Method for predicting relative correction factor of monomer raw material in electronic atomization liquid and storage medium

Abstract

The present application relates to the field of electronic atomization liquid, and relates to a method for predicting relative correction factors of monomer raw materials in electronic atomization liquid and a storage medium, comprising: obtaining a target molecular structural formula of a monomer raw material to be detected and a quantitative detection parameter of the monomer raw material to be detected; performing molecular descriptor calculation on the target molecular structural formula to obtain a target molecular descriptor; inputting the target molecular descriptor into a target integrated prediction model for prediction to obtain a target relative correction factor, which is used for quantitative calculation with the quantitative detection parameter of the monomer raw material to be detected to obtain a target concentration of the monomer raw material to be detected. The present application uses a trained target integrated prediction model to accurately predict the relative correction factor of the monomer, realizes accurate correction factors without standard products, and further accurately calculates the concentration of the monomer raw material to be detected, effectively improving the efficiency and accuracy of quantitative analysis.

Description

Technical Field

[0001] This invention relates to the field of electronic atomizing fluid technology, specifically to a method for predicting the relative correction factor of monomer raw materials in electronic atomizing fluid and a storage medium thereon. Background Technology

[0002] Electronic vapor fluid is a complex mixture composed of propylene glycol (PG), vegetable glycerin (VG), nicotine, and dozens to hundreds of monomeric ingredients. With the diversification of e-cigarette flavors, accurate determination of flavor content is crucial for product quality and safety. Currently, gas chromatography-flame ionization detector (GC-FID) is commonly used to detect flavor components. Quantification typically employs the internal standard method, using n-heptadecane as the internal standard, but this requires knowing the correction factor for each flavor component relative to n-heptadecane. Due to the wide variety of flavor components, traditional methods require determining the correction factor for each one using standards, which is costly, time-consuming, and lacks standards for some new flavor components, failing to address the problem of obtaining correction factors when standards are unavailable, thus limiting the accurate detection of flavor components in e-cigarette fluids. Summary of the Invention

[0003] One objective of this invention is to provide a method, apparatus, electronic atomizing fluid quality control and analysis system, and computer storage medium for predicting the relative correction factor of monomer raw materials in electronic atomizing fluids, in order to solve the technical problem that the relative correction factor of monomer raw materials in electronic atomizing fluids cannot be accurately obtained under the condition of no standard.

[0004] In a first aspect, embodiments of the present invention provide a method for predicting the relative correction factor of monomer raw materials in electronic atomizing liquids, the method comprising: Obtain the target molecular structure of the monomer raw material to be tested and the quantitative detection parameters of the monomer raw material to be tested; Molecular descriptor calculation is performed on the target molecule structure to obtain the target molecule descriptor; The target molecule descriptor is input into the target integrated prediction model for prediction to obtain the target relative correction factor. The target relative correction factor is used to perform quantitative calculation with the quantitative detection parameters of the monomer raw material to be detected to obtain the target concentration of the monomer raw material to be detected.

[0005] In a second aspect, a device for predicting the relative correction factor of monomer raw materials in an electronic atomizing liquid is provided, the device comprising: The acquisition unit is used to acquire the target molecular structure formula of the monomer raw material to be detected and the quantitative detection parameters of the monomer raw material to be detected. A computing unit is used to perform molecular descriptor calculation on the target molecular structure to obtain the target molecular descriptor; The prediction unit is used to input the target molecule descriptor into the target integrated prediction model for prediction to obtain the target relative correction factor. The target relative correction factor is used to perform quantitative calculation with the quantitative detection parameters of the monomer raw material to be detected to obtain the target concentration of the monomer raw material to be detected.

[0006] In a third aspect, an electronic atomizing liquid quality control and analysis system is provided, the electronic atomizing liquid quality control and analysis system including an electronic device, the electronic device including a memory and a processor, the memory being connected to the processor, the processor being configured to execute one or more computer programs stored in the memory, the processor, when executing the one or more computer programs, causing the electronic atomizing liquid quality control and analysis system to implement the relative correction factor prediction method for monomer raw materials in electronic atomizing liquid as described in the first aspect.

[0007] In a fourth aspect, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the method for predicting the relative correction factor of monomer raw materials in an electronic atomizing liquid as described in the first aspect.

[0008] In the embodiments of the above-mentioned method, apparatus, electronic atomizing liquid quality control and analysis system, and computer storage medium for predicting the relative correction factor of monomer raw materials in electronic atomizing liquids, this embodiment obtains the molecular structural formula of the monomer raw material to be detected and calculates the corresponding molecular descriptor, and accurately predicts the relative correction factor of the monomer using a target integrated prediction model. Quantitative analysis of monomer raw materials can be completed without the need for standards, reducing experimental costs and time consumption. Furthermore, by combining the predicted relative correction factor with quantitative detection parameters, the concentration of the monomer raw material to be detected can be accurately calculated, improving the efficiency and accuracy of quantitative detection. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the structure of an electronic atomizing liquid quality control and analysis system according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for predicting the relative correction factor of monomer raw materials in an electronic atomizing liquid according to an embodiment of the present invention. Figure 3This is a ranking diagram of the importance of the top 15 features of a random forest model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a relative correction factor prediction device for monomer raw materials in electronic atomizing liquid according to an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0012] It should be noted that, unless otherwise specified, the various features in the embodiments of this invention can be combined with each other, all of which are within the protection scope of this invention. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. Moreover, the terms "first," "second," and "third" used in this invention do not limit the data or execution order, but only distinguish identical or similar items with essentially the same function and effect.

[0013] In existing technologies, the quantitative detection and correction factor acquisition of fragrance components in electronic atomizing liquids mainly rely on the following techniques: First, the external standard method, which requires preparing a series of standard solutions of different concentrations for each fragrance component to be tested and plotting a standard curve. However, this method consumes a large amount of standards and is completely ineffective for compounds without standards. Second, the peak area normalization method, which directly estimates the relative content using the percentage of peak area. However, since the response differences of different compounds on the FID detector can be several times or even tens of times (the response is closely related to the effective carbon number and functional group type), this method has a large quantitative error and can only make a rough estimate. Third, the internal standard method, which involves adding a known amount of n-heptadecane as an internal standard and using the peak area ratio of the analyte to the internal standard for quantification. Although this method can correct for injection fluctuations and instrument drift to some extent, it still requires prior experimental determination of the relative correction factor of each compound relative to n-heptadecane.

[0014] However, there are many types of monomer raw materials used in electronic atomizing liquids, with more than 400 commonly used fragrances. If the relative correction factor of each monomer is determined one by one using traditional experimental methods, there are problems such as high cost of purchasing standard products, long experimental cycle and large workload. More importantly, some new monomer raw materials or rare natural extracts lack commercial standard products, which makes it impossible to obtain their relative correction factors directly through experiments.

[0015] In summary, existing technologies all rely on experimental determination of standards and cannot solve the problem of how to obtain relative correction factors when standards are unavailable.

[0016] Please see Figure 1 , Figure 1 This is a schematic diagram of an electronic atomizing fluid quality control and analysis system. Figure 1 In the electronic atomizing liquid quality control and analysis system 10, there is an electronic device 20, which includes at least one processor 201 and a memory 202.

[0017] Among them, the electronic device 20 can be a chromatography data workstation, a laboratory computing terminal, an industrial Internet of Things edge computing terminal, or a cloud quality control server, etc., and is not limited to one specific device.

[0018] Specifically, the chromatographic data workstation or laboratory computing terminal can be used to communicate with a gas chromatograph (GC-FID) to directly obtain the quantitative detection parameters (such as peak area) of the monomer raw material to be detected, and perform molecular descriptor calculation and target integrated prediction model prediction locally; the cloud quality control server can be used to centrally deploy the target integrated prediction model, receive molecular structure formulas and quantitative detection parameters sent from terminal devices for remote prediction and concentration calculation, and realize centralized management and analysis of electronic atomized liquid quality data.

[0019] The processor 201 is configured to support the electronic atomizing fluid quality control and analysis system in performing the corresponding functions of the relative correction factor prediction method for monomer raw materials in the electronic atomizing fluid described in the above method embodiments, such as instantiating the parsing processing module, querying the strategy mapping relationship, and performing rule-based data transformation and mapping calculations. The processor 201 can be a central processing unit (CPU), a microcontroller, or a microprocessor (suitable for edge computing gateway scenarios). In an optional embodiment, when the number of external devices connected is large and the data concurrency is high, the processor 201 may further include a graphics processing unit (GPU) or a neural network processor for parallel format parsing and numerical transformation acceleration of the first object model data from multiple devices. The aforementioned hardware chip can also be an application-specific integrated circuit (ASIC) or any combination thereof.

[0020] The memory 202 stores program code and various data required to support the prediction method for the relative correction factor of monomer raw materials in the electronic atomizing liquid. The memory 202 is coupled to the processor 201 and stores the program code for execution by the processor 201, as well as the strategy mapping list, first rule, second rule, third rule, and first intermediate data and second intermediate data generated during processing. The memory 202 may include volatile memory, such as random access memory (RAM), to provide operating space and data processing cache for the parsing processing module; the memory 202 may also include non-volatile memory, such as read-only memory, flash memory, hard disk, or solid-state drive, to persistently store the preset strategy mapping list and the final target business data; the memory 202 may also include combinations of the above types of memory.

[0021] See Figure 2 , Figure 2 A flowchart illustrating a method for predicting the relative correction factor of monomer raw materials in electronic atomizing liquids according to an embodiment of the present invention is provided. The method includes the following steps: S10. Obtain the target molecular structure of the monomer raw material to be tested and the quantitative detection parameters of the monomer raw material to be tested.

[0022] Among them, the quantitative detection parameters refer to the quantitative information obtained by experimentally measuring the analyte monomer raw material using gas chromatography-flame ionization detector (GC-FID) technology. These parameters mainly include the peak area of ​​the analyte monomer raw material, the peak area of ​​the n-heptadecane internal standard, and the concentration of the n-heptadecane internal standard.

[0023] The target molecular structure refers to the standardized expression of the chemical molecular structure of the monomer raw material to be detected, usually using the Simplified Molecular Input Line Entry System (SMILES) format. SMILES is a standardized method for representing molecular structures using ASCII strings, which can accurately describe the types of atoms, bond connections, ring structures, and branching information in a molecule.

[0024] In practice, for electronic atomizing liquid monomer raw materials that require quantitative analysis but lack physical standards, their identification information is determined. This identification information includes the Chemical Abstracts Service Number or the standard name of the compound. Based on this identification information, a chemical structure database is used to search and match, and the standard simplified molecular linear input standard string corresponding to the monomer raw material to be tested is obtained as the target molecular structure.

[0025] For example, for the monomer raw material to be tested, d-decyl lactone, its Chemical Abstracts Service Number is determined to be 705-86-2. Based on this Chemical Abstracts Service Number, a search is performed in the Chemical Identifier Resolver Database to obtain its canonical simplified molecular linear input canonical string as CCCCCCC1CCC(=O)O1. This string is used as the target molecular structure of the monomer raw material to be tested.

[0026] In practice, a gas chromatograph paired with a flame ionization detector is used to perform chromatographic analysis on an electronic atomized liquid sample containing the monomer raw material to be detected and n-heptadecane internal standard. The chromatographic peak area corresponding to the monomer raw material to be detected is extracted from the gas chromatography data workstation as the peak area of ​​the analyte, and the chromatographic peak area corresponding to the n-heptadecane internal standard is extracted as the peak area of ​​the internal standard. The known concentration of n-heptadecane internal standard added when preparing the electronic atomized liquid sample is obtained as the internal standard concentration. The peak area of ​​the analyte, the peak area of ​​the internal standard, and the concentration of the internal standard are used together to form the quantitative detection parameters of the monomer raw material to be detected.

[0027] As can be seen, in this example, by retrieving the target molecular structure from the chemical structure database and extracting the measured peak area of ​​the analyte, the peak area of ​​the internal standard, and the concentration of the internal standard from the gas chromatography data workstation as quantitative detection parameters, a complete data foundation was established for subsequent calculation of the relative correction factor based on the molecular structure and concentration calculation in combination with the actual detection parameters. This allows the analyte monomer raw material without a standard to directly enter the quantitative analysis process based on its inherent structural information and conventional chromatographic detection data.

[0028] S20. Perform molecular descriptor calculation on the target molecular structure to obtain the target molecular descriptor.

[0029] The target molecular structure refers to the chemical structure expression of a single fragrance molecule to be predicted or analyzed, usually in SMILES format or other standardized molecular representation methods.

[0030] Among them, molecular descriptors refer to quantitative features calculated based on molecular structure, covering physicochemical properties, topological structure, number of functional groups, etc., which can reflect the essential characteristics of molecules.

[0031] The target molecule descriptor, or multidimensional feature vector, is a set of numerical feature parameters extracted after systematically analyzing the target molecule's structural formula using cheminformatics tools. In practice, cheminformatics computing tools, such as open-source software like RDKit and ChemAxon, are selected. These tools can automatically analyze molecular structural formulas and calculate various molecular descriptors. The target molecule's structural formula is input into the computing tool, and using predefined calculation methods, molecular descriptors of, but not limited to, the following types are extracted: basic physicochemical properties (molecular weight, lipid-water partition coefficient, polar surface area, etc.), ring system structure information (number of aromatic rings, number of aliphatic rings, etc.), functional group counts (ester groups, hydroxyl groups, ether bonds, etc.), molecular fingerprint features (such as the number of active sites in Morgan fingerprints), and domain-specific descriptors (such as the effective carbon number, the number of carbon atoms corrected according to the detection mechanism). The calculated values ​​are then arranged and combined in a preset order to form the target molecule descriptor corresponding to the monomer raw material to be predicted.

[0032] S30. Input the target molecule descriptor into the target integrated prediction model for prediction to obtain the target relative correction factor. The target relative correction factor is used to perform quantitative calculation with the quantitative detection parameters of the monomer raw material to be detected to obtain the target concentration of the monomer raw material to be detected.

[0033] Among them, the target ensemble prediction model refers to a comprehensive prediction model composed of at least two pre-trained basic machine learning models, which are combined with a preset ensemble strategy to fuse the prediction results of multiple models.

[0034] Among them, the target relative correction factor refers to the factor output by the target ensemble model for quantitative correction, which corrects the differences in detection response of different monomers.

[0035] The target concentration refers to the final quantitative result of the monomer raw material to be detected in the sample.

[0036] In one embodiment, the target ensemble prediction model includes at least two pre-trained base models and a preset ensemble strategy. The step of inputting the target molecule descriptor into the target ensemble prediction model for prediction to obtain the target relative correction factor includes: inputting the target molecule descriptor into each base model for prediction to obtain a first relative correction factor output by each base model; and performing ensemble processing on the first relative correction factors output by each base model according to the preset ensemble strategy to obtain the target relative correction factor.

[0037] The base model refers to a single machine learning model, such as a random forest or a support vector machine. The base model can independently predict the relative correction factor.

[0038] Among them, the preset integration strategy is a rule or algorithm that integrates the prediction results of multiple basic models, such as simple averaging, weighted averaging, or stacked integration.

[0039] Here, the first relative correction factor refers to the predicted relative correction factor independently output by each base model for the input target molecule descriptor. Therefore, in this embodiment, the target relative correction factor refers to the final predicted relative correction factor obtained after integrating multiple first relative correction factors.

[0040] Specifically, the target molecule descriptor is input into each base model in the target ensemble prediction model for prediction. Each base model independently calculates the target molecule descriptor based on its learned mapping relationship and outputs the corresponding first relative correction factor. At least two first relative correction factors are integrated according to the preset integration strategy. When the preset integration strategy is arithmetic mean, the arithmetic mean of all first relative correction factors is calculated as the target relative correction factor. When the preset integration strategy is weighted summation, the weights are calculated based on the determination coefficients of each base model, and the first relative correction factors are weighted summated to obtain the target relative correction factor. When the preset integration strategy is meta-learner fusion, each first relative correction factor is used as a meta-feature input into the meta-learner for secondary calculation to obtain the target relative correction factor.

[0041] Further, quantitative detection parameters of the monomer raw material to be detected are obtained, including the peak area of ​​the monomer raw material, the peak area of ​​the internal standard, and the concentration of the internal standard. The target relative correction factor and the quantitative detection parameters are substituted into the internal standard method quantitative calculation formula to obtain the target concentration of the monomer raw material to be detected; the internal standard method quantitative calculation formula is: Cᵢ = (Aᵢ × C s ) / (A s × fpred), where Cᵢ is the target concentration of the monomer raw material to be detected, Aᵢ is the peak area of ​​the monomer raw material to be detected, and C s For the concentration of the internal standard, A s is the peak area of ​​the internal standard, and fpred is the target relative correction factor; For example, for the monomer raw material to be detected—proninolactone—its corresponding 25-dimensional target molecule descriptor is input into a random forest basic model, an extreme gradient boosting tree basic model, and a support vector machine basic model, respectively. The first relative correction factors output independently by the three basic models are 1.18, 1.20, and 1.13, respectively. When the preset integration strategy is arithmetic mean, (1.18+1.20+1.13) / 3=1.17 is calculated, and 1.17 is determined as the target relative correction factor fpred. The quantitative detection parameters obtained by GC-FID detection are obtained. The peak area Aᵢ of proninolactone is 230000, and the peak area Aᵢ of the n-heptadecane internal standard is... s The concentration C of the n-heptadecane internal standard is 200,000. s The target concentration of nonanolide is 0.1 mg / mL. Substituting the above parameters into the internal standard method calculation formula: Cᵢ = (230000 × 0.1) / (200000 × 1.17) ≈ 0.0983 mg / mL, the target concentration of nonanolide is calculated to be approximately 0.0983 mg / mL. This allows for the acquisition of the relative correction factor for unknown monomer raw materials without the need for standards, and enables quantitative analysis.

[0042] As can be seen, in this embodiment, by inputting the high-dimensional target molecule descriptor into multiple pre-trained basic models and using a preset integration strategy to fuse the prediction results, the target relative correction factor is obtained, which significantly improves the scientificity and reliability of quantitative analysis. Furthermore, by combining the GC-FID internal standard method, accurate concentration determination of unknown monomer raw materials can be achieved without the need for standards, simplifying the detection process and reducing experimental costs.

[0043] In one embodiment, before inputting the target molecule descriptor into the target integrated prediction model for prediction to obtain the target relative correction factor, the method implementation process can be referred to A1-A7.

[0044] A1. Obtain an initial dataset under the same chromatographic conditions, wherein the initial dataset includes multiple monomeric raw materials and a relative correction factor corresponding to each monomeric raw material.

[0045] The relative correction factor refers to the ratio correction value between the response value of the analyte monomer and the response value of the internal standard n-heptadecane in gas chromatography flame ionization detector analysis. It is used to eliminate the influence of detector response differences for different compounds.

[0046] Among them, the same chromatographic conditions refer to the experimental environment in which the gas chromatograph is equipped with a hydrogen flame ionization detector and the detection is carried out under the set fixed detector temperature, hydrogen flow rate, air flow rate and tail gas flow rate.

[0047] The initial dataset refers to the original data set composed of the identification information of various monomeric raw materials and their corresponding measured relative correction factors.

[0048] In the specific implementation of A1, ≥300 commonly used monomer raw materials for electronic atomizing liquids were selected as training samples to ensure that the selected monomer raw materials covered esters, aldehydes, ketones, alcohols, acids, lactones, thioethers, cooling agents, and matrix components such as propylene glycol, vegetable glycerin, and nicotine. Gas chromatography with a flame ionization detector was used, with n-heptadecane as an internal standard, to perform gas chromatography analysis on each monomer raw material under the same set chromatographic conditions. The relative correction factor of each monomer raw material relative to n-heptadecane was determined, and each monomer raw material was associated with and stored with its corresponding relative correction factor to construct the initial dataset.

[0049] The same chromatographic conditions can be set as follows: the temperature of the hydrogen flame ionization detector is set to 250~300℃, the hydrogen flow rate is set to 30~40 mL / min, the air flow rate is set to 300~400 mL / min, and the tail gas flow rate is set to 20~30 mL / min.

[0050] For example, 409 commonly used monomer raw materials for electronic atomizing liquids were selected and analyzed using gas chromatography with a flame ionization detector. Under the same chromatographic conditions of detector temperature 250-300℃, hydrogen flow rate 30-40 mL / min, air flow rate 300-400 mL / min, and make-up gas flow rate 20-30 mL / min, using n-heptadecane as an internal standard, the relative correction factor for butyldecyl lactone was determined to be 0.547, and the relative correction factor for ethyl acetate was determined to be 0.431. The relative correction factor of butyldecyl lactone with its relative correction factor of 0.547 was associated and stored, and the relative correction factor of ethyl acetate with its relative correction factor of 0.431 was associated and stored. In this way, each of the 409 monomer raw materials was measured and associated and stored, thus constructing an initial dataset containing 409 sets of corresponding relationships.

[0051] Therefore, the systematic determination of large-scale, multi-category monomer raw materials under unified chromatographic conditions ensured the uniformity of data sources and the comparability of experimental data.

[0052] A2. Obtain the molecular structural formula corresponding to each monomer raw material.

[0053] In a specific implementation of A2, obtaining the molecular structural formula corresponding to each monomer raw material includes: obtaining the chemical identifier of each monomer raw material; using the chemical identifier of each monomer raw material as a query identifier, querying in a preset chemical structure database to obtain the molecular structural formula corresponding to each monomer raw material.

[0054] Among them, monomeric raw materials refer to fragrance compounds with a single chemical structure, which are the basic units that constitute the flavor of e-liquids. Each monomeric raw material has a unique Chemical Abstracts Service Number (CAS Number), which is a globally recognized unique chemical identifier assigned by the Chemical Abstracts Service (CASS) to accurately and unambiguously identify compounds.

[0055] Among them, the most commonly used structural representation of molecular structures is the Simplified Molecular Input Line Entry System (SMILES) format. SMILES uses a specific string of characters to describe the atomic composition and their connections in a molecule, which can simply and accurately represent complex molecular structures, facilitating computer recognition and subsequent processing.

[0056] Chemical identifiers refer to text codes used to uniquely identify chemical substances in computer systems and databases. These chemical identifiers include, but are not limited to, the CAS number and / or standard name of a monomer (such as the IUPAC nomenclature or common name). Since the same flavoring monomer may have multiple common names or names in different languages, using the CAS number or standard name as a chemical identifier eliminates ambiguity in natural language and ensures the uniqueness and accuracy of subsequent queries.

[0057] Among them, the preset chemical structure database refers to a retrieval platform that stores the identification of chemical substances and their corresponding molecular structure data. The preset chemical structure database may include, but is not limited to, PubChem (a public chemical structure database maintained by the National Center for Biotechnology Information in the United States) and Chemical Identifier Resolver (CIR for short, an online tool for resolving and converting chemical identifiers).

[0058] Specifically, the obtained chemical identifier in text form (such as "123-92-2" or "Isoamyl acetate") is used as the query keyword and sent to a preset chemical structure database (such as the CIR database) via API interface or web page request. After receiving the query request, the preset chemical database performs a search and matching in its internal index. When a unique corresponding chemical substance record is matched, the preset chemical database extracts the SMILES string corresponding to the substance and returns it as the query result. Finally, the returned result is received, which yields the molecular structure formula corresponding to the monomer raw material that can be directly analyzed and calculated by the computer.

[0059] For example, the monomeric raw material isoamyl acetate, whose CAS number is known to be 123-92-2. By entering this CAS number into the PubChem database, the corresponding SMILES structural formula "CC (=O) OCCC (C) C" can be obtained. This string accurately reflects the atomic arrangement and chemical bond connections of its molecule.

[0060] It is evident that by retrieving the chemical structure database based on the unique chemical identifier of each monomer raw material, the standardized simplified molecular linear input standardized structural expression can be accurately obtained as the molecular structural formula, thus realizing the accurate conversion of the physical identity of the monomer raw material into standardized digital structural information and avoiding ambiguity caused by differences in compound naming.

[0061] A3. Perform molecular descriptor calculation on the molecular structure formula corresponding to each monomer raw material to obtain the multidimensional feature vector corresponding to each monomer raw material.

[0062] In the specific implementation of A3, the step of calculating the molecular descriptor for the molecular structure formula corresponding to each monomer raw material to obtain the multidimensional feature vector corresponding to each monomer raw material includes: performing multidimensional analysis on the molecular structure formula corresponding to each monomer raw material to obtain multiple types of numerical feature parameters corresponding to each monomer raw material; calculating the molecular structure formula corresponding to each monomer raw material according to the effective carbon number calculation rules to obtain the effective carbon number descriptor corresponding to each monomer raw material; and arranging and combining the multiple types of numerical feature parameters and the effective carbon number descriptor corresponding to each monomer raw material according to a preset order to obtain the multidimensional feature vector corresponding to each monomer raw material.

[0063] Molecular descriptor computation refers to the process of converting molecular structural formulas into numerical characteristic parameters using computational chemistry techniques.

[0064] Furthermore, molecular descriptors are quantitative expressions of molecular structure and physicochemical properties, reflecting characteristics such as molecular composition, shape, charge distribution, functional groups, and topological structure. By calculating molecular descriptors, each monomeric raw material is represented as a vector composed of multiple numerical features, i.e., a multidimensional feature vector.

[0065] Among them, the multidimensional feature vector refers to a multidimensional numerical combination composed of the basic physicochemical descriptor, ring structure descriptor, functional group count descriptor, molecular fingerprint descriptor, and effective carbon number descriptor of a single monomer raw material arranged in sequence.

[0066] The multi-dimensional analysis refers to using cheminformatics tools (such as RDKit) to deconstruct and extract features from molecular structures from different structural granularities and property perspectives. The various types of numerical feature parameters may include, but are not limited to, basic physicochemical descriptors, ring system structure descriptors, functional group count descriptors, and molecular fingerprint descriptors.

[0067] Specifically, the process of calculating the basic physicochemical descriptor involves extracting the molecular weight, lipid-water partition coefficient, topological polar surface area, number of hydrogen bond acceptors, number of hydrogen bond donors, number of rotatable bonds, and number of heavy atoms of the molecule using RDKit.

[0068] The basic physicochemical descriptors include the following: Molecular weight: the sum of the masses of all atoms in the molecule, reflecting the size of the molecule; Lipid-water partition coefficient: estimates the partition balance of the molecule in the lipid and aqueous phases, affecting the molecule's response in the detector; Topological polar surface area: the size of the polar regions on the molecule's surface, affecting the interaction between the molecule and its environment; Number of hydrogen bond acceptors and hydrogen bond donors: reflect the molecule's ability to form hydrogen bonds, affecting molecular activity and detection characteristics; Number of rotatable bonds: the number of freely rotatable single bonds in the molecule, affecting the conformational flexibility of the molecule; Number of heavy atoms: the number of non-hydrogen atoms in the molecule, affecting molecular complexity.

[0069] Specifically, the process of calculating the ring system descriptor involves identifying the ring structures in the molecule using RDKit and extracting the number of aromatic rings and aliphatic rings.

[0070] Among them, the number of aromatic rings and the number of aliphatic rings reflect the number of different types of ring structures in the molecule, which affect molecular stability and detection response.

[0071] Specifically, the process of calculating the functional group count descriptor involves substructure matching of the molecular structure using RDKit, and counting the number of ester groups, aldehyde groups, ketone groups, hydroxyl groups, ether bonds, carboxyl groups, lactone rings, methyl groups, carbon-carbon double bonds, benzene rings, furan rings, pyridine rings, nitro groups, and amino groups.

[0072] Specifically, the process of calculating the molecular fingerprint descriptor involves using RDKit to calculate a Morgan fingerprint with a radius of 2 and a bit length of 1024, and then counting the number of activated sites in the Morgan fingerprint.

[0073] Among them, Morgan fingerprint is a high-dimensional binary feature based on the molecular environment, with a radius of 2 and a bit depth of 1024. It calculates the number of activation sites as a feature to reflect the local structural features of the molecule.

[0074] Among them, the effective carbon number calculation rule is an empirical quantitative rule established based on the response mechanism of the gas chromatography flame ionization detector (FID).

[0075] Therefore, the effective carbon number descriptor is a value obtained by empirically correcting for the actual chemical environment of carbon atoms in a compound, based on the response mechanism of a gas chromatographic flame ionization detector (FID). Since FID primarily responds to carbon-containing compounds, and the degree of response of carbon atoms varies under different chemical environments, the effective carbon number can more accurately reflect the contribution of a substance to the electrical signal in the detector.

[0076] Specifically, the effective carbon number calculation rule includes: First, determining the initial count value, i.e., counting the total number of carbon atoms in the molecular structure formula, which is used as the initial count value for the effective carbon number; Second, identifying deduction items and deduction amounts, i.e., traversing the molecular structure formula, identifying specific functional groups or heteroatom environments, and performing numerical deductions according to the following fixed standards: deducting 1.0 for each carbonyl carbon (including carbonyl carbons in aldehyde, ketone, carboxyl, and ester groups), deducting 0.6 for each hydroxyl group, deducting 1.0 for each oxygen atom in an ether bond, deducting 0.6 for each primary and secondary amine, deducting 0.6 for each nitro group, deducting 1.0 for each halogen atom, and deducting 1.0 for each carbon atom in a nitrile group; Finally, subtracting the sum of the deduction amounts corresponding to all identified deduction items from the initial count value, the deducted value is the effective carbon number descriptor.

[0077] The preset order refers to a pre-defined feature arrangement index rule. Therefore, the preset order arrangement and combination refers to the vertical concatenation of various numerical feature parameters and effective carbon number descriptors according to a fixed feature column order to form a structured vector sequence.

[0078] For example, the specific arrangement rules of the preset order are as follows: Basic physicochemical descriptors take priority (positions 1-7): Parameters reflecting the overall size and macroscopic properties of the molecule are arranged first, in the following order: molecular weight (position 1), lipid-water partition coefficient (position 2), topological polar surface area (position 3), number of hydrogen bond acceptors (position 4), number of hydrogen bond donors (position 5), number of rotatable bonds (position 6), and number of heavy atoms (position 7); Ring system descriptors follow next (positions 8-9): Parameters reflecting the ring formation of the skeleton are arranged in the following order: number of aromatic rings (position 8) and number of aliphatic rings (position 9); Functional group count descriptors follow immediately (positions 10-23): Parameters are arranged according to the polarity of the functional groups and common chemical... The categories are arranged in a fixed order: ester group (position 10), aldehyde group (position 11), ketone group (position 12), hydroxyl group (position 13), ether bond (position 14), carboxyl group (position 15), lactone ring (position 16), methyl group (position 17), carbon-carbon double bond (position 18), benzene ring (position 19), furan ring (position 20), pyridine ring (position 21), nitro group (position 22), amino group (position 23); molecular fingerprint feature parameter (position 24): the number of Morgan fingerprint activation sites; effective carbon number descriptor (position 25): the effective carbon number calculated based on the detection mechanism is placed at the end of the vector as a mechanism-level compensation feature for the overall feature vector. During permutation and combination, the calculated values ​​are filled in according to the above preset order; if a monomer does not contain a certain functional group, 0 is entered. For example, for d-decyl lactone, if its lactone ring count is 1, then fill in position 16; if its aromatic ring count is 0, then fill in position 8; if its effective carbon count is 9.0, then fill in position 25.

[0079] The above-mentioned preset order can ensure that the multidimensional feature vectors generated by different monomer raw materials have a strict alignment relationship in the feature dimension, avoiding the model learning the wrong mapping rules due to feature misalignment.

[0080] Therefore, the multidimensional feature vector corresponding to each monomer raw material obtained by the molecular descriptor calculation of A3 contains 25 numerical features, which can reflect its chemical structure and properties.

[0081] For example, using the molecular structure CCCCCCC1CCC(=O)O1 for butyl-decalactone, analytical calculations using the cheminformatics tool RDKit yielded the following results: molecular weight 170.25, lipid-water partition coefficient 2.66, topological polar surface area 26.30, number of hydrogen bond acceptors 2, number of hydrogen bond donors 0, number of rotatable bonds 5, number of heavy atoms 12, number of aromatic rings 0, number of aliphatic rings 1, number of ester groups 0, number of aldehyde groups 0, number of ketone groups 0, number of hydroxyl groups 0, number of ether bonds 0, number of carboxyl groups 0, and number of lactone rings 1. The methyl group is 1, the carbon-carbon double bond is 0, the benzene ring is 0, the furan ring is 0, the pyridine ring is 0, the nitro group is 0, the amino group is 0, and the number of Morgan fingerprint activation sites is 27. According to the effective carbon number descriptor calculation rules, the butyl-decyl lactone molecule contains 10 carbon atoms, with an initial count value of 10. One of them is the carbonyl carbon in the lactone ring, which is deducted by 1.0. The ether oxygen in the lactone ring is part of the ester and is not deducted additionally. There are no other groups that need to be deducted, resulting in an effective carbon number of 9.0. The above 25 values ​​are combined to form the 25-dimensional feature vector corresponding to the butyl-decyl lactone.

[0082] As can be seen, this embodiment transforms complex chemical structural information into multidimensional feature vectors that can be recognized and processed by machine learning models by systematically calculating molecular descriptors of monomer raw material molecular structures.

[0083] A4. Based on the multi-dimensional feature vectors corresponding to each monomer raw material, construct the target feature matrix.

[0084] The target feature matrix refers to a two-dimensional numerical set composed of the multi-dimensional feature vectors corresponding to all monomer raw materials arranged in rows. In this matrix, each row represents a monomer raw material and each column represents a feature dimension.

[0085] In practice, the multidimensional feature vector corresponding to each monomer raw material is taken as a row, and all multidimensional feature vectors are vertically concatenated according to the arrangement order of the monomer raw materials to form a two-dimensional numerical matrix with the number of rows equal to the total number of monomer raw materials and the number of columns equal to the feature dimension. This two-dimensional numerical matrix is ​​determined as the target feature matrix.

[0086] For example, assuming there are 409 monomer raw materials, each with a 25-dimensional feature vector, the target feature matrix would be 409×25 in size. This matrix integrates the molecular information of all samples.

[0087] Therefore, the implementation process of A1-A4 can be understood through the following complete case: For ethyl 2-methylbutyrate (CAS: 629-78-7), a common monomer raw material in 409 electronic atomizing liquids, gas chromatography with flame ionization detector (GC-FID) was used for analysis. Using n-heptadecane as an internal standard, its relative correction factor f was measured to be 0.65313 under the same chromatographic conditions. The relative correction factor 0.65313 of ethyl 2-methylbutyrate was associated with and stored. The canonical simplified molecular linear input canonical structural expression of this compound was retrieved from the chemical identifier resolver database as CCOC(=O)C(C)CC. The cheminformatics tool RDKit was used to analyze and calculate the structure, yielding a molecular weight of 130.19, a lipid-water partition coefficient of 1.60, a topological polar surface area of ​​26.30, 2 hydrogen bond acceptors, 0 hydrogen bond donors, 4 rotatable bonds, 9 heavy atoms, 0 aromatic rings, 0 aliphatic rings, 1 ester group, and a number of aldehyde groups. The initial count is 6, with 1.0 for ester carbonyl groups, 0 for ketone groups, 0 for hydroxyl groups, 0 for ether bonds, 0 for carboxyl groups, 0 for lactone rings, 2 for methyl groups, 0 for carbon-carbon double bonds, 0 for benzene rings, 0 for furan rings, 0 for pyridine rings, 0 for nitro groups, 0 for amino groups, and 18 for Morgan fingerprint activation sites. According to the effective carbon number descriptor calculation rules, this molecule contains 6 carbon atoms, with an initial count of 6. One of these is an ester carbonyl group, minus 1.0. There are no hydroxyl, ether oxygen, amino, nitro, halogen, or nitrile groups, therefore the effective carbon number is 6 - 1.0 = 5.0. These 25 values ​​are combined to form a 25-dimensional feature vector corresponding to ethyl 2-methylbutyrate. Using this method, all 409 monomer raw materials in the training set are calculated one by one to form their respective 25-dimensional feature vectors. These 409 25-dimensional feature vectors are then vertically concatenated by rows to construct a 409-row, 25-column target feature matrix.

[0088] A5. Divide the target feature matrix and the relative correction factor corresponding to each monomer raw material into a dataset according to a preset ratio to obtain a training set and a test set.

[0089] In the specific implementation of A5, the step of dividing the target feature matrix and the relative correction factor corresponding to each monomer raw material into a training set and a test set according to a preset ratio includes: aligning and combining the feature vectors of each row in the target feature matrix with the relative correction factor corresponding to each monomer raw material to obtain a set of sample pairs; and dividing the set of sample pairs according to the preset ratio to obtain the training set and the test set.

[0090] The preset ratio refers to the pre-set numerical ratio used to divide the overall sample set, such as 8:2 or 7:3, and is not a unique limitation here.

[0091] Dataset partitioning refers to the process of randomly splitting the overall sample containing feature data and target variables into two non-overlapping subsets according to a preset ratio.

[0092] The training set refers to a subset of samples used to train the mapping relationship between features and target variables in a machine learning model.

[0093] The test set refers to a subset of samples used to evaluate the generalization performance and prediction accuracy of the trained machine learning model. The samples in this subset do not participate in the model training process.

[0094] The alignment combination refers to aligning the target feature matrix based on the unique identifier of the monomer raw material (such as row index or CAS number). eigenvectors of rows With the Relative correction factor for each monomer raw material Perform horizontal splicing or association binding to form a data whole.

[0095] Specifically, iterate through the index of each row of the target feature matrix. ( From 1 to Extract the first eigenvectors of rows At the same time, read the first Relative correction factor for each monomer raw material Combine the two into an ordered pair After the traversal is complete, all The set of ordered pairs constitutes the sample pair set. In this sample pair set, each element contains a complete characteristic representation of a monomeric raw material and its corresponding actual measured target value.

[0096] Specifically, the statistical sample is the total number of samples in the set. The number of training set samples is calculated based on a preset ratio (e.g., 8:2). and the number of test set samples Random sampling without replacement: Using a random sampling algorithm (such as randomly shuffling the index or sampling without replacement), samples are drawn from the set of sample pairs. Each sample pair. During this process, it is necessary to ensure that the sample pairs are maintained. The integrity of the feature vector means that the feature vector and its corresponding relative correction factor must be extracted or retained simultaneously, and cannot be separated; furthermore, the extracted... The sample pairs are combined into a training set, and the remaining unsampled samples are... Each sample pair is combined into a test set; the training set and test set are decoupled, and the feature vectors in the training set are re-stacked into a training feature submatrix. The relative correction factor is extracted as a subset of the training target variable. Similarly, extract the test feature sub-matrix from the test set. and test target variable subset .

[0097] For example, for a target feature matrix consisting of 409 rows and 25 columns, composed of 409 monomer raw materials, and corresponding 409 measured relative correction factors, each row of feature vectors is aligned and associated with the relative correction factor of the corresponding monomer, forming 409 overall samples. A preset ratio of 8:2 is set, resulting in approximately 327 training samples and approximately 82 test samples. Based on this allocation, these 409 overall samples are randomly sampled without replacement. Approximately 327 samples constitute the training set, and the remaining approximately 82 samples constitute the test set. The training set contains a subset of training target variables consisting of approximately 327 rows and 25 columns of training feature sub-matrices and approximately 327 relative correction factors, used for subsequent training of each base learner model. The test set contains a subset of test target variables consisting of approximately 82 rows and 25 columns of test feature sub-matrices and approximately 82 relative correction factors, used only for final performance evaluation of the trained model and not involved in any model training process.

[0098] Therefore, by randomly dividing the sample set consisting of the feature matrix and the relative correction factor according to a preset ratio, mutually independent training and test sets are constructed, ensuring the physical isolation between model training data and performance evaluation data, and effectively preventing data leakage and overfitting artifacts during the model evaluation process.

[0099] A6. Train each initial model in the initial model set according to the training set to obtain multiple base models, wherein the base models are each initial model after training.

[0100] For details on the implementation of A6, please refer to B1-B2 below.

[0101] In one embodiment, training each initial model in the initial model set based on the training set to obtain multiple base models includes: B1. Determine the target feature matrix in the training set as the input variable and the corresponding relative correction factor in the training set as the target variable.

[0102] The input variables refer to the set of independent variable data used to train the machine learning model, that is, the target feature matrix composed of the multidimensional feature parameters of each monomer raw material in the training set.

[0103] The target variable refers to the set of dependent variable data that the machine learning model needs to predict, namely the relative correction factor values ​​of each monomer raw material measured in the training set.

[0104] Specifically, the training set contains several monomer raw material samples, each corresponding to a multidimensional molecular descriptor vector and a measured relative correction factor. From the partitioned training set, the set of multidimensional molecular descriptor vectors corresponding to all training samples is extracted, arranged in sample order, to form the training feature data set, which serves as the input variable X. Simultaneously, the measured relative correction factors corresponding to each sample in the training set are extracted to form the target variable set, which serves as the target variable y.

[0105] For example, a training set contains 327 individual raw material samples, and the multidimensional feature vector of each sample has a dimension of 25. A training feature set containing 327 25-dimensional vectors is extracted and used as the input variable X; the corresponding 327 measured relative correction factors form the target variable set and are used as the target variable y.

[0106] B2. Input the input variables and the target variables into each initial model in the initial model set for training, so as to learn the mapping relationship between the input variables and the target variables, and obtain the multiple basic models.

[0107] The initial model set refers to the collection of untrained initial prediction models constructed using various machine learning algorithms. The initial model set may include, but is not limited to, initial models of random forests, extreme gradient boosting trees, and support vector machines.

[0108] The mapping relationship refers to the nonlinear mathematical function relationship between the multidimensional feature parameters of the input variables and the relative correction factor of the target variable.

[0109] Among them, the base model refers to the machine learning prediction model that has been optimized and learned by using training set data.

[0110] In practice, for each initial model in the initial model set, the input variable X and the target variable y are input into the initial model. The initial model iteratively learns and optimizes the mapping relationship between the input variable X and the target variable y according to its own machine learning algorithm rules. When the convergence condition set by the algorithm is met or the set number of iterations is reached, the model training is completed, and a basic model that can predict the relative correction factor based on the input features is output. The above training process is performed on all initial models in the initial model set in sequence to obtain multiple basic models of different types.

[0111] Specifically, when training the initial random forest model, the number of decision trees is set to 300, the maximum depth to 12, and the minimum number of split samples to 4. The input variable X and the target variable y are input into the initial random forest model for training to obtain the basic random forest model.

[0112] For reference Figure 3 , Figure 3 This is a ranking of the importance of the top 15 features in a random forest model. Figure 3 The diagram illustrates the importance of each molecular descriptor in the random forest model to the prediction of the relative correction factor. The horizontal axis represents feature importance, with higher values ​​indicating a greater contribution of the feature to the model's prediction results. The vertical axis lists the top 15 most important features out of the 25 molecular descriptors. Figure 3 Among the features, the effective carbon number (ECN) has the highest feature importance, reaching 0.360, indicating that the effective carbon number is the most critical factor affecting the prediction of the relative correction factor. Secondly, the feature importance is 0.249, reflecting the influence of molecular hydrophobicity and volatility on chromatographic behavior. Other important features include the number of hydrogen bond acceptors, molecular weight, and the number of Morgan fingerprint activation sites. These features reflect the structure and properties of the molecule from different perspectives and all contribute to the prediction of the relative correction factor. Therefore... Figure 3 By quantifying the importance of each feature, a scientific basis is provided for model training and feature selection, which also enhances the interpretability of the model.

[0113] Specifically, when training the initial model of the extreme gradient boosting tree, the number of decision trees is set to 300, the maximum depth is 6, and the learning rate is 0.05. The input variable X and the target variable y are input into the initial model of the extreme gradient boosting tree for training, and the basic model of the extreme gradient boosting tree is obtained.

[0114] Specifically, when training the initial support vector machine model, the kernel function is set to radial basis function, the penalty coefficient C is 10, the kernel function coefficient gamma is "scale", the insensitive loss coefficient epsilon is 0.05, and before inputting the input variable X into the initial support vector machine model, the input variable X is standardized using a standard scaler to eliminate the influence of the difference in scale between different feature dimensions. The standardized input variable and the target variable y are then input into the initial support vector machine model for training to obtain the basic support vector machine model.

[0115] For example, using the 25-dimensional feature matrix of 327 individuals as input variable X and the 327 measured relative correction factors as target variable y, we input them into the initial random forest model. The algorithm constructs 300 decision trees based on bootstrap resampling, with a maximum depth limit of 12 for each tree and a minimum number of split samples of 4 per node. By continuously splitting nodes to minimize the variance between the predicted and true values, a basic random forest model applicable to all individuals is obtained after training. Similarly, using the same input variable X and target variable y, we input the initial extreme gradient boosting tree model and perform sequential iterative training to obtain a basic extreme gradient boosting tree model. After standardizing the input variable X, we input it into the initial support vector machine model in conjunction with the target variable y, and search for the optimal hyperplane in the high-dimensional space to obtain a basic support vector machine model. This process continues until various different types of basic models are obtained.

[0116] As can be seen, by determining the feature matrix in the training set as the input variable and the measured relative correction factor as the target variable, and inputting them into an initial model set containing various algorithms such as random forest, extreme gradient boosting tree and support vector machine for mapping relationship learning, the complementary advantages of different algorithms in feature extraction and bias-variance tradeoff are utilized. At the same time, standardization is introduced for support vector machine to eliminate the influence of dimensions, thereby accurately fitting the complex nonlinear relationship between molecular descriptor and relative correction factor, and training multiple basic models with differentiated prediction characteristics.

[0117] A7. Based on the multiple basic models and the preset integration strategy, the target integrated prediction model is obtained.

[0118] Among them, the target integrated prediction model refers to the final prediction model obtained by fusing the prediction results of multiple basic models according to preset rules.

[0119] Among them, the preset integration strategy refers to the fusion rule that integrates the prediction relative correction factors of multiple basic models to eliminate the bias of a single model.

[0120] In practice, the characteristic data of the monomer raw material to be predicted are input into multiple trained basic models, and each basic model independently outputs the corresponding prediction relative correction factor. According to the selected preset integration strategy, the multiple prediction relative correction factors are fused and calculated, and the final prediction value obtained after fusion calculation is determined as the output result of the target integrated prediction model.

[0121] In one embodiment, the preset integration strategy is a first preset integration strategy, a second preset integration strategy, or a third preset integration strategy; when the preset integration strategy is the first preset integration strategy, the first preset integration strategy is to perform an arithmetic average of the prediction relative correction factors output by the plurality of base models; or, when the preset integration strategy is the second preset integration strategy, the second preset integration strategy is to perform a weighted summation of the output prediction relative correction factors according to the weights corresponding to each base model, wherein the weights are determined by the determination coefficients of each base model in the cross-validation process; or, when the preset integration strategy is the third preset integration strategy, the third preset integration strategy is to perform fusion processing on the prediction relative correction factors output by the plurality of base models according to the meta-learner.

[0122] When the preset integration strategy is a first preset integration strategy, a second preset integration strategy, or a third preset integration strategy, it is a single preset integration strategy.

[0123] Among them, the first preset integration strategy can be a simple averaging strategy, the second preset integration strategy can be a weighted averaging strategy, and the third preset integration strategy can be a stacked integration strategy.

[0124] Specifically, when the preset integration strategy is the first preset integration strategy, the process of the first preset integration strategy to perform an arithmetic average of the prediction relative correction factors output by the multiple basic models is as follows: obtain the relative correction factor values ​​of all basic models for the same monomer raw material, perform an arithmetic average of the prediction relative correction factors, and determine the calculated arithmetic average as the output result of the target integrated prediction model.

[0125] The formula for calculating the arithmetic mean is: fintegral = (f1 + f2 + … + f3) n ) / n, f is the output of the target ensemble prediction model, f1 to f n These are the prediction relative correction factors output by the 1st to nth base models, respectively, where n is the total number of base models.

[0126] Specifically, when the preset integration strategy is the second preset integration strategy, the second preset integration strategy is to perform a weighted summation of the output prediction relative correction factors according to the weights corresponding to each base model. The weights are determined by the coefficients of determination of each base model in the cross-validation process as follows: Obtain the determination coefficient of each base model in the cross-validation process, calculate the weight corresponding to each base model based on the determination coefficient, and calculate the weighted sum of the prediction relative correction factors of the output of each base model based on the weight corresponding to each base model. The calculated weighted sum is determined as the output result of the target integrated prediction model.

[0127] The weighted summation formula is: fense = Σ(wᵢ × fᵢ), and the weight calculation formula is: wᵢ = R²ᵢ / ΣR²ⱼ, where wᵢ is the weight of the i-th base model, fᵢ is the prediction relative correction factor of the i-th base model output, R²ᵢ is the determination coefficient of the i-th base model in the cross-validation process, and ΣR²ⱼ is the sum of the determination coefficients of all base models.

[0128] Specifically, when the preset integration strategy is the third preset integration strategy, the process of fusing the predicted relative correction factors output by the multiple basic models according to the meta-learner is as follows: the meta-learner is used to perform secondary learning and fusion of the predicted relative correction factors output by the multiple basic models.

[0129] In practical implementation, out-of-bag predictions for each base model are generated through 5-fold cross-validation. The out-of-bag predictions of each base model corresponding to the same monomer are combined as meta-features, and the measured relative correction factors corresponding to each monomer are used as target variables. A meta-learner is used to train the meta-features and the target variables to obtain the trained meta-learner. During actual prediction, the prediction relative correction factors output by multiple base models are input into the trained meta-learner to obtain the output of the target ensemble prediction model. The fusion formula used by the meta-learner is: fensemble = β0 + β1×f1 + β2×f2 + … + β0 n ×f n β0 is the bias term, β1 to β n is the fusion coefficient of the predictions of each basic model obtained from training the meta-learner.

[0130] For example, three top-performing base models (random forest model, support vector machine base model, and support vector machine model with effective carbon count) are selected for independent ensemble. If the relative correction factors for the output of a certain raw material by these three base models are 1.20, 1.25, and 1.15, respectively; when using the first preset ensemble strategy, their arithmetic average is calculated as (1.20 + 1.25 + 1.15) / 3 = 1.20, and 1.20 is determined as the output result of the target ensemble prediction model; when using the second preset ensemble strategy, if the coefficients of determination (R²) of the three base models in the cross-validation process are 0.90, 0.85, and 0.88, respectively, and the sum of the coefficients of determination is 2.63, their weights are calculated as 0.342, 0.323, and 0.335, respectively, and the weighted sum is calculated as 0.342 × 1.20 + 0.323 × 1.25 + 0.335 × 1.15 = 1.198, and 1.198 is determined as the output of the target ensemble prediction model. When the third preset ensemble strategy is adopted, ridge regression is used as the meta-learner. The out-of-bag prediction value generated by 5-fold cross-validation is used as the meta-feature to train the ridge regression to determine the fusion coefficient. Then, the actual outputs of the three basic models, 1.20, 1.25, and 1.15, are substituted into the formula for secondary learning and fusion. The final result is determined as the output of the target ensemble prediction model.

[0131] As can be seen, this embodiment significantly improves the prediction accuracy of the relative correction factor and the generalization ability of the model by fusing the prediction results of multiple base models using a single preset fusion strategy. The simple averaging strategy has a clear structure and high computational efficiency, and is suitable for scenarios where the performance of each base model is similar, effectively reducing the impact of individual model errors. The weighted averaging strategy makes reasonable use of the performance differences of each model, giving greater weight to the high-performing model, thereby improving the stability and reliability of the overall prediction. The stacked fusion strategy further captures the complex correlation between base models through the meta-learner, achieving more refined fusion optimization and enhancing the model's prediction ability. Therefore, the application of a single strategy is simple and highly targeted, suitable for different practical needs and data characteristics.

[0132] In one embodiment, the preset integration strategy is a combination strategy, which consists of any two or more of a first preset integration strategy, a second preset integration strategy, and a third preset integration strategy; wherein, the first preset integration strategy is to perform an arithmetic average of the prediction relative correction factors output by the plurality of base models; the second preset integration strategy is to perform a weighted summation of the output prediction relative correction factors according to the weights corresponding to each base model, wherein the weights are determined by the determination coefficients of each base model in the cross-validation process; and the third preset integration strategy is to perform a fusion processing on the prediction relative correction factors output by the plurality of base models according to the meta-learner.

[0133] Specifically, when the preset integration strategy is a combination strategy, the prediction relative correction factors output by multiple basic models are subjected to multi-level cascade fusion processing according to the execution order of the preset integration strategies included in the combination strategy, and the value obtained by the last fusion processing is determined as the output result of the target integrated prediction model.

[0134] Optionally, when the combination strategy consists of a second preset integration strategy and a third preset integration strategy, the weights are first determined based on the determination coefficients of each base model in the cross-validation process. The prediction relative correction factors output by each base model are weighted and summed using the second preset integration strategy to select the base models with better performance. Then, the meta-learner in the third preset integration strategy is used to perform secondary learning and fusion on the prediction relative correction factors output by the base models to obtain the output results of the target integrated prediction model.

[0135] Optionally, when the combination strategy consists of a first preset integration strategy and a third preset integration strategy, the first preset integration strategy is first used to perform an arithmetic average of the prediction relative correction factors output by multiple basic models to obtain a preliminary prediction result of rapid fusion; then, the meta-learner in the third preset integration strategy is used to perform fine optimization of stacking and fusion of the preliminary prediction result and the prediction relative correction factors of each basic model to obtain the output result of the target integrated prediction model.

[0136] For example, in the second case, assuming there are 5 base models, a combined strategy consisting of a second preset ensemble strategy and a third preset ensemble strategy is adopted. First, weights are assigned to the 5 base models according to their coefficients of determination during cross-validation, and a weighted sum is calculated. Based on the weighted sum and the coefficients of determination, the 3 base models with the highest coefficients of determination (such as the random forest model, the support vector machine base model, and the support vector machine model with effective carbon count) are selected. Then, only the prediction relative correction factors output by these 3 selected base models are used as input, and ridge regression is used as the meta-learner for stacked fusion of the third preset ensemble strategy. That is, the out-of-bag prediction values ​​of these 3 models are generated through 5-fold cross-validation as meta-features to train the ridge regression model. In actual prediction, the outputs of these 3 models are input into the trained ridge regression model to obtain the final target ensemble prediction model output.

[0137] As can be seen, the combined preset integration strategy in this embodiment further improves the robustness and predictive performance of the model by flexibly combining the advantages of multiple methods such as simple averaging, weighted averaging, and stacked integration. First, weighted averaging is used to quickly screen out high-performing basic models, and then stacked integration is used to deeply fuse the screening results, effectively preventing overfitting and refining the prediction results; alternatively, simple averaging is used to quickly obtain preliminary results, and then stacked integration is used for fine-tuning, balancing efficiency and accuracy. This combined strategy integrates multiple fusion mechanisms, overcomes the limitations of single models and single integration methods, and enhances the predictive accuracy and applicability of the relative correction factor of electronic atomizing liquid monomer raw materials.

[0138] In one embodiment, after obtaining the target ensemble prediction model based on the multiple base models and the preset ensemble strategy, the method further includes: predicting the target ensemble prediction model based on the test set to obtain a test relative correction factor; calculating the error between the test relative correction factor and the corresponding true relative correction factor to obtain an evaluation index of the target ensemble prediction model, the evaluation index including the coefficient of determination and the root mean square error; and determining that the target ensemble prediction model has passed the test when the coefficient of determination is greater than or equal to a first preset threshold and the root mean square error is less than or equal to a second preset threshold.

[0139] The test relative correction factor refers to the predicted value output after the test set features are input into the target integrated prediction model.

[0140] The true relative correction factor refers to the relative correction factor value determined by actual experiments on the monomer raw materials in the test set.

[0141] Among them, the evaluation index refers to the statistical index used to quantify the degree of deviation between the model's predicted value and the actual value, as well as the goodness of fit.

[0142] The coefficient of determination is a statistic that reflects the proportion of variation in the model's predicted values ​​that is explained by the actual values. It typically ranges from 0 to 1, and the closer the value is to 1, the better the model fits.

[0143] The root mean square error (RMSE) is the square root of the ratio of the sum of squares of the deviations between the predicted and actual values ​​to the number of observations. It measures the absolute dispersion of the predicted values ​​from the actual values, and the smaller the value, the higher the accuracy of the model prediction.

[0144] The first preset threshold refers to the minimum acceptable limit of the predetermined determination coefficient.

[0145] The second preset threshold refers to the maximum permissible limit of the root mean square error.

[0146] In the specific implementation, firstly, a test feature sub-matrix is ​​extracted from the test set. This test feature sub-matrix is ​​then input into the trained target ensemble prediction model. The model outputs the predicted relative correction factor corresponding to each test sample, and this predicted relative correction factor is determined as the test relative correction factor. Secondly, the measured relative correction factor corresponding to each test sample is extracted from the test set as the true relative correction factor. The test relative correction factor and the true relative correction factor of the same test sample are paired. Then, the coefficient of determination and root mean square error are calculated based on the paired data. The process for calculating the coefficient of determination is as follows: calculate the mean of the true relative correction factor, calculate the sum of squares of the differences between each true relative correction factor and its mean to obtain the total sum of squares, calculate the sum of squares of the differences between each test relative correction factor and its corresponding true relative correction factor to obtain the residual sum of squares, and subtract the ratio of the residual sum of squares to the total sum of squares from 1 to obtain the coefficient of determination. The process for calculating the root mean square error is as follows: calculate the square of the difference between each test relative correction factor and its corresponding true relative correction factor, sum all the squared differences and divide by the total number of test samples to obtain the mean square error, and take the square root of the mean square error to obtain the root mean square error. Finally, the calculated coefficient of determination is compared with the first preset threshold, and the root mean square error is compared with the second preset threshold. When the coefficient of determination is greater than or equal to the first preset threshold and the root mean square error is less than or equal to the second preset threshold, the goodness of fit and prediction accuracy of the target integrated prediction model are determined to meet the requirements, and the target integrated prediction model is determined to have passed the test; otherwise, the model test is determined to have failed, and the model parameters or integration strategy need to be readjusted. For example, for a test set consisting of approximately 82 monomeric raw materials, the 25-dimensional feature matrix of these 82 monomers is input into the target ensemble prediction model to obtain 82 test relative correction factors; the experimental measured values ​​corresponding to these 82 monomers are extracted as true relative correction factors; the coefficient of determination R² is calculated to be 0.85, and the root mean square error RMSE is 0.06; a first preset threshold is set to 0.80, and a second preset threshold is set to 0.07; since R²=0.85≥0.80 and RMSE=0.06≤0.07, the goodness of fit of the target ensemble prediction model is determined to meet the standard of explaining more than 80% of the variation, and the absolute prediction error is controlled within 0.07, thus the target ensemble prediction model is determined to have passed the test and has acceptable prediction accuracy.

[0147] As can be seen, this embodiment uses an independent test set to validate the target ensemble prediction model, which can objectively and accurately evaluate the model's predictive performance and ensure that the model has good generalization ability and stability. Furthermore, using indicators such as the coefficient of determination and root mean square error for multi-dimensional evaluation effectively avoids model overfitting and prediction bias, improving the reliability and accuracy of the model in practical applications.

[0148] The following four case studies illustrate this solution. It's important to note that the target ensemble prediction model described in these cases is built upon at least two pre-trained foundational machine learning models. These foundational models include, but are not limited to, Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost). Each foundational model independently completes its training using multidimensional molecular descriptors of spice monomers and their corresponding relative correction factors.

[0149] Based on this, a pre-defined fusion strategy is used to fuse the individual prediction outputs of each base model to form a comprehensive prediction result. The pre-defined fusion strategy can be a single strategy, specifically including: a first pre-defined fusion strategy—simple arithmetic mean method, which takes the arithmetic mean of the prediction relative correction factors of each base model output; a second pre-defined fusion strategy—weighted average method, which uses the coefficient of determination (R²) of each base model in the cross-validation process as weights to sum the prediction results in a weighted manner; and a third pre-defined fusion strategy—stacking, which uses a meta-learner (such as linear regression or ridge regression) to perform secondary learning and fusion of the prediction results from multiple base models.

[0150] Case 1 uses γ-Decalactone (CAS No.: 706-14-9), a peach-scented lactone commonly used in electronic atomizing fluids, as an example to verify the predictive performance and practical application effect of the target integrated prediction model.

[0151] First, the molecular structure of β-decanolactone was retrieved from the Chemical Identifier Resolver (CIR) database, yielding its SMILES expression as "CCCCCCC1CCC(=O)O1". Subsequently, the molecular descriptor was calculated using the RDKit cheminformatics tool. Specific results included: molecular weight (MolWt) of 170.25, lipid-water partition coefficient (MolLogP) of 2.66, topological polar surface area (TPSA) of 26.30, number of hydrogen bond acceptors (HBA) of 2, number of hydrogen bond donors (HBD) of 0, number of rotatable bonds (NumRotatableBonds) of 5, number of heavy atoms (NumHeavyAtoms) of 12, number of aromatic rings (NumAromaticRings) of 0, number of aliphatic rings (NumAliphaticRings) of 1, number of lactone rings (FGlactone) of 1, number of methyl rings (FGmethyl) of 1, and number of Morgan fingerprint activation sites (MorganOnBits) of 27. According to the effective carbon number (ECN) calculation rules, this molecule contains 10 carbon atoms. Subtracting one carbon atom from the lactone carbonyl group, and without additional subtraction of the ether oxygen atom within the lactone ring, the ECN value is 9.0. The above molecular descriptors constitute the multidimensional feature vector of this fragrance, serving as the prediction input.

[0152] Based on the architecture of the aforementioned target ensemble prediction model, the multidimensional molecular descriptor of gamma-decanolide is input into each trained basic model for prediction, and the corresponding relative correction factor prediction values ​​are obtained. Table 1 below compares the individual prediction results of each basic model with the comprehensive prediction results fused by the first preset ensemble strategy and the measured values.

[0153] Table 1 Comparison of multi-model prediction results for β-decanolide As shown in Table 1, the base SVM model alone significantly overestimates the relative correction factor for β-decanolactone (error +8.85%), while the model combining Random Forest with effective carbon number features has a smaller prediction error (+2.07%). In this embodiment, the target ensemble prediction model adopts the aforementioned first preset ensemble strategy (simple arithmetic mean method), which takes the arithmetic mean of the prediction results of the three base models: RF+ECN, SVM(Base), and SVM+ECN: (0.542+0.578+0.561) / 3=0.560. By fusing the prediction results of each base model using this ensemble strategy, the relative error is successfully reduced to +5.46%, which is significantly better than the prediction result of the single SVM model.

[0154] Therefore, this result shows that the ensemble strategy can effectively combine the complementary advantages of different algorithm models in capturing molecular structure features, thereby improving the accuracy and robustness of predictions.

[0155] Case 2: In this embodiment, we take the monocyclic terpene compound menthol (CAS No.: 89-78-1) as the object to verify the predictive performance and practical application effect of the target integrated prediction model.

[0156] First, the molecular structure of menthol was retrieved from the Chemical Identifier Resolver (CIR) database, yielding its SMILES expression as "CC1CCC(C(C)C)C(O)C1". Subsequently, the molecular descriptor was calculated using the RDKit cheminformatics tool. Specific results include: molecular weight (MolWt) of 156.27, lipid-water partition coefficient (MolLogP) of 2.44, topological polar surface area (TPSA) of 20.23, number of hydrogen bond acceptors (HBA) of 1, number of hydrogen bond donors (HBD) of 1, number of rotatable bonds (NumRotatableBonds) of 1, number of heavy atoms (NumHeavyAtoms) of 11, number of aromatic rings (NumAromaticRings) of 0, number of aliphatic rings (NumAliphaticRings) of 1, number of ester groups (FGester) of 0, and aldehydes... The molecular descriptors are as follows: FGaldehyde = 0, FGketone = 0, FGhydroxyl = 1, FGether = 0, FGcarboxylicacid = 0, FGlactone = 0, FGmethyl = 3, FGdoublebond = 0, FGbenzene = 0, FGfuran = 0, FGpyridine = 0, FGnitro = 0, FGamine = 0, and the number of Morgan fingerprint activation sites (MorganOnBits) is 21. According to the effective carbon number (ECN) calculation rules, this molecule contains 10 carbon atoms, including one hydroxyl group (subtracting 0.6), and no carbonyl, etheroxy, amino, nitro, halogen, or nitrile groups. Therefore, the ECN value is 10 - 0.6 = 9.4. These molecular descriptors constitute the multidimensional feature vector of this fragrance, serving as the prediction input.

[0157] Based on the architecture of the aforementioned target ensemble prediction model, the multidimensional molecular descriptor of menthol is input into each trained basic model for prediction, and the corresponding relative correction factor prediction values ​​are obtained. Table 2 shows a comparison between the individual prediction results of each basic model, the comprehensive prediction results fused by the first preset ensemble strategy, and the measured values.

[0158] Table 2 Comparison of multi-model prediction results for menthol. As shown in Table 2, the model predictions of the Random Forest combined with the effective carbon number feature are slightly lower (error -2.19%), while the predictions of the Support Vector Machine series models are slightly higher (error +1.82% to +2.67%). In this embodiment, the target ensemble prediction model adopts the first preset ensemble strategy (simple arithmetic mean method) mentioned above, that is, the arithmetic mean of the prediction results of the three basic models RF+ECN, SVM(Base) and SVM+ECN is taken: (0.805+0.845+0.838) / 3=0.829. By combining the lower predictions of Random Forest and the higher predictions of Support Vector Machine with this ensemble strategy, the relative error is successfully offset and reduced to only +0.73%.

[0159] Therefore, this result shows that the integrated strategy can effectively combine the complementary advantages of different models in capturing molecular structure features. For cyclic alcohols such as menthol, it can improve the prediction accuracy, further demonstrating the applicability and robustness of the relative correction factor prediction method for monomer raw materials in electronic atomizing liquids in the quantitative analysis of flavor monomers in electronic atomizing liquids.

[0160] Case 3 uses vanillin (CAS No.: 121-33-5) in electronic atomizing liquid as an example to verify the predictive performance and practical application effect of the target integrated prediction model for multifunctional aromatic compounds.

[0161] First, the molecular structure of vanillin was retrieved from the CIR database, yielding its SMILES expression as "COc1cc(C=O)ccc1O". Subsequently, the RDKit cheminformatics tool was used to calculate the molecular descriptor for this molecule. Specific results included: molecular weight (MolWt) of 152.15, lipid-water partition coefficient (MolLogP) of 1.21, topological polar surface area (TPSA) of 46.53, number of hydrogen bond acceptors (HBA) of 3, number of hydrogen bond donors (HBD) of 1, number of rotatable bonds (NumRotatableBonds) of 2, number of heavy atoms (NumHeavyAtoms) of 11, number of aromatic rings (NumAromaticRings) of 1, number of aliphatic rings (NumAliphaticRings) of 0, number of aldehyde groups (FG_aldehyde) of 1, number of hydroxyl groups (FG_hydroxyl) of 1, number of ether bonds (FG_ether) of 1, number of methyl groups (FG_methyl) of 1, number of benzene rings (FG_benzene) of 1, and number of Morgan fingerprint activation sites (MorganOnBits) of 25, etc. According to the effective carbon number (ECN) calculation rules, this molecule contains 8 carbon atoms. One is an aldehyde carbonyl group (deduct 1.0), one is a phenolic hydroxyl group (deduct 0.6), and one is a methoxy ether bond (deduct 1.0). Therefore, the ECN value is 8 - 1.0 - 0.6 - 1.0 = 5.4. The above molecular descriptors constitute the multidimensional feature vector of this fragrance, serving as the prediction input.

[0162] Based on the architecture of the aforementioned target ensemble prediction model, the multidimensional molecular descriptor of vanillin is input into each trained basic model for prediction, and the corresponding relative correction factor prediction values ​​are obtained. Table 3 below compares the individual prediction results of each basic model with the comprehensive prediction results fused by the first preset ensemble strategy and the measured values.

[0163] As shown in Table 3, vanillin, as an aromatic aldehyde compound, contains multiple functional groups such as methoxy and phenolic hydroxyl groups, resulting in a relatively complex structure. The basic support vector machine model alone overestimates the relative correction factor for vanillin to some extent (error +4.62%), while the SVM+ECN model incorporating effective carbon number features shows a slight underestimation (error -1.20%). The random forest model incorporating effective carbon number features has a prediction error of +2.22%.

[0164] In this embodiment, the target ensemble prediction model employs the aforementioned first preset ensemble strategy (simple arithmetic mean method), which involves taking the arithmetic mean of the prediction results from the three base models: RF+ECN, SVM(Base), and SVM+ECN: (0.598+0.612+0.578) / 3=0.596. By fusing the prediction results from each base model using this ensemble strategy, the relative error is successfully reduced to +1.88%, significantly better than the overestimation results of a single support vector machine base model, and the error is controlled at an extremely low level. Therefore, by introducing effective carbon number features and combining them with the ensemble strategy, the target ensemble prediction model also exhibits excellent prediction accuracy and good generalization ability for multifunctional aromatic compounds.

[0165] Case 4 uses cinnamaldehyde (CAS No.: 104-55-2) in electronic atomizing liquid as an example to verify the predictive performance and practical application effect of the target integrated prediction model for α,β-unsaturated aldehydes containing carbon-carbon double bonds and aromatic rings.

[0166] First, the molecular structure of cinnamaldehyde was retrieved from the CIR database, yielding its SMILES expression as "O=CC=Cc1ccccc1". Subsequently, the RDKit cheminformatics tool was used to calculate the molecular descriptor for this molecule. Specific results included: molecular weight (MolWt) of 132.16, lipid-water partition coefficient (MolLogP) of 1.90, topological polar surface area (TPSA) of 17.07, number of hydrogen bond acceptors (HBA) of 1, number of hydrogen bond donors (HBD) of 0, number of rotatable bonds (NumRotatableBonds) of 2, number of heavy atoms (NumHeavyAtoms) of 10, number of aromatic rings (NumAromaticRings) of 1, number of aliphatic rings (NumAliphaticRings) of 0, number of aldehyde groups (FG_aldehyde) of 1, number of carbon-carbon double bonds (FG_double_bond) of 1, number of benzene rings (FG_benzene) of 1, and number of Morgan fingerprint activation sites (MorganOnBits) of 18, etc. According to the effective carbon number (ECN) calculation rules, this molecule contains 9 carbon atoms, one of which is an aldehyde carbonyl group (subtract 1.0). There are no hydroxyl, ether oxygen, amino, nitro, halogen, or nitrile groups, so the ECN value is 9 - 1.0 = 8.0. The above molecular descriptors constitute the multidimensional feature vector of this fragrance, serving as the prediction input.

[0167] Based on the architecture of the aforementioned target ensemble prediction model, the multidimensional molecular descriptor of cinnamaldehyde is input into each trained basic model for prediction, and the corresponding relative correction factor prediction values ​​are obtained. Table 4 below compares the individual prediction results of each basic model with the comprehensive prediction results fused by the first preset ensemble strategy and the measured values.

[0168] As shown in Table 4, cinnamaldehyde, a typical α,β-unsaturated aromatic aldehyde, contains carbon-carbon double bonds and an aromatic ring structure. The basic support vector machine model alone significantly overestimates the relative correction factor for cinnamaldehyde (error +5.54%), while the random forest model combined with the effective carbon number feature shows a slight underestimation (error -2.40%). The prediction error of the support vector machine model combined with the effective carbon number feature is +1.95%.

[0169] Therefore, the target ensemble prediction model adopted the first pre-set ensemble strategy (simple arithmetic mean method), which involves taking the arithmetic mean of the prediction results of the three basic models: RF+ECN, SVM(Base), and SVM+ECN: (0.652+0.705+0.681) / 3=0.679. By fusing the prediction results of each basic model using this ensemble strategy, the relative error was successfully reduced significantly to +1.65%, far less than the allowable error range of 10%. Therefore, by introducing the effective carbon number feature and combining it with the ensemble strategy, the target ensemble prediction model also has accurate prediction ability and good stability for complex aldehydes containing carbon-carbon double bonds and aromatic rings.

[0170] This embodiment obtains the molecular structure of the monomer raw material to be detected and calculates the corresponding molecular descriptor. It then uses a target integrated prediction model to accurately predict the relative correction factor of the monomer. Quantitative analysis of the monomer raw material can be completed without the need for standards, reducing experimental costs and time consumption. Furthermore, by combining the predicted relative correction factor with quantitative detection parameters, the concentration of the monomer raw material to be detected can be accurately calculated, improving the efficiency and accuracy of quantitative detection.

[0171] It should be noted that in the above embodiments, there is no necessarily a certain order between the steps. Those skilled in the art can understand from the description of the embodiments of this application that the above steps may have different execution orders in different embodiments, that is, they may be executed in parallel or in turn, etc.

[0172] As another aspect of the embodiments of this application, this application provides a device for predicting the relative correction factor of monomer raw materials in electronic atomizing liquids. The device can be a software module, comprising several instructions stored in a memory. A processor can access the memory and execute the instructions to complete the method for predicting the relative correction factor of monomer raw materials in electronic atomizing liquids as described in the various embodiments above.

[0173] See Figure 4 , Figure 4 This is a schematic diagram of a device for predicting the relative correction factor of monomer raw materials in electronic atomizing liquid, provided in an embodiment of this application. Figure 4 As shown, the relative correction factor prediction device 400 for monomer raw materials in the electronic atomizing liquid includes: The acquisition unit 401 is used to acquire the target molecular structure formula of the monomer raw material to be detected and the quantitative detection parameters of the monomer raw material to be detected. The calculation unit 402 is used to perform molecular descriptor calculation on the target molecular structure to obtain the target molecular descriptor; The prediction unit 403 is used to input the target molecule descriptor into the target integrated prediction model for prediction to obtain the target relative correction factor. The target relative correction factor is used to perform quantitative calculation with the quantitative detection parameters of the monomer raw material to be detected to obtain the target concentration of the monomer raw material to be detected.

[0174] This embodiment obtains the molecular structure of the monomer raw material to be detected and calculates the corresponding molecular descriptor. It then uses a target integrated prediction model to accurately predict the relative correction factor of the monomer. Quantitative analysis of the monomer raw material can be completed without the need for standards, reducing experimental costs and time consumption. Furthermore, by combining the predicted relative correction factor with quantitative detection parameters, the concentration of the monomer raw material to be detected can be accurately calculated, improving the efficiency and accuracy of quantitative detection.

[0175] It should be noted that the aforementioned device for predicting the relative correction factor of monomeric raw materials in electronic atomizing liquids can execute the method for predicting the relative correction factor of monomeric raw materials in electronic atomizing liquids provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in the embodiments of the device for predicting the relative correction factor of monomeric raw materials in electronic atomizing liquids can be found in the method for predicting the relative correction factor of monomeric raw materials in electronic atomizing liquids provided in the embodiments of this application.

[0176] This application also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the method for predicting the relative correction factor of monomer raw materials in electronic atomizing liquid as described in the foregoing embodiments.

[0177] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0178] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A method for predicting the relative correction factor of monomer raw materials in electronic atomizing liquids, characterized in that, The method includes: Obtain the target molecular structure of the monomer raw material to be tested and the quantitative detection parameters of the monomer raw material to be tested; Molecular descriptor calculation is performed on the target molecule structure to obtain the target molecule descriptor; The target molecule descriptor is input into the target integrated prediction model for prediction to obtain the target relative correction factor. The target relative correction factor is used to perform quantitative calculation with the quantitative detection parameters of the monomer raw material to be detected to obtain the target concentration of the monomer raw material to be detected.

2. The method according to claim 1, characterized in that, The target ensemble prediction model includes at least two pre-trained base models and a preset ensemble strategy. The step of inputting the target molecule descriptor into the target ensemble prediction model for prediction to obtain the target relative correction factor includes: The target molecule descriptor is input into each base model for prediction, and the first relative correction factor output by each base model is obtained. The first relative correction factor output by each basic model is integrated according to the preset integration strategy to obtain the target relative correction factor.

3. The method according to claim 1, characterized in that, Before inputting the target molecule descriptor into the target ensemble prediction model for prediction to obtain the target relative correction factor, the method further includes: Obtain an initial dataset under the same chromatographic conditions, the initial dataset including multiple monomer feedstocks and the relative correction factor corresponding to each monomer feedstock; Obtain the molecular structural formula corresponding to each of the monomer raw materials; Molecular descriptors are calculated for the molecular structural formulas corresponding to each monomer raw material to obtain multidimensional feature vectors for each monomer raw material. Based on the multi-dimensional feature vectors corresponding to each monomer raw material, a target feature matrix is ​​constructed; The target feature matrix and the relative correction factor corresponding to each monomer raw material are divided into a training set and a test set according to a preset ratio. Training is performed on each initial model in the initial model set based on the training set to obtain multiple base models, wherein each initial model is trained. The target integrated prediction model is obtained based on the multiple basic models and the preset integration strategy.

4. The method according to claim 3, characterized in that, The preset integration strategy is either a first preset integration strategy, a second preset integration strategy, or a third preset integration strategy. When the preset integration strategy is a first preset integration strategy, the first preset integration strategy is to perform an arithmetic average of the prediction relative correction factors output by the multiple basic models; or... When the preset integration strategy is the second preset integration strategy, the second preset integration strategy is to perform a weighted summation of the output prediction relative correction factors according to the weights corresponding to each base model, wherein the weights are determined by the determination coefficients of each base model in the cross-validation process; or... When the preset integration strategy is the third preset integration strategy, the third preset integration strategy is to fuse the prediction relative correction factors output by the meta-learner from the multiple basic models.

5. The method according to claim 3, characterized in that, The preset integration strategy is a combination strategy, which consists of any two or more of the first preset integration strategy, the second preset integration strategy, and the third preset integration strategy; wherein, the first preset integration strategy is to perform an arithmetic average of the prediction relative correction factors output by the multiple base models; the second preset integration strategy is to perform a weighted summation of the output prediction relative correction factors according to the weights corresponding to each base model, wherein the weights are determined by the determination coefficients of each base model in the cross-validation process; and the third preset integration strategy is to perform a fusion processing on the prediction relative correction factors output by the multiple base models according to the meta-learner.

6. The method according to claim 3, characterized in that, After obtaining the target ensemble prediction model based on the multiple base models and the preset ensemble strategy, the method further includes: Based on the test set, the target integrated prediction model is used to make predictions to obtain the test relative correction factor; Error calculation is performed on the test relative correction factor and the corresponding true relative correction factor to obtain the evaluation index of the target integrated prediction model. The evaluation index includes the coefficient of determination and the root mean square error. When the determination coefficient is greater than or equal to the first preset threshold and the root mean square error is less than or equal to the second preset threshold, the target integrated prediction model is determined to have passed the test.

7. The method according to claim 3, characterized in that, The step involves training each initial model in the initial model set based on the training set to obtain multiple base models, including: The target feature matrix in the training set is determined as the input variable, and the corresponding relative correction factor in the training set is determined as the target variable. The input variables and the target variables are input into each initial model in the initial model set for training, so as to learn the mapping relationship between the input variables and the target variables and obtain the multiple basic models.

8. The method according to claim 3, characterized in that, Obtaining the molecular structural formula corresponding to each monomer raw material includes: Obtain the chemical identifier for each of the monomer raw materials; Using the chemical identifier of each monomer raw material as the query identifier, a query is performed in a preset chemical structure database to obtain the molecular structure formula corresponding to each monomer raw material.

9. The method according to claim 3, characterized in that, The step of calculating molecular descriptors for the molecular structural formulas corresponding to each monomer raw material to obtain a multidimensional feature vector for each monomer raw material includes: Multi-dimensional analysis of the molecular structure formula corresponding to each monomer raw material was performed to obtain multiple types of numerical characteristic parameters corresponding to each monomer raw material. The molecular structure of each monomer raw material is calculated according to the effective carbon number calculation rules to obtain the effective carbon number descriptor for each monomer raw material; According to a preset order, the various types of numerical feature parameters corresponding to each monomer raw material and the effective carbon number descriptor corresponding to each monomer raw material are arranged and combined to obtain a multidimensional feature vector corresponding to each monomer raw material.

10. The method according to claim 3, characterized in that, The step of dividing the dataset according to a preset ratio of the target feature matrix and the relative correction factor corresponding to each monomer raw material to obtain a training set and a test set includes: Align and combine the feature vectors of each row in the target feature matrix with the relative correction factor corresponding to each monomer raw material to obtain a set of sample pairs; The sample set is divided according to the preset ratio to obtain the training set and the test set.

11. A device for predicting the relative correction factor of monomer raw materials in electronic atomizing liquid, characterized in that, The device includes: The acquisition unit is used to acquire the target molecular structure formula of the monomer raw material to be detected and the quantitative detection parameters of the monomer raw material to be detected. A computing unit is used to perform molecular descriptor calculation on the target molecular structure to obtain the target molecular descriptor; The prediction unit is used to input the target molecule descriptor into the target integrated prediction model for prediction to obtain the target relative correction factor. The target relative correction factor is used to perform quantitative calculation with the quantitative detection parameters of the monomer raw material to be detected to obtain the target concentration of the monomer raw material to be detected.

12. A quality control and analysis system for electronic atomizing fluid, characterized in that, The system includes an electronic device comprising a memory and a processor, the memory being connected to the processor, the processor being configured to execute one or more computer programs stored in the memory, wherein, when executing the one or more computer programs, the processor enables the electronic atomizing liquid quality control and analysis system to implement the method for predicting the relative correction factor of monomer raw materials in electronic atomizing liquids as described in any one of claims 1-10.

13. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which includes program instructions that, when executed by a processor, cause the processor to perform the method for predicting the relative correction factor of monomer raw materials in electronic atomizing liquid as described in any one of claims 1-10.

Images