Method and device for predicting properties of polyether polyol

Abstract

The invention discloses a polyether polyol property prediction method and device, and the method comprises the steps: collecting a set input field and a set actual measurement property field, and obtaining a formula data set; the method comprises the following steps: converting initiator molecules SMILES in a set input field into graph structure representation based on initiator molecular structure representation requirements; based on the formula data set, determining a polyether molecular structure and assigning a force field; constructing a plurality of simulation systems based on a polyether molecular structure; extracting a property descriptor of the simulation system through molecular dynamics simulation; constructing a multi-task learning neural network model by taking splicing vectors of graph structure representation, property descriptors and formula input characteristics as double inputs and taking a set actually measured property field as an output; training is carried out through a self-defined loss function, and property prediction is carried out on polyether polyol through the trained multi-task learning neural network model. The method solves the problems that in the prior art, prediction precision is low, feature dimensions are single, and correlation between properties is not utilized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of polyether polyol property prediction, and in particular to a polyether polyol property prediction method and device. BACKGROUND

[0002] Polyether polyols, as a kind of polymer containing ether bonds in the molecular backbone, are generated by ring-opening polymerization of compounds containing epoxy structures such as ethylene oxide and propylene oxide with initiators containing active hydrogen. Their performance is regulated by multiple factors such as initiator type, epoxy monomer ratio, polymerization method, and molecular weight. In the fields of polyurethane synthesis, coatings, adhesives, and many other polymer materials, the key properties of polyether polyols such as hydroxyl value, viscosity, and cloud point directly determine the performance of the end products. Therefore, fast and accurate property prediction is of great practical significance for formula optimization, process adjustment, and research and development efficiency improvement. Currently, the industry mainly uses single-property modeling methods based on experimental data for prediction, i.e., independent models are constructed for each property such as hydroxyl value and viscosity, and regression analysis or simple machine learning modeling is carried out combined with formula parameters.

[0003] However, the existing technology has significant defects: on the one hand, there are often missing values in the polyether formula and property data sets, which leads to a significant decrease in prediction accuracy of properties with more missing values due to insufficient effective sample size, making it difficult to meet the actual application requirements; on the other hand, traditional modeling methods only rely on macro formula data and do not fully exploit the internal relationship between polyether microstructure and macro properties, resulting in single feature dimension and limiting the improvement of model generalization ability. At the same time, single-property modeling does not utilize the potential correlation between different properties, resulting in low resource utilization and further restricting the breakthrough in prediction performance, which cannot meet the urgent needs of high polymer material research and development for efficient and accurate property prediction.

[0004] Therefore, there is an urgent need for a polyether polyol property prediction method to solve the problems of low prediction accuracy, single feature dimension, and lack of utilization of property correlation in the existing technology. SUMMARY

[0005] To this end, the present application provides a polyether polyol property prediction method and device to solve the problems of low accuracy, single feature dimension, and lack of utilization of property correlation in single-property modeling in existing polyether polyol property prediction, and to realize accurate and simultaneous prediction of multiple properties.

[0006] To achieve the above purpose, the present application provides the following technical solution: a polyether polyol property prediction method, characterized in that it comprises:

[0007] Collecting the set input field and the set measured property field to obtain a formula data set; converting the initiator molecule SMILES in the set input field into a graph structure representation based on the requirement of initiator molecule structure representation;

[0008] Determining a polyether molecule structure and specifying a force field based on the formula data set; constructing a plurality of simulation systems based on the polyether molecule structure; extracting property descriptors of the simulation systems through molecular dynamics simulation;

[0009] Taking the graph structure representation, the property descriptors and the splicing vector of formula input characteristics as double inputs, and taking the set measured property field as output, a multi-task learning neural network model is constructed; the multi-task learning neural network model is trained through a self-defined loss function to obtain a trained multi-task learning neural network model; the trained multi-task learning neural network model is used to predict the properties of polyether polyols.

[0010] As a preferred scheme of the polyether polyol property prediction method, the set input field includes initiator molecular weight, initiator dosage, functionality, polyether theoretical hydroxyl value, propylene oxide dosage and ethylene oxide dosage; the set measured property field includes polyether measured hydroxyl value, viscosity and cloud point; and the graph structure representation includes an atomic feature vector matrix, a bond feature vector matrix and an adjacency matrix.

[0011] As a preferred scheme of the polyether polyol property prediction method, in the process of determining the polyether molecule structure based on the formula data set, the polyether molecular weight is calculated through the hydroxyl value and the initiator functionality; based on the polyether molecular weight, the polymerization amount of propylene oxide and ethylene oxide is calculated by combining the initiator molecular weight, the molar ratio of propylene oxide to ethylene oxide; the polymerization sequence is determined according to the addition sequence of propylene oxide and ethylene oxide to generate the polyether molecule structure;

[0012] The calculation formula of the polyether molecular weight is: ; In the formula, is the polyether molecular weight;

[0013] The calculation formula of the polymerization amount of propylene oxide is: ; In the formula, is the polymerization amount of propylene oxide in one polyether molecule; represents the nearest integer to the following number; is the molecular weight of the initiator; is the molar ratio of propylene oxide to ethylene oxide in the polyether;

[0014] The formula for calculating the polymerization amount of ethylene oxide is:

[0015]

[0016] In the formula, This represents the amount of ethylene oxide polymerized in one polyether molecule.

[0017] As a preferred method for predicting the properties of polyether polyols, the simulated systems include: a pure polyether system, a polyether aqueous solution system, and a 25% mass fraction diethylene glycol butyl ether aqueous solution system; the property descriptors of the simulated systems include: density, dielectric constant, viscosity, isobaric heat capacity, isochoric heat capacity, coefficient of thermal expansion, adiabatic compressibility, mean square fluctuation, and diffusion coefficient.

[0018] As a preferred method for predicting the properties of polyether polyols, during the training of the multi-task learning neural network model using the custom loss function, the custom loss function is the mean squared error loss function; the multi-task learning neural network model includes two input heads and three output heads; the two input heads process the graph structure representation and the concatenated vector respectively; the three output heads correspond to the single prediction of the set measured property field respectively.

[0019] The present invention also provides a polyether polyol property prediction device, employing the above-mentioned polyether polyol property prediction method, comprising:

[0020] The dataset processing module is used to collect the set input fields and set measured property fields to obtain the formulation dataset; based on the requirements for characterizing the initiator molecular structure, the initiator molecules SMILES in the set input fields are converted into a graph structure representation;

[0021] The property descriptor acquisition module is used to determine the polyether molecular structure and specify the force field based on the formulation dataset; construct several simulation systems based on the polyether molecular structure; and extract the property descriptors of the simulation systems through molecular dynamics simulation.

[0022] The multi-task learning neural network model training and application module is used to construct a multi-task learning neural network model by using the concatenated vector of the graph structure representation, the property descriptor, and the formula input features as dual inputs and the set measured property fields as outputs; to train the multi-task learning neural network model by using a custom loss function to obtain the trained multi-task learning neural network model; and to predict the properties of polyether polyols using the trained multi-task learning neural network model.

[0023] As a preferred embodiment of a polyether polyol property prediction device, the dataset processing module includes the following input fields: initiator molecular weight, initiator dosage, functionality, theoretical hydroxyl value of polyether, propylene oxide dosage, and ethylene oxide dosage; the measured property fields include: measured hydroxyl value of polyether, viscosity, and cloud point; and the graph structure representation includes: atomic feature vector matrix, bond feature vector matrix, and adjacency matrix.

[0024] As a preferred embodiment of a polyether polyol property prediction device, in the property descriptor acquisition module, during the process of determining the polyether molecular structure based on the formulation dataset, the polyether molecular weight is calculated using the hydroxyl value and the functionality of the initiator; based on the polyether molecular weight, combined with the initiator molecular weight and the molar ratio of propylene oxide to ethylene oxide, the polymerization quantities of propylene oxide and ethylene oxide are calculated; and according to the order of addition of propylene oxide and ethylene oxide, the polymerization sequence is determined to generate the polyether molecular structure.

[0025] The formula for calculating the molecular weight of the polyether is:

[0026]

[0027] In the formula, The molecular weight of the polyether is [value missing].

[0028] The formula for calculating the polymerization amount of propylene oxide is:

[0029]

[0030] In the formula, The amount of propylene oxide polymerized in one polyether molecule; This indicates taking the nearest integer following the given number; The molecular weight of the initiator; This represents the molar ratio of propylene oxide to ethylene oxide in the polyether.

[0031] The formula for calculating the polymerization amount of ethylene oxide is:

[0032]

[0033] In the formula, This represents the amount of ethylene oxide polymerized in one polyether molecule.

[0034] As a preferred embodiment of a polyether polyol property prediction device, the property descriptor acquisition module includes several simulation systems: a pure polyether system, a polyether aqueous solution system, and a 25% mass fraction diethylene glycol butyl ether aqueous solution system; the property descriptors of the simulation systems include: density, dielectric constant, viscosity, isobaric heat capacity, isochoric heat capacity, coefficient of thermal expansion, adiabatic compressibility, mean square fluctuation, and diffusion coefficient.

[0035] As a preferred embodiment of a polyether polyol property prediction device, in the multi-task learning neural network model training and application module, during the training of the multi-task learning neural network model using the custom loss function, the custom loss function is the mean squared error loss function; the multi-task learning neural network model includes two input heads and three output heads; the two input heads respectively process the graph structure representation and the concatenated vector; the three output heads respectively correspond to the single prediction of the set measured property field.

[0036] The present invention has the following advantages:

[0037] First, this invention integrates macroscopic formulation data with microscopic simulation features, enriches feature dimensions, strengthens the correlation between structure and activity, and improves prediction accuracy;

[0038] Second, this invention solves the problem of insufficient samples caused by missing data by learning a shared feature extraction layer through multi-task learning and making full use of the potential correlation between properties.

[0039] Third, this invention improves data utilization and enhances model robustness by adapting a custom loss function to scenarios where labels are missing.

[0040] Fourth, the present invention adopts a dual-input head design that takes into account both molecular structure and multi-dimensional characteristics, adapts to complex polyether systems, and has a wide range of applications;

[0041] Fifth, this invention enables simultaneous prediction of hydroxyl value, viscosity, and cloud point, simplifying the prediction process and improving R&D and production efficiency. Attached Figure Description

[0042] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0043] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0044] Figure 1 This is a flowchart illustrating a method for predicting the properties of polyether polyols provided in Embodiment 1 of the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of the multi-task learning property prediction model in the polyether polyol property prediction method provided in Embodiment 1 of the present invention;

[0046] Figure 3 This is a schematic diagram of the molecular structure of the polyether in one possible embodiment provided in Embodiment 1 of the present invention;

[0047] Figure 4 This is a schematic diagram of a single polyether molecule structure and multiple simulation systems in one possible embodiment of Embodiment 1 of the present invention; wherein, (A) is the molecular structure of a single polyether molecule; (B) is a pure polyether simulation system, with the blue box representing the periodic boundary; (C) is an aqueous solution simulation system of polyether, with the polyether highlighted by a sphere; (D) is a BDG solution simulation system of polyether, with the polyether highlighted by a sphere.

[0048] Figure 5 This is a schematic diagram of the architecture of a polyether polyol property prediction device provided in Embodiment 2 of the present invention. Detailed Implementation

[0049] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] See Figure 1 Example 1 of the present invention provides a method for predicting the properties of polyether polyols, comprising the following steps:

[0052] S1. Collect the set input fields and set measured property fields to obtain the formulation dataset; based on the requirements for characterizing the initiator molecular structure, convert the initiator molecule SMILES in the set input fields into a graph structure representation;

[0053] S2. Based on the formulation dataset, determine the polyether molecular structure and specify the force field; based on the polyether molecular structure, construct several simulation systems; extract the property descriptors of the simulation systems through molecular dynamics simulations;

[0054] S3. Using the concatenated vector of the graph structure representation, the property descriptor, and the formula input features as dual inputs, and the set measured property fields as outputs, construct a multi-task learning neural network model; train the multi-task learning neural network model using a custom loss function to obtain the trained multi-task learning neural network model; predict the properties of polyether polyols using the trained multi-task learning neural network model.

[0055] In this embodiment, in step S1, the set input fields and set measured property fields are collected to obtain the formulation dataset; based on the requirements for characterizing the initiator molecular structure, the initiator molecule SMILES in the set input fields is converted into a graph structure representation.

[0056] Specifically, firstly, the scope of the input fields and the measured property fields was clearly defined to ensure the completeness and relevance of data collection. The input fields covered core parameters of polyether synthesis, including: initiator molecules (SMILES), initiator molecular weight, initiator dosage, functionality, theoretical hydroxyl value of the polyether, propylene oxide dosage, and ethylene oxide dosage. The measured property fields focused on key application performance of the polyether, including: measured hydroxyl value, viscosity, and cloud point. Data for these fields was collected through multiple channels, including literature review, experimental data accumulation, and industrial production records. After removing outliers and duplicates, the data was integrated to form a standardized formulation dataset, providing a reliable data foundation for subsequent analysis.

[0057] To meet the need for precise characterization of initiator molecular structures, the collected initiator molecules SMILES underwent structural transformation. Using molecular structure processing tools such as rdkit, the chemical structures corresponding to the SMILES strings were analyzed and converted into machine-recognizable graph structures. These graph structures comprise three core components: atomic feature vector matrices, bond feature vector matrices, and adjacency matrices. This enables the quantitative expression of the initiator molecule's microstructure, laying the foundation for models to capture the correlation between molecular structure and polyether properties.

[0058] In this embodiment, in step S2, the polyether molecular structure is determined and the force field is specified based on the formulation dataset; several simulation systems are constructed based on the polyether molecular structure; and the property descriptors of the simulation systems are extracted through molecular dynamics simulation.

[0059] Specifically, based on the formulation dataset, the complete structure of the polyether molecule is derived step by step: First, the molecular weight of the polyether is calculated using the formula for the functionality of the initiator and the hydroxyl value; then, the polymerization quantity of the two epoxy monomers in a single polyether molecule is calculated by combining the initiator molecular weight and the molar ratio of propylene oxide to ethylene oxide; finally, the monomer polymerization sequence is determined according to the order of addition of propylene oxide and ethylene oxide in the formulation, and simultaneously, using the hydroxyl, amino, and imino groups in the initiator as linking sites, excess hydrogen atoms are removed and the corresponding number of epoxy monomer units are connected to form the complete polyether molecular structure. Subsequently, the force field and charge scheme required for molecular simulation are specified, providing the physicochemical basis parameters for molecular dynamics simulation.

[0060] The formula for calculating the molecular weight of the polyether is as follows:

[0061]

[0062] In the formula, The molecular weight of the polyether is [value missing].

[0063] The formula for calculating the polymerization amount of propylene oxide is:

[0064]

[0065] In the formula, The amount of propylene oxide polymerized in one polyether molecule; This indicates taking the nearest integer following the given number; The molecular weight of the initiator; This represents the molar ratio of propylene oxide to ethylene oxide in the polyether.

[0066] The formula for calculating the polymerization amount of ethylene oxide is:

[0067]

[0068] In the formula, This represents the amount of ethylene oxide polymerized in one polyether molecule.

[0069] In this embodiment, based on the constructed polyether molecular structure, three representative simulation systems were built to cover different application scenarios: a pure polyether system, a polyether aqueous solution system, and a 25% (w / w) diethylene glycol butyl ether aqueous solution system. Reasonable molecular ratios and boundary conditions were set for each simulation system to ensure that the simulation results closely approximate actual application conditions.

[0070] Molecular dynamics simulations were conducted using specialized molecular simulation software, such as GROMACS. First, energy minimization was performed using a 10,000-step gradient descent method to eliminate unreasonable atomic forces within the system. Then, NVT ensemble simulations (250 ps, ​​temperature gradually increased from 0°C to 300°C, time step 1 fs) were performed to achieve temperature equilibrium. Subsequently, NPT ensemble simulations (500 ps, ​​300 K, 1 bar, time step 2 fs) were used to adjust the system volume to stability. Finally, NPT ensemble equilibrium state sampling was performed for 10 ns (time step 2 fs) to record the simulation trajectory. Nine property categories were extracted from the simulation trajectory, including density, dielectric constant, viscosity, isobaric heat capacity, isochoric heat capacity, coefficient of thermal expansion, adiabatic compressibility, mean square fluctuation, and diffusion coefficient. A total of 27 property descriptors were obtained from the three simulation systems, comprehensively capturing the microscopic physicochemical characteristics of polyether under different environments.

[0071] The formula for calculating density is as follows:

[0072]

[0073] In the formula, To obtain the density; The total mass of the simulated system; Let V be the volume of the simulated system under pressure bath conditions.

[0074] The formula for calculating the dielectric constant is:

[0075]

[0076] In the formula, Dielectric constant; The dipole moment of the system in the simulation; This represents an ensemble average of the simulated trajectories; It is the vacuum conductivity constant; Let V be the volume of the simulated system under pressure bath conditions. Boltzmann's constant; To simulate temperature.

[0077] The formula for calculating viscosity is:

[0078]

[0079] In the formula, Viscosity; Let V be the volume of the simulated system under pressure bath conditions. Boltzmann's constant; To simulate temperature; To calculate the length of the interval; This represents the shear stress in the simulation. The average of the representative comprehensive courses.

[0080] The formulas for calculating isobaric heat capacity and isochoric heat capacity are as follows:

[0081]

[0082]

[0083] In the formula, It is the isobaric heat capacity; It is an isochoric heat capacity; It simulates the fluctuations in the enthalpy of the system; Boltzmann's constant; To simulate temperature; To simulate the volume of the system; The coefficient of thermal expansion; It represents the adiabatic compressibility.

[0084] The formula for calculating the coefficient of thermal expansion is:

[0085]

[0086] In the formula, The coefficient of thermal expansion; It refers to the fluctuations in the system volume during the simulation; This is to simulate the fluctuations in the enthalpy of the system.

[0087] The formula for calculating adiabatic compressibility is:

[0088]

[0089] In the formula, It represents the adiabatic compressibility.

[0090] The formula for calculating mean square fluctuations is:

[0091]

[0092] In the formula, The mean square fluctuation is represented by N; N is the total number of atoms in the simulated system. This represents the deviation between the position and the average position of atom i.

[0093] The formula for calculating the diffusion coefficient is:

[0094]

[0095] In the formula, D is the diffusion coefficient, and t is the calculation interval. For atoms The location.

[0096] In this embodiment, in step S3, a multi-task learning neural network model is constructed using the concatenated vector of the graph structure representation, the property descriptor, and the formula input features as dual inputs, and the set measured property field as the output. The multi-task learning neural network model is trained using a custom loss function to obtain the trained multi-task learning neural network model. The properties of polyether polyols are predicted using the trained multi-task learning neural network model.

[0097] Specifically, such as Figure 2 As shown, the input and output dimensions of the model are clearly defined, and a targeted multi-task learning neural network architecture is constructed: the two inputs correspond to two types of core features: the first input is the molecular graph structure representation of the initiator obtained in step S1, used to capture molecular microstructure information; the second input is the concatenated vector of the formulation input features and the 27 property descriptors obtained in step S2, integrating macroscopic formulation parameters and microscopic simulation features. The model contains two input heads and three output heads: the first input head is followed by two graph convolutional layers and one fully connected hidden layer to extract features and transform dimensions of the graph structure data; the second input head is followed by two fully connected hidden layers to fuse features of the concatenated vector; the hidden layer representations of the two input heads are concatenated column-wise, and then the shared features are further extracted through two fully connected hidden layers. Finally, the three independent output heads correspond to the prediction tasks of measured hydroxyl value, viscosity, and cloud point, respectively, to achieve simultaneous prediction of multiple properties.

[0098] In this embodiment, a custom loss function adapted to the data characteristics is designed: a mean squared error loss function that can mask missing label values ​​is adopted, and the loss weight vector is related to the effective sample ratio of each measured property field. The expression of the weighted loss function is as follows:

[0099]

[0100] In the formula, Loss is the weighted loss function; This is the loss weight vector with three properties; This is the loss vector with three properties; The proportion of the number of samples after removing missing values ​​for the i-th property to the total number of samples.

[0101] This loss function automatically ignores the interference of missing labeled samples on training, making full use of effective data and improving the stability and reliability of model training. By employing appropriate optimizers, such as the Adam optimizer and hyperparameters, the recipe dataset is proportionally divided into training and validation sets. Model parameters are iteratively updated using the training set, and model performance is monitored using the validation set to avoid overfitting. Ultimately, a convergent and stable well-trained multi-task learning neural network model is obtained.

[0102] In the property prediction stage, for the polyether sample to be predicted, the input field data is collected and set according to the standard procedure in S1, the initiator molecule SMILES is converted into a graph structure representation, and then the molecular structure is determined, the simulation system is constructed and the property descriptors are extracted by molecular dynamics simulation in step S2. The processed dual-input features are input into the trained model. The model quickly outputs the predicted results of measured hydroxyl value, viscosity and cloud point through the learned structure-activity relationship mapping, providing an efficient reference for polyether formulation optimization and production process adjustment.

[0103] In one possible implementation, a verification example is provided as follows:

[0104] Table 1 shows some of the experimentally measured properties of the formulation and molecular simulation.

[0105] Hydroxyl number Initiator Amount of propylene oxide / ethylene oxide 66.3 Ethylene diamine Propylene oxide 935

[0106] Table 1 Measured property field data

[0107] The ethylenediamine has a functionality of 4 and a molecular weight of 60 g / mol. Based on the hydroxyl value and the functionality of ethylenediamine, the molecular weight of the polyether can be calculated to be 3384.6 g / mol. Therefore, one polyether molecule contains 57 propylene oxide units. The molecular structure of the polyether corresponding to this formulation is as follows: Figure 3 As shown.

[0108] The 3D structure of the polyether molecule was obtained using molecular editing tools such as rdkit, and the force field in the molecular simulation was specified. The simulation used the GAFF / TIP3P force field and the bcc charge scheme.

[0109] Based on this, simulation systems were established for pure polyether, aqueous solutions of polyether, and a 25% diethylene glycol butyl ether aqueous solution of polyether (hereinafter referred to as BDG solution), such as... Figure 4 As shown, the simulation system for pure polyether consists of 12 polyether molecules; the simulation for polyether aqueous solution consists of 4400 water molecules and 1 polyether molecule; and the simulation system for polyether in BDG solution consists of 3384 water molecules, 125 diethylene glycol butyl ether molecules, and 1 polyether molecule. The simulation system was established under periodic boundary conditions with a density of 1000 kg / m³.

[0110] Molecular simulations were performed on pure polyether systems, polyether aqueous solutions, and polyether BDG solutions. First, a 10,000-step energy minimization was performed using the gradient descent method. For the minimized structure, an NVT ensemble simulation was performed for 250 ps. In the NVT simulation, the temperature was gradually increased from 0 K to 300 K with a time step of 1 fs. Then, in the NPT ensemble, a 500-p simulation was performed at 300 K and 1 bar to adjust the box size, with a time step of 2 fs. Finally, in the NPT ensemble, a 10 ns equilibrium sampling was performed at 300 K and 1 bar with a time step of 2 fs to obtain the properties of the simulated system. In the NVT and NPT ensemble simulations, the V-rescale method was used to control the temperature; in the NPT ensemble simulation, the isotropic C-rescale method was used to control the pressure. The LINCS method was used to constrain the structures containing hydrogen atoms in the simulations. Molecular simulations were performed using GROMACS software.

[0111] Density, dielectric constant, viscosity, isobaric heat capacity, isochoric heat capacity, coefficient of thermal expansion, adiabatic compressibility, mean square fluctuation, and diffusion coefficient were extracted from the equilibrium simulation trajectories of the three simulation systems. The results are shown in Table 2.

[0112] System Density kg / m 3 ]] Dielectric coefficient Viscosity mPa*s Isothermal heat capacity J / kg / K Isochoric heat capacity J / kg / K Thermal expansion coefficient 1 / K Adiabatic compression ratio m 3 / J]]> Mean square fluctuation nm Diffusion coefficient 10 -5 cm 2 / s]]> Pure polyether 921.32 2.04 31.37 4673.04 4055.47 1.24e-3 9.00e-10 0.25 0.02 Aqueous solution 982.53 68.30 0.49 4673.22 4413.57 6.59e-4 5.63e-10 3.22 3.50 BDG solution 976.18 52.65 0.75 4792.07 4362.38 9.08e-4 6.49e-10 1.02 0.25

[0113] Table 2 Results of extracted property descriptors

[0114] For all the recipes in the dataset, perform molecular simulations using the method described above to obtain the properties listed in the table as descriptors for the recipes, totaling 27.

[0115] Following the standard procedure in S1, data from the input fields are collected and set, and the initiator molecules SMILES are converted into a graph structure representation. Then, in step S2, the molecular structure is determined, the simulation system is constructed, and property descriptors are extracted through molecular dynamics simulation. The processed dual-input features are input into the trained model. The model quickly outputs the predicted results of measured hydroxyl value, viscosity, and cloud point through the learned structure-activity relationship mapping, providing an efficient reference for polyether formulation optimization and production process adjustment.

[0116] In summary, the application scenarios of this invention are as follows:

[0117] In the development of polyether polyol formulations, this invention can quickly predict the hydroxyl value, viscosity, and cloud point under different initiator and epoxy monomer ratios, significantly shortening the formulation iteration cycle and reducing experimental costs.

[0118] In the context of optimizing polyurethane material production processes, this invention can deduce the optimal formulation parameters in reverse based on the target properties, helping to precisely control the production process and improve the quality stability of end products.

[0119] In the development of waterborne polyether products, this invention accurately predicts the turbidity point and water solubility compatibility by simulating the characteristics of aqueous solution systems, providing data support for expanding product application scenarios.

[0120] In the context of quality control in polyether production in chemical enterprises, this invention can predict the key properties of production batches in advance, adjust process parameters in a timely manner, and avoid the production of unqualified products.

[0121] In the context of synergistic optimization of multiple properties of polymer materials, this invention simultaneously predicts multiple core properties, helping to balance the processability and performance of polyether polyols and meet the needs of complex applications.

[0122] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0123] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0124] Example 2

[0125] See Figure 5 Embodiment 2 of the present invention also provides a polyether polyol property prediction device, comprising:

[0126] The dataset processing module 001 is used to collect the set input fields and set measured property fields to obtain the formulation dataset; based on the requirements for characterizing the initiator molecular structure, the initiator molecules SMILES in the set input fields are converted into a graph structure representation;

[0127] The property descriptor acquisition module 002 is used to determine the polyether molecular structure and specify the force field based on the formulation dataset; construct several simulation systems based on the polyether molecular structure; and extract the property descriptors of the simulation systems through molecular dynamics simulation.

[0128] The multi-task learning neural network model training and application module 003 is used to construct a multi-task learning neural network model by using the concatenated vector of the graph structure representation, the property descriptor, and the formula input features as dual inputs and the set measured property fields as outputs; to train the multi-task learning neural network model by using a custom loss function to obtain the trained multi-task learning neural network model; and to predict the properties of polyether polyols using the trained multi-task learning neural network model.

[0129] In this embodiment, the set input fields in the dataset processing module 001 include: initiator molecular weight, initiator dosage, functionality, theoretical hydroxyl value of polyether, propylene oxide dosage, and ethylene oxide dosage; the set measured property fields include: measured hydroxyl value of polyether, viscosity, and cloud point; the graph structure representation includes: atomic feature vector matrix, bond feature vector matrix, and adjacency matrix.

[0130] In this embodiment, in the property descriptor acquisition module 002, during the process of determining the polyether molecular structure based on the formulation dataset, the molecular weight of the polyether is calculated by the hydroxyl value and the functionality of the initiator; based on the molecular weight of the polyether, combined with the molecular weight of the initiator and the molar ratio of propylene oxide to ethylene oxide, the polymerization quantities of propylene oxide and ethylene oxide are calculated; according to the order of addition of propylene oxide and ethylene oxide, the polymerization order is determined to generate the polyether molecular structure;

[0131] The formula for calculating the molecular weight of the polyether is:

[0132]

[0133] In the formula, The molecular weight of the polyether is [value missing].

[0134] The formula for calculating the polymerization amount of propylene oxide is:

[0135]

[0136] In the formula, The amount of propylene oxide polymerized in one polyether molecule; This indicates taking the nearest integer following the given number; The molecular weight of the initiator; This represents the molar ratio of propylene oxide to ethylene oxide in the polyether.

[0137] The formula for calculating the polymerization amount of ethylene oxide is:

[0138]

[0139] In the formula, This represents the amount of ethylene oxide polymerized in one polyether molecule.

[0140] In this embodiment, the property descriptor acquisition module 002 includes several simulation systems: a pure polyether system, a polyether aqueous solution system, and a 25% mass fraction diethylene glycol butyl ether aqueous solution system; the property descriptors of the simulation systems include: density, dielectric constant, viscosity, isobaric heat capacity, isochoric heat capacity, coefficient of thermal expansion, adiabatic compressibility, mean square fluctuation, and diffusion coefficient.

[0141] In this embodiment, in the multi-task learning neural network model training and application module 003, during the training of the multi-task learning neural network model using the custom loss function, the custom loss function is the mean squared error loss function; the multi-task learning neural network model includes two input heads and three output heads; the two input heads respectively process the graph structure representation and the concatenated vector; the three output heads respectively correspond to the single-item prediction of the set measured property field.

[0142] It should be noted that the information interaction and execution process between the modules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0143] Example 3

[0144] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a method of predicting the properties of polyether polyols. The program code includes instructions for executing the method of predicting the properties of polyether polyols according to Embodiment 1 or any possible implementation thereof.

[0145] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0146] Example 4

[0147] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0148] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute a polyether polyol property prediction method according to Embodiment 1 or any possible implementation thereof by calling the program instructions.

[0149] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0150] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0151] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0152] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for predicting the properties of polyether polyols, characterized in that, include: The formula dataset is obtained by collecting data from the defined input fields and the defined measured property fields. Based on the requirement for characterizing the molecular structure of initiators, the initiator molecules SMILES in the specified input fields are converted into a graph structure representation; Based on the aforementioned formulation dataset, the polyether molecular structure was determined and the force field was specified; Based on the polyether molecular structure, several simulation systems were constructed; property descriptors of the simulation systems were extracted through molecular dynamics simulations. A multi-task learning neural network model is constructed using the concatenated vector of the graph structure representation, the property descriptor, and the formula input features as dual inputs, and the set measured property fields as outputs. The multi-task learning neural network model is trained using a custom loss function to obtain a trained multi-task learning neural network model. The properties of polyether polyols are predicted using the trained multi-task learning neural network model.

2. The method for predicting the properties of polyether polyols according to claim 1, characterized in that, The input fields include: initiator molecular weight, initiator dosage, functionality, theoretical hydroxyl value of polyether, propylene oxide dosage, and ethylene oxide dosage; the measured property fields include: measured hydroxyl value of polyether, viscosity, and cloud point; the graph structure representation includes: atomic feature vector matrix, bond feature vector matrix, and adjacency matrix.

3. The method for predicting the properties of polyether polyols according to claim 2, characterized in that, In determining the polyether molecular structure based on the formulation dataset, the polyether molecular weight is calculated using hydroxyl value and initiator functionality. Based on the molecular weight of the polyether, combined with the molecular weight of the initiator and the molar ratio of propylene oxide to ethylene oxide, the polymerization quantities of propylene oxide and ethylene oxide are calculated; according to the order of addition of propylene oxide and ethylene oxide, the polymerization sequence is determined to generate the polyether molecular structure. The formula for calculating the molecular weight of the polyether is: ； In the formula, The molecular weight of the polyether is [value missing]. The formula for calculating the polymerization amount of propylene oxide is: ； In the formula, The amount of propylene oxide polymerized in one polyether molecule; This indicates taking the nearest integer following the given number; The molecular weight of the initiator; This represents the molar ratio of propylene oxide to ethylene oxide in the polyether. The formula for calculating the polymerization amount of ethylene oxide is: ； In the formula, This represents the amount of ethylene oxide polymerized in one polyether molecule.

4. The method for predicting the properties of polyether polyols according to claim 3, characterized in that, The simulated systems include: a pure polyether system, a polyether aqueous solution system, and a 25% mass fraction diethylene glycol butyl ether aqueous solution system; the property descriptors of the simulated systems include: density, dielectric constant, viscosity, isobaric heat capacity, isochoric heat capacity, coefficient of thermal expansion, adiabatic compressibility, mean square fluctuation, and diffusion coefficient.

5. The method for predicting the properties of polyether polyols according to claim 4, characterized in that, During the training of the multi-task learning neural network model using the custom loss function, the custom loss function is the mean squared error loss function; the multi-task learning neural network model includes two input heads and three output heads; the two input heads process the graph structure representation and the concatenated vector respectively; the three output heads correspond to the single-item prediction of the set measured property field respectively.

6. A device for predicting the properties of polyether polyols, employing the method for predicting the properties of polyether polyols according to any one of claims 1-5, characterized in that, include: The dataset processing module is used to collect data from the set input fields and the set measured property fields to obtain the formula dataset; Based on the requirement for characterizing the molecular structure of initiators, the initiator molecules SMILES in the specified input fields are converted into a graph structure representation; The property descriptor acquisition module is used to determine the polyether molecular structure and specify the force field based on the formulation dataset; Based on the polyether molecular structure, several simulation systems were constructed; property descriptors of the simulation systems were extracted through molecular dynamics simulations. The multi-task learning neural network model training and application module is used to construct a multi-task learning neural network model by using the concatenated vector of the graph structure representation, the property descriptor, and the formula input features as dual inputs and the set measured property fields as outputs; to train the multi-task learning neural network model by using a custom loss function to obtain the trained multi-task learning neural network model; and to predict the properties of polyether polyols using the trained multi-task learning neural network model.

7. The polyether polyol property prediction device according to claim 6, characterized in that, In the dataset processing module, the set input fields include: initiator molecular weight, initiator dosage, functionality, theoretical hydroxyl value of polyether, propylene oxide dosage, and ethylene oxide dosage; the set measured property fields include: measured hydroxyl value of polyether, viscosity, and cloud point; the graph structure representation includes: atomic feature vector matrix, bond feature vector matrix, and adjacency matrix.

8. The polyether polyol property prediction device according to claim 7, characterized in that, In the property descriptor acquisition module, during the process of determining the polyether molecular structure based on the formulation dataset, the polyether molecular weight is calculated using the hydroxyl value and the functionality of the initiator. Based on the molecular weight of the polyether, combined with the molecular weight of the initiator and the molar ratio of propylene oxide to ethylene oxide, the polymerization quantities of propylene oxide and ethylene oxide are calculated; according to the order of addition of propylene oxide and ethylene oxide, the polymerization sequence is determined to generate the polyether molecular structure. The formula for calculating the molecular weight of the polyether is: ； In the formula, The molecular weight of the polyether is [value missing]. The formula for calculating the polymerization amount of propylene oxide is: ； In the formula, The amount of propylene oxide polymerized in one polyether molecule; This indicates taking the nearest integer following the given number; The molecular weight of the initiator; This represents the molar ratio of propylene oxide to ethylene oxide in the polyether. The formula for calculating the polymerization amount of ethylene oxide is: ； In the formula, This represents the amount of ethylene oxide polymerized in one polyether molecule.

9. The polyether polyol property prediction device according to claim 8, characterized in that, In the property descriptor acquisition module, the simulated systems include: a pure polyether system, a polyether aqueous solution system, and a 25% mass fraction diethylene glycol butyl ether aqueous solution system; the property descriptors of the simulated systems include: density, dielectric constant, viscosity, isobaric heat capacity, isochoric heat capacity, coefficient of thermal expansion, adiabatic compressibility, mean square fluctuation, and diffusion coefficient.

10. The polyether polyol property prediction device according to claim 9, characterized in that, In the multi-task learning neural network model training and application module, during the training of the multi-task learning neural network model using the custom loss function, the custom loss function is the mean squared error loss function; the multi-task learning neural network model includes two input heads and three output heads; the two input heads process the graph structure representation and the concatenated vector respectively; the three output heads correspond to the single-item prediction of the set measured property field respectively.

Images