CN-121999911-A - Intelligent physical property prediction method, device, equipment and medium

CN121999911ACN 121999911 ACN121999911 ACN 121999911ACN-121999911-A

Abstract

The application discloses an intelligent physical property prediction method, device, equipment and medium, and relates to the field of physical property prediction. The method comprises the steps of calculating the eccentricity of a target refrigerant, inputting the group number and the eccentricity of the target refrigerant into a physical property prediction model to obtain thermodynamic physical property prediction values, wherein a determination method of the physical property prediction model comprises the steps of constructing an augmented group contribution prediction model, model parameters in the model comprise a group contribution value, an eccentricity contribution value and a bias term, constructing an objective function and constraint conditions according to the model parameters in the augmented group contribution prediction model, introducing regularization terms of physical property parameters into the objective function, introducing a physical property nonlinear mapping function into the constraint conditions, inputting training data into the constraint conditions, solving the objective function by adopting an interior point method, and taking the augmented group contribution prediction model after the model parameters are determined as the physical property prediction model. The application can improve the prediction precision, the interpretability and the isomer distinguishing capability of physical property prediction.

Inventors

TAO SHAOHUI
MENG XIANGMING
WANG XIJUN
LI ZHENGYONG
SUN XIAOYAN
ZHAO WENYING
XIA LI
XIANG SHUGUANG

Assignees

青岛科技大学

Dates

Publication Date: 20260508
Application Date: 20260211

Claims (10)

1. An intelligent physical property prediction method is characterized by comprising the following steps: obtaining molecular structure parameters of a target refrigerant, wherein the molecular structure parameters comprise the number of groups; calculating the eccentricity of the target refrigerant according to the molecular structure of the target refrigerant; inputting the group number and the eccentricity of the target refrigerant into a physical property prediction model to obtain a thermodynamic physical property prediction value of the target refrigerant molecule; the method for determining the physical property prediction model comprises the following steps: establishing an augmented group contribution prediction model, wherein model parameters in the augmented group contribution prediction model comprise a group contribution value, an eccentricity contribution value and a bias term; Constructing an objective function and a constraint condition according to model parameters in the augmented group contribution prediction model, wherein a regularization item of physical property parameters is introduced into the objective function, and a physical property nonlinear mapping function is introduced into the constraint condition; Inputting training data into the constraint condition, solving the objective function by adopting an interior point method, determining model parameters in the augmented group contribution prediction model, and taking the augmented group contribution prediction model after the model parameters are determined as the physical property prediction model, wherein the training data comprises molecular structure parameters, eccentricity and thermodynamic physical property experimental values of different refrigerants for training.
2. The intelligent physical property prediction method according to claim 1, wherein the expression of the augmented group contribution prediction model is: ; Wherein, the Representing predicted values of thermodynamic properties; representing a group contribution value; Representing an eccentricity contribution value; Represent the first Number of groups of the species group type; Representing eccentricity; Representing a bias term; indicating the number of group types.
3. The intelligent physical property prediction method according to claim 2, wherein the expression of the objective function and the constraint condition is: ; ; Wherein, the Representing physical parameters; regularization term representing physical parameters; is the amount of organics in the refrigerant; Is a positive relaxation variable; is a negative relaxation variable; And Is a super parameter; Is the thermodynamic physical property experimental value; representing a physical nonlinear mapping function; representing thermodynamic physical property experimental values after nonlinear mapping; representing the first of the training data The thermodynamic physical property experimental values of the samples are subjected to nonlinear mapping, wherein L represents the number of the samples in training data; Representation of And (3) with Vector inner product operation result of (2); Representation of And (3) with Vector inner product operation result of (2).
4. The method for predicting physical properties intelligently according to claim 3, wherein if the thermodynamic physical property is boiling point The regularization term of the physical property parameter is , The expression of (2) is Wherein, the method comprises the steps of, As a parameter of the boiling point, Super parameters which are boiling point regularization items; if the thermodynamic property is critical temperature The regularization term of the physical property parameter is , The expression of (2) is Wherein, the method comprises the steps of, As a parameter of the critical temperature, the temperature of the material, Super parameters which are critical temperature regularization items; If the thermodynamic property is critical pressure The regularization term of the physical property parameter is , The expression of (2) is Wherein, the method comprises the steps of, And All of the parameters are the critical pressure parameters, And Are all super parameters of a regularization term of the critical pressure; if the thermodynamic property is critical molar volume The regularization term of the physical property parameter is , The expression of (2) is Wherein, the method comprises the steps of, Is a critical molar volume parameter; super-parameters that are critical molar volume regularization terms; If the thermodynamic property is an eccentric factor The regularization term of the physical property parameter is , The expression of (2) is Wherein, the method comprises the steps of, Is an eccentric factor parameter; the term is regularized for the eccentricity factor.
5. The intelligent physical property prediction method according to claim 1, wherein calculating the eccentricity of the target refrigerant based on the molecular structure of the target refrigerant specifically comprises: acquiring SMILES character strings of molecular structures of target refrigerants; based on the SMILES character string, calling an open source chemical informatics tool bag to perform three-dimensional conformation generation and inertia tensor calculation to obtain the eccentricity of the target refrigerant, wherein the eccentricity is used as a three-dimensional geometric configuration descriptor to distinguish isomers.
6. The intelligent physical property prediction method according to claim 1, wherein training data is input into the constraint condition, the objective function is solved by an interior point method, and model parameters in the augmented group contribution prediction model are determined, specifically including: Inputting training data into the constraint condition, taking an interior point method as a constraint optimization framework, solving the objective function by adopting a sparse Newton iteration method and a finite memory Newton method, and determining model parameters in the augmented group contribution prediction model.
7. The intelligent physical property prediction method according to claim 1, wherein the different refrigerants used for training in the training data comprise a plurality of sets of cis-trans isomer pairs and structural isomer pairs.
8. An intelligent physical property prediction device is characterized in that, the intelligent physical property prediction device comprises: The data acquisition module is used for acquiring molecular structure parameters of the target refrigerant, wherein the molecular structure parameters comprise the number of groups; an eccentricity calculation module for calculating eccentricity of the target refrigerant according to the molecular structure of the target refrigerant; The physical property prediction module is used for inputting the group number and the eccentricity of the target refrigerant into a physical property prediction model to obtain a thermodynamic physical property prediction value of the target refrigerant molecule; the method for determining the physical property prediction model comprises the following steps: establishing an augmented group contribution prediction model, wherein model parameters in the augmented group contribution prediction model comprise a group contribution value, an eccentricity contribution value and a bias term; Constructing an objective function and a constraint condition according to model parameters in the augmented group contribution prediction model, wherein a regularization item of physical property parameters is introduced into the objective function, and a physical property nonlinear mapping function is introduced into the constraint condition; Inputting training data into the constraint condition, solving the objective function by adopting an interior point method, determining model parameters in the augmented group contribution prediction model, and taking the augmented group contribution prediction model after the model parameters are determined as the physical property prediction model, wherein the training data comprises molecular structure parameters, eccentricity and thermodynamic physical property experimental values of different refrigerants for training.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the intelligent physical property prediction method of any one of claims 1-7.
10. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program, when executed by a processor, implements the intelligent physical property prediction method of any of claims 1-7.

Description

Intelligent physical property prediction method, device, equipment and medium Technical Field The application relates to the field of physical property prediction, in particular to an intelligent physical property prediction method, device, equipment and medium. Background Novel green refrigerants (such as HFOs, HFCs and the like) are used as key materials, and development of the novel green refrigerants is highly dependent on accurate thermodynamic property data. However, a large number of novel molecules have serious defects in experimental physical data due to novel structures, difficult synthesis or high measurement cost. The existing physical property prediction methods are mainly divided into a group contribution method, a quantum chemical method (based on a real solvent-like conductor shielding model COSMO, a quantum chemical-based method, a molecular simulation method), a machine learning method and the like. The first class is a traditional group contribution method (Group Contribution Method, GCM), such as Joback, lydersen, constantinou-Gani (CG), marrero-Gani (MG) and other models, and has the advantages of clear physical mechanism, additive linearity, high calculation efficiency and complete interpretation. However, the fundamental defect is that isomers (particularly cis-trans isomers) with the same group composition but different spatial configurations cannot be distinguished based on a two-dimensional topological structure, so that the key working media such as R1234ze (Z) and R1234ze (E) are provided with the identical prediction results, and the molecular prediction and design are severely restricted. The second type is quantum chemistry and molecular simulation methods, including quantum chemistry methods based on electronic structure calculations such as density functional theory (Density Functional Theory, DFT), physical property prediction methods based on a true solvent-like conductor-shielding model (COSMO/COSMO-SAC), and molecular dynamics (Molecular Dynamics, MD) or Monte Carlo (MC) simulations. The method has the advantages of firm physical foundation, high prediction precision (especially no experience dependence on new molecules), capability of providing microscopic mechanism explanation of electron/conformation level, natural support of three-dimensional configuration differential modeling and capability of theoretically distinguishing any type of isomer. However, it is extremely computationally expensive, poorly adaptable to engineering, and lacks a linear additivity and interpretable design interface. The third category is machine learning methods (MACHINE LEARNING, ML), including Random Forest (RF), support vector machines (Support Vector Regression, SVR), multi-layer perceptrons (Multilayer Perceptron, MLP), and the like. Although the fitting precision of the method is always better than that of GCM, the method has three inherent defects of (1) black box property and unexplained property, namely opaque model decision process, incapability of providing quantitative contribution of group level and difficulty in guiding molecular design, (2) isomer distinguishing capability loss, namely that when input characteristics (such as the number of groups) are completely the same for isomers, the black box model inevitably outputs the same result, (3) weak generalization capability, easiness in overfitting (such as KNN in a training set AARD (approximately 0.0001 percent), a testing set >18 percent), high prediction error dispersion and poor engineering reliability. Therefore, how to improve the prediction accuracy, the interpretability, and the isomer discrimination ability of the physical property prediction is a problem to be solved. Disclosure of Invention The application aims to provide an intelligent physical property prediction method, device, equipment and medium, which can improve the prediction precision, the interpretability and the isomer distinguishing capability of physical property prediction. In order to achieve the above object, the present application provides the following. In a first aspect, the present application provides an intelligent physical property prediction method, including the following steps. And obtaining the molecular structure parameters of the target refrigerant, wherein the molecular structure parameters comprise the number of groups. The eccentricity of the target refrigerant is calculated from the molecular structure of the target refrigerant. Inputting the group number and the eccentricity of the target refrigerant into a physical property prediction model to obtain a thermodynamic physical property prediction value of the target refrigerant molecule. The method for determining the physical property prediction model comprises the steps of constructing an augmented group contribution prediction model, wherein model parameters in the augmented group contribution prediction model comprise a group contribution value, an eccentricity contribution val