adaptive weight assignment.

[0110] Step S3: Embed the constraints of the laws of thermodynamics and physical laws into the model structure to ensure that the predicted results of physical properties meet the requirements of thermodynamic consistency; Specifically, this step first embeds the basic laws of thermodynamics into the model structure, including implementing the energy conservation constraint of the first law of thermodynamics, implementing the entropy increase principle constraint of the second law of thermodynamics, embedding the phase equilibrium condition constraint, and verifying the thermodynamic consistency relationship. Then, the equation of state constraints of substances are embedded, including embedding the cubic equation of state constraints, embedding the SAFT series of advanced equation of state constraints, implementing the mixing rule constraints, and making corrections to the equation of state under special conditions. Next, the physical constraints describing the mass transfer phenomenon of substances are embedded, including embedding the viscosity model constraints, embedding the heat conduction constraints, embedding the diffusion coefficient constraints, and embedding the surface property constraints. Finally, various physical constraints are optimized and integrated, including optimizing the weights of each physical constraint, combining the hard and soft constraint mechanisms, coordinating the multi-scale constraint relationships, and evaluating the effect of constraint embedding.

[0111] Step S4: Precisely capture the complex interactions between different components in the petrochemical material mixture, predict the non-ideal mixing effects between components, and construct a component interaction knowledge base; Specifically, this step first precisely captures the interactions between two components, including calculating binary interaction parameters, predicting binary mixture excess properties, predicting binary phase diagrams, and simulating special binary interaction effects. Then, it deals with the complex interactions between three or more components, including identifying higher-order interaction terms, optimizing multi-component mixing rules, predicting multi-component phase behavior, and simulating multi-component dynamic interactions. Next, it captures the complex interactions between components through neural networks, including constructing a component graph representation learning framework, implementing a concentration-aware interaction network, constructing a condition-sensitive interaction model, and capturing long-range interactions. Finally, it accurately describes the mixing effects deviating from the ideal behavior, including implementing activity coefficient models, comprehensively modeling excess properties, dealing with special non-ideal systems, and predicting phase separation phenomena.

[0112] Step S5: Evaluate the reliability of the physical property prediction results, quantify the sources and magnitudes of the prediction uncertainties, and generate the confidence intervals of the prediction results; Specifically, this step first identifies and analyzes the sources of uncertainties, including evaluating data uncertainties, analyzing model uncertainties, quantifying physical approximation uncertainties, and evaluating extrapolation uncertainties of prediction conditions. Then, it realizes uncertainty quantification through Bayesian methods, including implementing Bayesian neural networks, implementing Monte Carlo Dropout, constructing Bayesian ensemble learning, and implementing deep Gaussian processes. Next, it tracks the propagation of uncertainties during the prediction process, including analyzing the uncertainty propagation chain, aggregating multi-source uncertainties, analyzing conditional dependence uncertainties, and dealing with extreme condition uncertainties. Finally, it ensures the accuracy of uncertainty estimation, including implementing uncertainty calibration methods, calculating reliability evaluation metrics, realizing uncertainty visualization, and detecting abnormal predictions.

[0113] Step S6: Based on the above steps, predict various physical and chemical properties of petrochemical materials, analyze the relationship between molecular structure and physical properties, and provide visualization and interpretation functions.

[0114] Specifically, this step first predicts various thermodynamic properties, including state properties, phase equilibrium properties, thermodynamic derivative properties, and mixing properties. Then, it predicts various transport properties, including viscosity, heat conduction properties, diffusion coefficients, and surface properties. Next, it predicts the properties related to combustion, including fuel evaluation indices, fire safety properties, combustion thermodynamic properties, and combustion kinetic parameters. Finally, it analyzes the relationship between molecular structure and physical properties, including identifying key structural features, predicting the effects of structural modifications, analyzing component interaction mechanisms, and analyzing the driving factors of physical property changes. At the same time, it provides intuitive result display and interpretation, including multi-dimensional physical property visualization, contribution analysis visualization, uncertainty visualization, and interactive exploration tools.

[0115] The collaborative implementation of the above steps constitutes a complete physical information-constrained physical property calculation method for petrochemical material mixing, which can accurately predict various physical and chemical properties of petrochemical materials, while providing reliable uncertainty assessment and intuitive interpretation functions.

[0116] To more intuitively illustrate the working principle and performance of the present invention, a specific application example is given below.

[0117] In this example, the system of the present invention is used to predict the physical properties of a petrochemical mixture. The mixture consists of n-hexane (40 wt%), cyclohexane (30 wt%), and toluene (30 wt%), and it is required to predict its density, viscosity, surface tension, and octane number in the temperature range of 293.15 - 373.15 K and pressure range of 0.1 - 5.0 MPa.

[0118] First, the data processing and management module collects pure substance physical property data and partial mixture data of the three components from the NIST ThermoData Engine, DIPPR database, and public literature. After cleaning and quality control, a standard format data set is formed. Then, the multi-scale feature characterization module constructs feature characterizations at the quantum, molecular, mesoscopic, and macroscopic levels for the three components, and realizes feature fusion through the attention mechanism. Next, the physical constraint modeling module embeds thermodynamic laws and equation of state constraints into the model to ensure that the prediction results satisfy thermodynamic consistency. The component interaction modeling module accurately captures the interactions between the three components, including three pairs of binary interactions of n-hexane - cyclohexane, n-hexane - toluene, and cyclohexane - toluene, as well as ternary synergistic effects. The uncertainty quantification module evaluates the prediction uncertainty through Bayesian neural network and Monte Carlo Dropout methods. Finally, the physical property prediction and analysis module predicts the required physical properties and analyzes the structure-property relationship based on the above processing.

[0119] The prediction results show that the density of the mixture at 298.15 K and 0.1 MPa is 783.2 ± 1.5 kg / m³, the viscosity is 0.52 ± 0.02 mPa·s, the surface tension is 24.3 ± 0.5 mN / m, and the octane number is 72.5 ± 1.2. Compared with the experimental measurement values, the relative error of density prediction is 0.3%, the relative error of viscosity prediction is 2.5%, the relative error of surface tension prediction is 1.8%, and the absolute error of octane number prediction is 0.8 units. These results are significantly better than traditional methods (such as ideal mixing rules or empirical correlations), and the errors of the latter are usually 2 - 5 times that of this system.

[0120] In addition, structure-property correlation analysis shows that for the viscosity of this mixture, the benzene ring structure of toluene is the most important influencing factor (contributing approximately 45%), followed by the cyclic structure of cyclohexane (contributing approximately 30%) and the chain length of n-hexane (contributing approximately 15%). The synergistic effect among the three components contributes approximately 10%, which explains why the simple mixing rule is difficult to accurately predict the viscosity of this mixture. This in-depth structure-property correlation analysis provides important guidance for the optimization of the mixture formulation.

Claims

1. A physical information-constrained petrochemical material mixing modeling physical property calculation system, characterized in that Including: A data processing and management module, configured to: collect multi-source data of petrochemical materials; clean and perform quality control on the multi-source data; integrate the multi-source data into standard format data; A multi-scale feature characterization module, communicatively connected to the data processing and management module, configured to: receive the standard format data; construct a multi-level petrochemical material feature characterization from the quantum scale to the macroscopic scale; realize the fusion of features at different scales; A physical constraint modeling module, communicatively connected to the multi-scale feature characterization module, configured to: receive the multi-level petrochemical material feature characterization; embed thermodynamic laws and physical law constraints into the model structure; ensure that the physical property prediction results meet the requirements of thermodynamic consistency; A component interaction modeling module, communicatively connected to the multi-scale feature characterization module and the physical constraint modeling module, configured to: accurately capture the complex interactions between different components in a petrochemical material mixture; predict the non-ideal mixing effect between components; construct a component interaction knowledge base; An uncertainty quantification module, communicatively connected to the physical constraint modeling module and the component interaction modeling module, configured to: evaluate the reliability of the physical property prediction results; quantify the sources and magnitudes of the predicted uncertainties; generate a confidence interval for the prediction results; A physical property prediction and analysis module, communicatively connected to the physical constraint modeling module, the component interaction modeling module and the uncertainty quantification module, configured to: predict various physical and chemical properties of petrochemical materials based on the outputs of the above modules; analyze the relationship between molecular structure and physical properties; provide visualization and interpretation functions.

2. The system according to claim 1, wherein The data processing and management module includes: a multi-source data collection unit, configured to collect petrochemical material data from an experimental database, theoretical calculation results, and industrial process data; a data cleaning and quality control unit, communicatively connected to the multi-source data collection unit, configured to: detect and process outliers; verify the thermodynamic consistency of the data; perform a quality score on the data; a data integration and standardization unit, communicatively connected to the data cleaning and quality control unit, configured to: unify the physical property data format; standardize the molecular representation; convert the physical property units; a knowledge graph construction unit, communicatively connected to the data integration and standardization unit, configured to: extract substance entities and substance relationships; construct a substance-property-condition triple; realize knowledge reasoning and completion.

3. The system according to claim 1, wherein The multi-scale feature characterization module includes: a quantum chemical feature characterization unit for extracting molecular electronic structure features, calculating molecular dipole moments and multipole moments, characterizing intermolecular interactions; a molecular structure feature characterization unit communicatively connected to the quantum chemical feature characterization unit for extracting molecular topological structure features, constructing molecular geometric structure features, identifying and characterizing functional groups, and generating molecular fingerprints; a mesoscopic scale feature characterization unit communicatively connected to the molecular structure feature characterization unit for characterizing molecular aggregate features, extracting phase behavior features, and constructing statistical thermodynamics features; a macroscopic thermodynamics feature characterization unit communicatively connected to the mesoscopic scale feature characterization unit for constructing state equation parameter features, extracting thermodynamic property features, and characterizing transport property features; and a multi-scale feature fusion unit communicatively connected to the above-mentioned feature characterization units for realizing effective fusion and weight optimization of features at different scales.

4. The system according to claim 1, wherein The physical constraint modeling module includes: a thermodynamics constraint embedding unit for implementing the energy conservation constraint of the first law of thermodynamics, implementing the entropy increase principle constraint of the second law of thermodynamics, embedding phase equilibrium condition constraints, and verifying thermodynamic consistency relations; an equation of state constraint unit communicatively connected to the thermodynamics constraint embedding unit for embedding cubic equation of state constraints, embedding SAFT series advanced equation of state constraints, implementing mixing rule constraints, and performing equation of state corrections under special conditions; a transport process constraint unit communicatively connected to the thermodynamics constraint embedding unit for embedding viscosity model constraints, embedding heat conduction constraints, embedding diffusion coefficient constraints, and embedding surface property constraints; and a physical constraint optimization and integration unit communicatively connected to the above-mentioned constraint units for optimizing the weights of various physical constraints, combining hard and soft constraint mechanisms, coordinating multi-scale constraint relations, and evaluating the effect of constraint embedding.

5. The system according to claim 1, wherein The component interaction modeling module includes: a binary component interaction unit for calculating binary interaction parameters, predicting binary mixture excess properties, predicting binary phase diagrams, and simulating special binary interaction effects; a multi-component interaction unit communicatively connected to the binary component interaction unit for identifying higher-order interaction terms, optimizing multi-component mixing rules, predicting multi-component phase behavior, and simulating multi-component dynamic interactions; a component interaction characterization network unit communicatively connected to the multi-component interaction unit for constructing a component graph representation learning framework, realizing a concentration-aware interaction network, constructing a condition-sensitive interaction model, and capturing long-range interactions; and a non-ideal mixture effect unit communicatively connected to the component interaction characterization network unit for implementing an activity coefficient model, comprehensively modeling excess properties, dealing with special non-ideal systems, and predicting phase separation phenomena.

6. The system according to claim 1, wherein The uncertainty quantification module includes: a prediction uncertainty source analysis unit for evaluating data uncertainty, analyzing model uncertainty, quantifying physical approximation uncertainty, evaluating prediction condition extrapolation uncertainty; a Bayesian uncertainty quantification unit communicatively connected to the prediction uncertainty source analysis unit for implementing a Bayesian neural network, implementing Monte Carlo Dropout, constructing a Bayesian ensemble learning, implementing a deep Gaussian process; an uncertainty propagation and aggregation unit communicatively connected to the Bayesian uncertainty quantification unit for analyzing an uncertainty propagation chain, aggregating multi-source uncertainties, analyzing conditional dependence uncertainties, handling extreme condition uncertainties; an uncertainty calibration and evaluation unit communicatively connected to the uncertainty propagation and aggregation unit for calibrating uncertainty estimates, evaluating prediction reliability, visualizing uncertainty distributions, detecting abnormal predictions.

7. The system according to claim 1, wherein The physical property prediction and analysis module includes: a thermodynamic property prediction unit for predicting state properties including density, specific heat capacity, enthalpy and entropy, predicting phase equilibrium properties including vapor pressure and critical parameters, calculating thermodynamic derivative properties, predicting mixture properties; a transport property prediction unit communicatively connected to the thermodynamic property prediction unit for predicting viscosity, predicting heat conduction properties, calculating diffusion coefficients, predicting surface properties; a combustion property prediction unit communicatively connected to the thermodynamic property prediction unit for predicting fuel evaluation indices including octane number and cetane number, predicting fire safety properties including flash point and ignition point, calculating combustion thermodynamic properties, predicting combustion kinetic parameters; a structure-property correlation analysis unit communicatively connected to the above-mentioned prediction units for identifying key structural features, predicting structural modification effects, analyzing component interaction mechanisms, analyzing driving factors for physical property changes.

8. The system according to claim 1, wherein The system further includes: a model training and optimization module communicatively connected to the multi-scale feature characterization module, the physical constraint modeling module, the component interaction modeling module and the uncertainty quantification module for implementing multi-task joint learning, optimizing physical constraint implementation, optimizing the model architecture, optimizing hyperparameters, evaluating model performance and selecting the optimal model.

9. The system according to claim 1, wherein The system further includes: a system integration and application module communicatively connected to the physical property prediction and analysis module for providing a user interaction interface, implementing integration with industrial systems, performing data exchange and standardization, supporting system deployment and expansion, and adapting to different application scenarios.

10. Physical information-constrained petrochemical material mixing modeling physical property calculation method, characterized in that, including: Collecting multi-source data of petrochemical materials, and cleaning, quality controlling and standardizing the multi-source data to obtain data in a standard format; Based on the standard format data, construct a multi-level characterization of petrochemical material characteristics from the quantum scale to the macroscopic scale, and achieve the fusion of characteristics at different scales; embed the constraints of the laws of thermodynamics and physical laws into the model structure to ensure that the predicted results of physical properties meet the requirements of thermodynamic consistency; accurately capture the complex interactions between different components in the petrochemical material mixture, predict the non-ideal mixing effects between components, and construct a component interaction knowledge base; evaluate the reliability of the predicted results of physical properties, quantify the sources and magnitudes of the prediction uncertainties, and generate the confidence intervals of the predicted results. Based on the above steps, predict various physical and chemical properties of petrochemical materials, analyze the relationship between molecular structure and physical properties, and provide visualization and interpretation functions.

Images