A method, apparatus, and medium for molecular prediction of liquid impregnant air breathing performance

Abstract

This application discloses a molecular prediction method, device, and medium for the upgassing performance of liquid impregnating agents, applied in the field of liquid impregnating agent performance evaluation technology. The method obtains molecular descriptors and signed upgassing response values ​​of molecular samples to construct a modeling dataset; it performs multiple stratified random partitions based on the sign of the response values, completing preprocessing, feature selection, and hyperparameter search only within the training set; it obtains out-of-bounds predicted values ​​through cross-validation using at least two different types of machine learning models as base learners, and then fuses these predictions using stacked meta-learners to obtain initial predicted values; finally, it combines these with confidence factor correction for the applicable domain to obtain target predicted values, forming candidate models; it determines and stores representative models through external validation on a validation set and multi-index evaluation; after performing the same preprocessing on candidate molecules, it calls the model to output upgassing prediction values ​​and priority ranking results. This application can complete quantitative prediction of upgassing performance and determination of upgassing or gas evolution direction in batches before molecular synthesis, with a short screening cycle and low cost.

Description

Technical Field

[0001] This application relates to the field of liquid impregnating agent performance evaluation technology, and in particular to a molecular prediction method, equipment and medium for the gas absorption performance of liquid impregnating agents. Background Technology

[0002] Liquid impregnating agents for high-voltage power capacitors are a key component of the capacitor insulation system, and their getter properties directly affect the partial discharge characteristics and operational reliability of the capacitor. Traditional aromatic hydrocarbon impregnating agents, especially aromatic systems such as monobenzyltoluene and dibenzyltoluene, have long been used as liquid dielectrics in high-voltage power capacitors due to their high aromaticity, strong polarization ability, and good getter behavior. With the development of green and low-carbon power equipment, ester-based impregnating agents, which are environmentally friendly and have excellent dielectric properties, have begun to attract attention. Studies have shown that the getter properties of ester liquids are closely related to their molecular structure, with different molecular structures exhibiting significant differences in getter properties.

[0003] However, traditional experimental testing methods require sample preparation, capacitor assembly, and standardized getter experiments for each candidate substance, which is time-consuming, costly, and difficult to obtain getter performance data for thousands of ester liquids in a short period of time. This cannot meet the needs of rapid evaluation and screening of a large number of candidate molecules in the molecular design stage.

[0004] In view of the above-mentioned technologies, finding a method to rapidly evaluate the getter performance of ester impregnating agents based on molecular structure at the molecular design stage is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a molecular prediction method, device, and medium for the getter properties of liquid impregnating agents. This addresses the problems of existing traditional experimental testing methods being time-consuming, costly, and unable to obtain getter property data for thousands of ester liquids in a short time, thus failing to meet the need for rapid evaluation and screening of a large number of candidate molecules in the molecular design stage.

[0006] To address the aforementioned technical problems, this application provides a molecular prediction method for the getter properties of liquid impregnating agents, comprising: Obtain the molecular descriptors of the impregnating agent molecular samples and measure the signed inspiratory response values ​​of the impregnating agent molecular samples, and construct the corresponding modeling dataset; The modeling dataset is randomly divided into multiple stratified partitions according to the sign of the signed inspiratory response values. Data preprocessing, feature selection, and hyperparameter search are performed in the training set obtained from each partition to obtain the preprocessing rules, molecular descriptor set, and at least one set of model parameter combinations for that partition, which are then applied to the corresponding validation set. Using at least two different types of machine learning models as base learners, cross-validation is performed based on the training set partitioned each time and each set of model parameter combinations to obtain the out-of-fold prediction value, and the initial prediction value of the training set is obtained by stacking and fusing meta learners. Based on the initial predicted value and the corresponding signed inspiratory response value, the target predicted value is obtained by correcting the initial predicted value with the confidence factor of the applicable domain of the molecular structure, and the candidate models corresponding to each division and each combination of model parameters are obtained. External validation of each candidate model is performed using the corresponding validation set, evaluation metrics are obtained, and representative models are determined and stored according to preset rules. Obtain the molecular descriptors of the predicted candidate molecules, perform the same preprocessing as the molecular descriptor set according to the preprocessing rules corresponding to the representative model, call the representative model to determine the corresponding asspirability prediction value, and output the priority ranking result according to each asspirability prediction value.

[0007] Preferably, the molecular descriptor of the impregnating agent molecular sample is obtained, and the signed inspiratory response value of the impregnating agent molecular sample is measured, and the corresponding modeling dataset is constructed, including: Obtain the molecular structural formula or simplified linear input specification of the impregnating agent molecular sample, use descriptor calculation software to determine the corresponding molecular descriptor, and form a molecular descriptor matrix based on the molecular descriptor; Measure and record the signed inhalation response values ​​of the impregnating agent molecular samples, where a signed inhalation response value greater than zero indicates inhalation, and a signed inhalation response value less than zero indicates gas evolution. The molecular descriptor matrix is ​​cleaned to remove descriptors whose quality does not meet the preset conditions in order to obtain a modelable descriptor matrix. The modelable descriptor matrix is ​​associated with the signed inspiratory response values ​​to construct the modeling dataset.

[0008] Preferably, the molecular descriptor matrix is ​​cleaned to remove descriptors whose quality does not meet preset conditions, in order to obtain a modelable descriptor matrix, including: Descriptors with a missing rate higher than a first preset threshold are removed; Descriptors with variance below a second preset threshold are removed; For any two descriptors with a correlation coefficient higher than the third preset threshold, the one with a lower correlation to the signed inspiratory response value is removed.

[0009] Preferably, data preprocessing, feature selection, and hyperparameter search are performed within each training set obtained from the partitioning, including: The modelable descriptor matrix in the training set is imputed by missing median values ​​and standardized to obtain the preprocessing rules for this partitioning. The recursive feature elimination method is used to filter the features of the standardized modelable descriptor matrix to obtain the molecular descriptor set for this partition. Hyperparameter search is performed based on the set of molecular descriptors to obtain the model parameter combination for this partition.

[0010] Preferably, at least two different types of machine learning models are used as base learners, and cross-validation is performed based on the training set of each partition and each set of model parameter combinations to obtain the out-of-bounds prediction value, including: The training set obtained from each partition is divided into multiple subsets; One subset is used as the validation subset and the remaining subsets are used as the training subsets to train each base learner and obtain the prediction value corresponding to the validation subset. Iterate through each subset to obtain the out-of-place prediction values ​​of each base learner for all samples in the training set, where the out-of-place prediction value of each training sample comes from the base learner that does not include that sample.

[0011] Preferably, the initial predicted values ​​of the training set for this partition are obtained by stacking and fusing meta-learners, including: The out-of-place predictions of each base learner are combined to form a two-level input matrix; The meta-learner is trained using the second-level input matrix as input and the corresponding signed inspiratory response value as output to obtain the initial prediction value.

[0012] Preferably, the target predicted value is obtained by correcting the initial predicted value based on the initial predicted value and the corresponding signed inspiratory response value, combined with the confidence factor of the applicable domain of the molecular structure, including: The residuals are calculated based on the initial predicted values ​​of the training set and the corresponding signed inspiratory response values. A residual correction model is established with the set of molecular descriptors as input and the residuals as output. Confidence factor for determining the applicable domain of molecular structure based on the structural space of the training set; The initial predicted value is corrected based on the applicable domain confidence factor and the residual correction model to obtain the target predicted value; The expression for the target predicted value is: ; in, Let i be the target predicted value for the i-th sample; This is the initial predicted value for the i-th sample; This is the residual shrinkage coefficient; Let be the confidence factor for the applicable domain of the i-th sample; Let be the residual correction model for the i-th sample.

[0013] Preferably, the process of acquiring evaluation indicators and determining and storing representative models according to preset rules includes: The evaluation metrics for each candidate model are obtained, including: coefficient of determination, root mean square error, mean absolute error, cross-validation prediction coefficient, external prediction coefficient, consistency correlation coefficient, and direction discrimination accuracy. After normalizing each evaluation indicator according to the preset weights, the weighted sum is obtained to get the corresponding evaluation result. Based on the evaluation results, candidate models that meet the comprehensive performance requirements are selected as representative models.

[0014] On the other hand, this application also provides an electronic device, including a memory for storing computer programs; A processor is used to execute a computer program to implement the steps of the molecular prediction method for the gas absorption properties of the liquid impregnating agent described above.

[0015] On the other hand, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described molecular prediction method for the gas absorption performance of liquid impregnating agents.

[0016] This application provides a molecular prediction method for the uptake performance of liquid impregnating agents. Based on the molecular microstructure characteristics represented by the impregnating agent molecule descriptor, the method performs multiple hierarchical random partitions of the modeling dataset according to the sign of the signed response values. This ensures that the proportion of impregnating agent molecule samples corresponding to the signed response values ​​in the training and validation sets obtained from each partition remains consistent with the overall population and is isolated from each other. Data preprocessing, feature selection, and hyperparameter search are completed in a closed loop only within the training set, and the resulting rules are then extrapolated and applied to the validation set. At least two different types of machine learning models are used as base learners. Leveraging the characteristics of different machine learning models, cross-validation is used to obtain extrapolated prediction values, which are then fused by stacking meta-learners to obtain the corresponding initial prediction values. The initial prediction values ​​are then corrected using the confidence factor of the applicable domain determined by the structure space of the training set. Each set of model parameter combinations under each partition is independently trained to form candidate models. These models are externally validated on the validation set and selected through multi-index screening, and representative models are determined and stored. The same preprocessing is applied to the candidate molecules, and their uptake prediction values ​​are obtained by calling the model. Compared to traditional physical testing methods, this application eliminates the need for impregnating agent synthesis and sample preparation. It enables the early, batch prediction of the uptake performance of various impregnating agent molecules and ranks different molecules based on their predicted uptake performance. This application offers rapid screening, requires minimal experimental consumables, and can process a large number of candidate molecules at once. Even novel molecules that have not yet been synthesized can have their performance predicted in advance. This overcomes the problems of traditional methods, which require sample preparation, capacitor assembly, and electric field gas evolution experiments, resulting in long processing times, high costs, and difficulty in completing large-scale screening of impregnating agent molecules in a short time. Attached Figure Description

[0017] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a molecular prediction method for the gas absorption performance of a liquid impregnating agent, provided for embodiments of this application; Figure 2 A schematic diagram of the fitting curve of the applicable domain-aware residual correction stacking model provided in the embodiments of this application; Figure 3 A flowchart illustrating a molecular prediction method for the gas absorption performance of a liquid impregnating agent, provided in an embodiment of this application. Figure 4 A block diagram of a molecular prediction device for the gas absorption performance of a liquid impregnating agent, provided for another embodiment of this application; Figure 5 A structural diagram of an electronic device provided in another embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0020] The core of this application is to provide a molecular prediction method, device, and medium for the gas absorption performance of liquid impregnating agents.

[0021] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Traditional aromatic hydrocarbon impregnating agents, especially aromatic systems such as monobenzyltoluene and dibenzyltoluene, have long been used as liquid dielectrics in high-voltage power capacitors due to their high aromaticity, strong polarization ability, and good getter behavior. With the development of green and low-carbon power equipment, ester impregnating agents, which are environmentally friendly and have excellent dielectric properties, have begun to attract attention. The getter performance of ester liquids is closely related to their molecular structure, and different molecular structures exhibit significant differences in getter performance. However, traditional experimental testing methods are insufficient to obtain the getter performance of thousands of ester liquids in a short time. Therefore, it is of great significance to develop a method for rapidly and accurately determining the getter performance of ester impregnating agent molecules and screening for molecules with excellent getter properties.

[0023] Existing experimental studies on the gas absorption and gas evolution performance of electrical insulating liquids indicate that the overall insulation performance of aromatic hydrocarbon impregnating media is highly correlated with the molecular aromatic structure. Through multi-index performance evaluation of monobenzyltoluene and dibenzyltoluene aromatic impregnating agents, comprehensively examining aromaticity, aromatic ring carbon content, kinematic viscosity, freezing point, dielectric constant, partial discharge characteristics, and gas evolution indicators, a pattern emerges: higher molecular aromaticity and lower system viscosity can simultaneously improve the dielectric's gas absorption capacity and low-temperature partial discharge tolerance. Comparative tests on the gas evolution behavior of insulating liquids with different hydrocarbon components under AC, DC, and AC / DC superimposed electric field conditions confirmed that the molecular aromatic hydrocarbon content is the core factor determining the gas absorption or gas evolution tendency of the insulating oil, and the electric field conditions significantly alter the gas evolution test results. Further research on the correlation between the gas evolution of insulating oil and aromatic components revealed significant differences in the degree to which monocyclic aromatic hydrocarbons, bicyclic aromatic hydrocarbons, and tricyclic and higher polycyclic aromatic hydrocarbons inhibit the gas evolution behavior of the dielectric. Studies on the modification of environmentally friendly natural ester insulating media show that compounding aromatic components into refined plant-based esters can effectively improve the absorption capacity of ester media for hydrogen and partial-emission pyrolysis gases of light hydrocarbons, thereby enhancing the gas absorption performance of ester media. Based on the above research, it is clear that existing studies rely on molded liquid physical samples to complete electric field gas evolution tests, allowing only performance verification on a small number of prepared samples. This approach cannot achieve rapid batch evaluation and preliminary screening of a large number of candidate ester impregnating agent molecules at the molecular design stage, resulting in technical limitations such as long experimental cycles, high R&D costs, and difficulty in large-scale molecular screening.

[0024] In other words, existing research mainly focuses on standardized experimental testing of prepared samples, comprehensive performance evaluation, macroscopic composition correlation, and additive or formulation optimization. While capable of determining the gas absorption or evolution behavior of specific samples, it still has significant shortcomings: First, the research subjects are mostly aromatic hydrocarbon impregnating agents and mineral oils, making it difficult to reveal the influence of ester single-molecule structural differences on gas absorption performance; second, existing evaluation methods typically rely on sample preparation, capacitors, or experimental testing, making it difficult to conduct large-scale rapid screening before candidate molecule synthesis or formulation; third, empirical rules based on aromatic hydrocarbon content, aromatic carbon content, or formulation ratios are difficult to directly apply to candidate molecules with ester structures. Therefore, current technology cannot simultaneously achieve numerical prediction of ester impregnating agent absorbency, determination of gas absorption or evolution direction, and priority screening of a large number of candidate molecules.

[0025] Based on the above problems, the molecular prediction method for the getter properties of liquid impregnating agents provided in this application is mainly used for the molecular prediction of the getter properties of ester-based liquid impregnating agents. The method is as follows... Figure 1 As shown, it includes the following steps: S10: Obtain the molecular descriptor of the impregnating agent molecular sample and measure the signed inhalation response value of the impregnating agent molecular sample, and construct the corresponding modeling dataset.

[0026] In a specific embodiment, firstly, samples of impregnating agent molecules (e.g., ester and ester-related impregnating agent molecules) are collected. Then, the molecular structural formula or SMILES (Simplified Molecular InputLine Entry System) expression of the impregnating agent molecules is obtained. Finally, the signed gas-escape response value Gas of the impregnating agent molecules is measured through a standard electric field gas evolution test. The Gas value is stored in a signed manner, where Gas>0 indicates gas absorption and Gas<0 indicates gas evolution.

[0027] After obtaining the molecular structural formula or SMILES expression, all structural feature descriptors of the molecule are extracted using relevant computational software (e.g., PaDEL-Descriptor). After data preprocessing, the processed effective molecular features, gas values, and basic molecular information (e.g., sample number) are integrated to form a modeling dataset.

[0028] S11: The modeling dataset is randomly divided into multiple stratified partitions according to the sign of the signed inspiratory response values. In the training set obtained from each partition, data preprocessing, feature selection, and hyperparameter search are performed to obtain the preprocessing rules, molecular descriptor set, and at least one set of model parameter combinations for that partition, and then applied to the corresponding validation set.

[0029] In a specific implementation, the modeling dataset is randomly divided into multiple stratified partitions based on the sign of the signed inspiratory response value (Gas value) during modeling. This ensures that both the training and validation datasets contain both inspiratory and exhaled gas samples, preventing a particular type of sample from being concentrated in a single dataset. This guarantees that the baseline conditions for model learning and performance verification are consistent.

[0030] Furthermore, all operations such as filling missing data (data preprocessing), standardizing features, selecting effective molecular features (feature selection), adjusting model parameters, and hyperparameter search are performed only on the training set. The obtained processing standards, selected features, and optimal parameters are then directly applied to the validation set, preventing the data features of the validation set from being prematurely leaked to the model and avoiding inflated or distorted model prediction results.

[0031] It's easy to understand that during this data preprocessing, feature selection, and hyperparameter search process, a recursive feature elimination method is used to remove molecular descriptors with minimal impact on inhalation performance. This simplifies the large number of original features, reducing the computational burden caused by redundant data and making it easier for the model to capture the correlation between molecular structure and inhalation performance. Furthermore, after multiple rounds of comparing the model performance across various feature dimensions, a set of 20 optimal molecular descriptors (the molecular descriptor set) is fixed. Subsequent validation and new molecule predictions only extract these features, strictly adhering to the training order. This consistent data processing standard further reduces data processing bias, ultimately making the model's prediction results more stable and reliable.

[0032] S12: Using at least two different types of machine learning models as base learners, cross-validate is performed based on the training set of each partition and each set of model parameter combinations to obtain the out-of-fold prediction value, and the initial prediction value of the training set is obtained by stacking and fusing the meta-learners.

[0033] S13: Based on the initial predicted value and the corresponding signed inspiratory response value, the target predicted value is obtained by correcting the initial predicted value with the confidence factor of the applicable domain of the molecular structure, and the candidate models corresponding to each division and each combination of model parameters are obtained.

[0034] In a specific embodiment, multiple different types of machine learning models are used as base models. Cross-validation is carried out for each dataset partition and each set of model parameters to obtain out-of-range prediction results that do not participate in their own training samples. Then, the output results of multiple base models are integrated through a meta-learner, which can give full play to the advantages of different models in capturing the nonlinear laws of molecular structure and inhalation performance, and make up for the defect of large fitting bias of a single model. Then, the correction logic is built by combining the difference between the actual inhalation value of the sample (signed inhalation response value) and the initial prediction result (initial prediction value). At the same time, the structural similarity between the molecule to be tested and the training sample is judged by the confidence factor of the molecular structure applicability domain, and the correction amplitude is adaptively adjusted. After correcting the initial prediction value, the final reliable prediction result (target prediction value) is obtained.

[0035] It is easy to understand that each set of divisions, combined with a set of parameters, can generate a complete and independent candidate model. This not only improves the accuracy of numerical prediction of molecular inhalation performance, but also significantly reduces the problem of prediction distortion for new molecules with structures that are significantly different from the training molecules. It also provides sufficient alternatives for subsequent selection of multiple models.

[0036] S14: Use the corresponding validation set to perform external validation on each candidate model, obtain evaluation indicators, and determine and store representative models according to preset rules.

[0037] S15: Obtain the molecular descriptors of the predicted candidate molecules, perform the same preprocessing as the molecular descriptor set according to the preprocessing rules corresponding to the representative model, call the representative model to determine the corresponding asspirability prediction value, and output the priority ranking result according to each asspirability prediction value.

[0038] In a specific embodiment, each group uses a corresponding validation set to conduct independent external tests on all candidate models, calculate multiple model performance evaluation indicators, and select the model with the best overall performance as a representative model based on preset evaluation criteria. The data processing rules, molecular descriptor set, and model parameter combination are then fully saved. When evaluating new molecules to be screened (predicted candidate molecules) in the future, only the molecular descriptor set needs to be extracted. The data processing flow and fixed feature set of the finalized model are strictly followed to complete unified preprocessing. The saved representative model is directly called to calculate the molecular inhalation performance value (inhalation prediction value). Then, inhalation and gas evolution molecules are distinguished based on the positive or negative inhalation prediction value and sorted according to the strength of inhalation ability.

[0039] This application provides a molecular prediction method for the uptake performance of liquid impregnating agents. Based on the molecular microstructure characteristics represented by the impregnating agent molecule descriptor, the method performs multiple hierarchical random partitions of the modeling dataset according to the sign of the signed response values. This ensures that the proportion of impregnating agent molecule samples corresponding to the signed response values ​​in the training and validation sets obtained from each partition remains consistent with the overall population and is isolated from each other. Data preprocessing, feature selection, and hyperparameter search are completed in a closed loop only within the training set, and the resulting rules are then extrapolated and applied to the validation set. At least two different types of machine learning models are used as base learners. Leveraging the characteristics of different machine learning models, cross-validation is used to obtain extrapolated prediction values, which are then fused by stacking meta-learners to obtain the corresponding initial prediction values. The initial prediction values ​​are then corrected using the confidence factor of the applicable domain determined by the structure space of the training set. Each set of model parameter combinations under each partition is independently trained to form candidate models. These models are externally validated on the validation set and selected through multi-index screening, and representative models are determined and stored. The same preprocessing is applied to the candidate molecules, and their uptake prediction values ​​are obtained by calling the model. Compared to traditional physical testing methods, this application eliminates the need for impregnating agent synthesis and sample preparation. It enables the early, batch prediction of the uptake performance of various impregnating agent molecules and ranks different molecules based on their predicted uptake performance. This application offers rapid screening, requires minimal experimental consumables, and can process a large number of candidate molecules at once. Even novel molecules that have not yet been synthesized can have their performance predicted in advance. This overcomes the problems of traditional methods, which require sample preparation, capacitor assembly, and electric field gas evolution experiments, resulting in long processing times, high costs, and difficulty in completing large-scale screening of impregnating agent molecules in a short time.

[0040] Based on the above embodiments, as a preferred embodiment, the specific implementation method for obtaining the molecular descriptor of the impregnating agent molecular sample and measuring the signed updraft response value of the impregnating agent molecular sample, and constructing the corresponding modeling dataset is as follows: Obtain the molecular structural formula or simplified linear input specification of the impregnating agent molecular sample, use descriptor calculation software to determine the corresponding molecular descriptor, and form a molecular descriptor matrix based on the molecular descriptor; Measure and record the signed inhalation response values ​​of the impregnating agent molecular samples, where a signed inhalation response value greater than zero indicates inhalation, and a signed inhalation response value less than zero indicates gas evolution. The molecular descriptor matrix is ​​cleaned to remove descriptors whose quality does not meet the preset conditions in order to obtain a modelable descriptor matrix. The modelable descriptor matrix is ​​associated with the signed inspiratory response values ​​to construct the modeling dataset.

[0041] The specific implementation of data cleaning of the molecular descriptor matrix, removing descriptors whose quality does not meet preset conditions to obtain a modelable descriptor matrix, is as follows: Descriptors with a missing rate higher than a first preset threshold are removed; Descriptors with variance below a second preset threshold are removed; For any two descriptors with a correlation coefficient higher than the third preset threshold, the one with a lower correlation to the signed inspiratory response value is removed.

[0042] In a specific embodiment, firstly, samples of impregnating agent molecules (e.g., ester and ester-related impregnating agent molecules) are collected. Then, the molecular structural formula or SMILES (Simplified Molecular InputLine Entry System) expression of the impregnating agent molecules is obtained. Finally, the signed gas-escape response value Gas of the impregnating agent molecules is measured through a standard electric field gas evolution test. The Gas value is stored in a signed manner, where Gas>0 indicates gas absorption and Gas<0 indicates gas evolution.

[0043] Then, relevant computational software (e.g., PaDEL-Descriptor) is used to calculate molecular descriptors and construct the corresponding molecular descriptor matrix. In a preferred embodiment, the initial sample dataset contains 65 ester-based liquid dielectric samples, and the initial numerical molecular descriptors are 1444.

[0044] The molecular descriptor matrix undergoes data cleaning, including deleting descriptors with high missing rates (those with missing rates exceeding a first preset threshold), deleting descriptors with constant or near-zero variance (those with variance below a second preset threshold), and deleting highly correlated descriptors (for any two descriptors with a correlation coefficient exceeding a third preset threshold, the one with lower correlation to the signed inspiratory response value is removed). After preprocessing, a modelable descriptor matrix is ​​obtained, which, along with sample numbers, gas values, and inspiratory or evolution labels, constitutes the modeling dataset.

[0045] The hierarchical sample division provided in this application can balance the distribution of the two types of performance samples, avoiding model learning bias towards a single type; multi-layer cleaning removes invalid and redundant features, which can reduce the amount of computation and enable the model to accurately grasp the intrinsic relationship between molecular structure and inhalation performance.

[0046] Based on the above embodiments, as a preferred embodiment, the specific implementation method for performing data preprocessing, feature selection, and hyperparameter search within each partitioned training set is as follows: The modelable descriptor matrix in the training set is imputed by missing median values ​​and standardized to obtain the preprocessing rules for this partitioning. The recursive feature elimination method is used to filter the features of the standardized modelable descriptor matrix to obtain the molecular descriptor set for this partition. Hyperparameter search is performed based on the set of molecular descriptors to obtain the model parameter combination for this partition.

[0047] In a specific embodiment, considering that the Gas value includes both gas intake and gas evolution samples, in order to ensure that a relatively stable directional distribution is maintained in both the training and validation sets, this application performs multiple stratified random partitioning based on the positive or negative sign of the Gas value. In a preferred embodiment, 100 stratified random partitionings are used, with 20% of the samples allocated to the validation set each time, and the remaining samples used as the training set.

[0048] In each partition, missing value median imputation, descriptor standardization, recursive feature elimination, and hyperparameter search are all performed solely on the training set. The median, standardized parameters, and feature subset rules learned from the training set are solidified and transferred to the validation set and subsequent candidate molecules to be predicted, thereby preventing validation information from being prematurely incorporated into the training process.

[0049] Feature selection employs a recursive feature elimination (RFE) method within the training set. RFE uses a 500-tree random forest as the feature ranking estimator, recursively eliminating 10% of low-contribution descriptors in each round, resulting in candidate feature subsets of 20, 15, 10, 8, and 5 dimensions. As a preferred approach, based on recursive feature elimination within the training set and model performance evaluation under different feature dimensions, the optimal method is to use the 20-dimensional molecular descriptor subset (molecular descriptor set) corresponding to the 38th partition as the fixed input variable for the applicable domain-aware residual correction stacked model. The 20-dimensional descriptors should be based on the RFE feature file output from this training process, and fixed according to the original column names and column order during training; they should not be re-selected or adjusted during the validation set or candidate molecule prediction stage.

[0050] The 20-dimensional descriptors are AATS4m, AATSC0v, AATSC3v, ATSC4p, BCUTP-1h, BCUTw-1h, ETA_dEpsilon_B, GATS1p, GATS5m, GATS5v, JGI10, JGI8, MATS1c, MLFER_E, R_TpiPCTPC, SHother, SIC5, maxHother, maxwHBa, and piPC10. For any candidate molecule to be predicted, the original molecular descriptor is first calculated using the same version of PaDEL-Descriptor software as used in the training phase. Then, only the fixed 20-dimensional descriptors are extracted, and the input vector is formed strictly according to the column order of the training phase.

[0051] The embodiment provided in this application can balance the distribution of two types of samples in the dataset, avoiding model training bias towards a single performance type; all data processing, feature selection, and parameter optimization processes rely solely on the training set generation standard and are uniformly reused, which can prevent data leakage caused by early participation of validation set information in training, and ensure the authenticity and reliability of the model generalization evaluation results; at the same time, recursive feature elimination simplifies the original features of thousands of dimensions into a relatively compact set of candidate structural representations that contribute to the inspiratory response (which can also be understood as high contribution features), which greatly reduces the computational load of the model and reduces the interference caused by irrelevant features.

[0052] Based on the above embodiments, as a preferred embodiment, using at least two different types of machine learning models as base learners, and performing cross-validation based on each partitioned training set and each set of model parameter combinations to obtain the out-of-bounds prediction value, the specific implementation method is as follows: The training set obtained from each partition is divided into multiple subsets; One subset is used as the validation subset and the remaining subsets are used as the training subsets to train each base learner and obtain the prediction value corresponding to the validation subset. Iterate through each subset to obtain the out-of-place prediction values ​​of each base learner for all samples in the training set, where the out-of-place prediction value of each training sample comes from the base learner that does not include that sample.

[0053] Correspondingly, The specific implementation method for obtaining the initial predicted values ​​of the training set for this partition through stacking and fusing meta-learners is as follows: The out-of-place predictions of each base learner are combined to form a two-level input matrix; The meta-learner is trained using the second-level input matrix as input and the corresponding signed inspiratory response value as output to obtain the initial prediction value.

[0054] In specific embodiments, this application uses three different types of machine learning models—Random Forest (RF), Extreme Gradient Boosting (XGB), and Neural Network (NN)—as base learners (which can also be understood as first-level base learners). Random Forest is used to capture nonlinear tree ensemble relationships between multiple descriptors; Extreme Gradient Boosting is used to characterize gradient-boosted nonlinear structure-property mappings; and Neural Networks are used to represent shallow nonlinear mappings. These three types of models have different biases, variance characteristics, and structural response modes for small sample molecular descriptor data; therefore, as heterogeneous base learners, they can improve the complementarity of stacking and fusion.

[0055] Specifically, the training set is divided into K non-overlapping subsets (here, K=3, using tri-fold cross-validation). For any one of the first-level base learners (RF, XGB, and NN), one subset is selected each time as the out-of-fold prediction subset (validation subset), and the first-level base learner is trained using the remaining K-1 subsets (training subsets), predicting the samples in the current out-of-fold prediction subset. After traversing all subsets, each training sample obtains an out-of-fold prediction value generated by the first-level model that was not trained using that sample. Thus, each training sample obtains RF out-of-fold prediction values, XGB out-of-fold prediction values, and NN out-of-fold prediction values, respectively. These three prediction values ​​are arranged in sample order to form a three-dimensional second-level input matrix, and the input matrices of all training samples together constitute the out-of-fold prediction matrix. Using this out-of-fold prediction matrix as input and the signed inspiratory response values ​​(measured inspiratory values) of the training samples as output, a Ridge (Ring Regression) meta-learner is trained, thereby obtaining the fusion relationship between the first-level base learners and the initial prediction results (initial prediction values) after stacking.

[0056] For example: RF can use 300 trees, with no limit on maximum depth and a minimum number of leaf node samples of 1; XGB can use the squared-error loss function, with a maximum depth of 3, a learning rate of 0.05, and both row sampling rate and column sampling rate of 0.8; NN can use a single hidden layer structure, 32 neurons, ReLU activation function, Adam optimizer, learning rate of 0.001, and a maximum number of iterations of 5000.

[0057] This application employs multiple machine learning models as base models, splitting the training set into multiple subsets for cross-validation. In each round, one subset is used for validation, while the remaining subsets are used to train the model. After all iterations are completed, the prediction result for each sample comes from the training process that did not include that sample, effectively preventing the base model from memorizing sample data in advance and reducing the risk of overfitting. The prediction results calculated by all base models are integrated as new input data, and a meta-learner is then trained to perform fusion prediction. This integrates the fitting advantages of different models, compensates for the limitations of a single model in capturing molecular and inspiratory performance patterns, and ensures that the final output initial prediction value closely matches the actual inspiratory response, thereby improving the overall prediction accuracy.

[0058] Based on the above embodiments, as a preferred embodiment, the specific implementation method for obtaining the target predicted value by correcting the initial predicted value according to the initial predicted value and the corresponding signed inspiratory response value, combined with the confidence factor of the applicable domain of the molecular structure, is as follows: The residuals are calculated based on the initial predicted values ​​of the training set and the corresponding signed inspiratory response values. A residual correction model is established with the set of molecular descriptors as input and the residuals as output. Confidence factor for determining the applicable domain of molecular structure based on the structural space of the training set; The initial predicted value is corrected based on the applicable domain confidence factor and the residual correction model to obtain the target predicted value; The expression for the target predicted value is: ; in, Let i be the target predicted value for the i-th sample; This is the initial predicted value for the i-th sample; This is the residual shrinkage coefficient; Let be the confidence factor for the applicable domain of the i-th sample; Let be the residual correction model for the i-th sample.

[0059] In a specific embodiment, after obtaining the initial predicted value after stacking, this application further introduces an applicable domain-aware residual correction module to improve the local fitting ability of the ordinary stacked model in the reliable structural space and control the extrapolation risk when far from the training structural space.

[0060] First, the residuals are calculated based on the initial predicted values ​​from the cross-fit stack of the training samples. For the i-th training sample, its experimental values ​​are... (Essentially Gas) and the stacked initial predictions obtained when the corresponding model was fitted without using this sample. The difference is defined as the residual. : Subsequently, the molecular descriptor vectors (set of molecular descriptors) from the training samples were solidified and preprocessed. As feature input, with residual Establish a residual correction model for the response variable. It is used to predict systematic biases that are not yet explained in the initial predictions after stacking.

[0061] Secondly, the confidence factor of the applicable domain of the query molecule is calculated based on the training set structure space. Its applicable domain confidence factor can be determined by at least one of the leverage value, standardized Euclidean distance, and Mahalanobis distance, or by the product, minimum, or weighted average of the above confidence factors. Its function is to evaluate the degree of consistency between the query sample and the structure space of the training set.

[0062] Finally, the prediction results are output according to the following attenuation double correction formula: ; in, Let i be the target predicted value for the i-th sample; is the initial predicted value for the i-th sample, representing the molecular descriptor of the queried sample; The residual shrinkage coefficient is preferably in the range of 0.30 to 0.50. is the confidence factor for the applicable domain of the i-th sample, and its value ranges from 0 to 1; This is the residual correction model for the i-th sample. This formula enables the model in this application to fully utilize residual information for local correction when the query molecule is within the training domain, and to automatically reduce the correction magnitude when the query molecule deviates from the training domain.

[0063] Based on the above embodiments, as a preferred embodiment, the specific implementation method for obtaining evaluation indicators and determining and storing representative models according to preset rules is as follows: The evaluation metrics for each candidate model are obtained, including: coefficient of determination, root mean square error, mean absolute error, cross-validation prediction coefficient, external prediction coefficient, consistency correlation coefficient, and direction discrimination accuracy. After normalizing each evaluation indicator according to the preset weights, the weighted sum is obtained to get the corresponding evaluation result. Based on the evaluation results, candidate models that meet the comprehensive performance requirements are selected as representative models.

[0064] In a specific embodiment, after constructing the stacked model and the domain-aware residual correction module, the performance of candidate models under different random seeds, different numbers of features, and different parameter combinations is comprehensively evaluated. Evaluation metrics include the coefficient of determination. Root mean square error Mean absolute error Cross-validation prediction coefficients External prediction coefficient Coherence correlation coefficient And the accuracy of determining the direction of inhalation or exhalation.

[0065] The expressions for each evaluation indicator are as follows: ; ; ; ; ; ; in, It is the number of samples; It is the experimental value of the i-th sample; It is the predicted inspiratory performance value obtained for the i-th sample (that is, the target predicted value of the model output for that sample). ); It is the average of all sample experimental values; It is the size of the training set; It is the experimental value of the inhalation of the i-th training sample (which is essentially Gas). This refers to the prediction value obtained by rebuilding the complete model using the remaining training samples after removing the i-th sample from the training set (that is, the final model prediction value obtained when performing leave-one-out cross-validation on the selected model). This represents the average of the experimental values ​​in the training set. It is the number of validation sets; It is the inspiratory test value of the i-th verification sample; It is the final predicted value obtained from the i-th validation sample; This represents the covariance between the experimental and predicted values. and It is the variance of the corresponding quantity; It is the average of the predicted values.

[0066] Based on the above evaluation indicators, the candidate models are comprehensively compared. That is, the evaluation indicators are normalized according to the preset weights and then weighted and summed to obtain the corresponding evaluation results (which can also be understood as the final evaluation score). Based on the evaluation results, the candidate model that meets the comprehensive performance requirements (the candidate model with the highest evaluation score) is selected as the representative model, and its molecular descriptor set, data preprocessing rules and model parameters are solidified for subsequent prediction and screening of the gas uptake of candidate impregnating agent molecules.

[0067] The embodiments provided in this application employ multiple indicators to evaluate the performance of candidate models from different dimensions. These include error indicators to assess the magnitude of numerical prediction deviations, various fitting coefficients and consistency coefficients to reflect the overall fitting effect, and a direction discrimination accuracy rate to specifically verify whether the model can distinguish between inhalation and exhalation properties, avoiding the problem of focusing only on predicted values ​​while ignoring the accuracy of gas direction judgment. After all indicators are normalized, they are comprehensively scored according to set weights, preventing evaluation imbalances caused by differences in the numerical ranges of different indicators. The scoring results can objectively quantify the overall performance of each model. By relying on the comprehensive score to select the optimal model, it is possible to balance the accuracy of predicted numerical values, structural generalization ability, and inhalation and exhalation direction recognition ability, selecting the model with the strongest overall stability as the final model, ensuring that subsequent quantitative predictions of the inhalation performance of impregnating agent molecules and gas direction judgments are sufficiently reliable.

[0068] After selecting the aforementioned model, for predicted candidate molecules (candidate ester impregnating agents to be screened), their structural formulas or SMILES expressions are first obtained, and molecular descriptors are generated using the same descriptor calculation software as in the training phase. Subsequently, homology processing is performed according to the median imputation rules, normalization rules, and feature sets fixed in the training phase. The processed candidate molecules are input into the fixed domain-aware residual correction stacking model to obtain Gas prediction values ​​(getting prediction values). When the prediction value is greater than 0, the candidate molecule is determined to be an getting candidate; when the prediction value is less than 0, the candidate molecule is determined to be an evolution candidate. For multiple candidate molecules, they can be sorted from largest to smallest according to the prediction value, and a priority list of getting capabilities can be output to guide subsequent synthesis, formulation design, and standard experimental verification.

[0069] In addition, to verify the fitting ability and external prediction ability of the selected representative model, the training set and validation set results were evaluated. The results show that the domain-aware residual correction stacking model achieves good overall performance under the conditions of 38 random partitions and 20-dimensional feature subsets. Among the evaluation metrics corresponding to its training set, the coefficient of determination is [missing information]. Cross-validation prediction coefficients Root mean square error Coherence correlation coefficient Among the evaluation metrics corresponding to the validation set: coefficient of determination External prediction coefficient Root mean square error Coherence correlation coefficient In addition, the mean absolute error The direction discrimination accuracy (verification) is 1.000. This result demonstrates that the method provided in this application can achieve high fitting accuracy and external prediction capability on small sample ester impregnating agent data. The fitting curve of the applicable domain-aware residual correction stacking model is shown below. Figure 2 As shown.

[0070] Further, several external ester liquid samples were selected for prediction and verification. The results are shown in Table 1. The model of this application can correctly identify the gas absorption or gas evolution direction of all external samples. Among them, 3 samples were identified as gas absorption type and 1 sample was identified as gas evolution type. The direction discrimination is consistent with the experimental measurement results, which shows that the method of this application can effectively serve the evaluation of the gas absorption of candidate ester molecules.

[0071] Table 1

[0072] Based on the above, the flowchart corresponding to the molecular prediction method for the getter performance of liquid impregnating agents provided in this application is as follows: Figure 3 As shown, the process includes four steps. The first step is to establish a sample dataset of ester impregnating agents' gas-up response; the second step is to construct an applicable domain-aware residual correction stacked model; the third step is to evaluate model performance and determine representative models; and the fourth step is to conduct gas-up performance evaluation and screening of candidate impregnating agent molecules.

[0073] For the first step (establishing a dataset of ester impregnating agent upgas response samples), an upgas response sample database of ester impregnating agents is established. Ester and ester-related impregnating agent molecular samples are collected, their molecular structural formulas or SMILES expressions are obtained, signed upgas response values ​​are measured and recorded, molecular descriptors are calculated, and after data preprocessing, a modeling dataset containing sample numbers, gas response values, a molecular descriptor matrix, and upgas or gas evolution labels is formed.

[0074] For the second step (constructing the applicable domain-aware residual correction stacked model - AD_DC_STACKING model), the sample data is divided into hierarchical random partitions, and feature selection and preprocessing rules are solidified within the training set. RF (random forest), XGB (extreme gradient boosting), and NN (neural network) are used as first-level base learners for stacking and fusion. Based on this, an applicable domain constraint calculation and residual correction module are introduced to form an inspiratory response prediction model with intra-domain enhancement correction and out-of-domain suppression extrapolation.

[0075] For the third step (model performance evaluation and representative model determination), the prediction results under different feature combinations and model parameters are repeatedly and randomly divided to determine the meta-learners (RF, XGB, NN, Ridge Regression) and their corresponding calibration parameters, and the coefficient of determination is used. Root mean square error Mean absolute error Cross-validation prediction coefficients External prediction coefficient Coherence correlation coefficient The model performance is evaluated based on indicators such as the accuracy of identifying the direction of gas inhalation or gas exhalation, and finally, a representative model is determined.

[0076] For the fourth step (conducting evaluation and screening of the gas absorption performance of candidate impregnating agent molecules), descriptors are generated and homology preprocessing is performed on the candidate new molecules. The solidified applicable domain-aware residual correction stacking model is called to calculate the gas absorption prediction value. The direction of gas absorption or gas evolution is determined according to the positive or negative sign of the prediction value, and suitable substances are screened according to the predicted gas absorption capacity.

[0077] Therefore, the molecular prediction method for the getter properties of liquid impregnating agents provided in this application has the following advantages: 1. Data-dependent operations such as missing value handling, standardization, feature selection, model parameter determination, and prediction generation are limited to the training set. This reduces data leakage caused by validation set information entering the modeling process prematurely, and improves the standardization, repeatability, and reliability of model construction under conditions of small samples and high-dimensional descriptors.

[0078] 2. By leveraging the complementary representation capabilities of random forests, extreme gradient boosting, and neural networks for nonlinear structure-property relationships, and reducing the prediction bias of a single model through stacking fusion and residual secondary correction, we can achieve synchronous output of numerical prediction of gas intake and determination of the direction of gas intake or gas evolution.

[0079] 3. The residual correction amplitude is adaptively adjusted by using the confidence factor of the applicable domain. When the candidate molecule is close to the training structure space, the local correction is enhanced, and when the candidate molecule deviates from the training structure space, the residual correction is automatically reduced. This avoids the unreliable extrapolation error from being further amplified, and provides a more robust technical basis for the priority screening of ester impregnating agent candidate molecules and subsequent experimental verification.

[0080] The significance of the molecular prediction method for the getter properties of liquid impregnating agents provided in this application lies in offering a method for evaluating and screening the getter properties of ester impregnating agents based on a domain-aware residual correction stacking model. This allows candidate molecules to directly obtain predictive values ​​for getter properties, determine the direction of getter or gas evolution, and prioritize results from their molecular structures before experimental verification, thereby shortening the early material screening cycle and reducing the cost of sample preparation and getter property testing. This method transforms traditional empirical judgment into data-driven structure-property prediction through molecular descriptors, feature screening within the training set, heterogeneous stacking fusion, and residual correction. Simultaneously, this application introduces domain constraints into the residual correction process, enhancing local correction when candidate molecules are close to the training structure space and suppressing unreliable extrapolation when they are far from the training structure space, thus improving the prediction robustness under small sample material database conditions. This method can serve the initial screening, design, material database pre-evaluation, and subsequent standard experimental verification sample selection of impregnating agents for high-voltage power capacitors, and has significant application value in reducing the number of blind experiments, improving R&D efficiency, and promoting the development of impregnating agents from traditional aromatic hydrocarbon systems to environmentally friendly ester systems.

[0081] In the above embodiments, a molecular prediction method for the gas-getter performance of liquid impregnating agents has been described in detail. This application also provides embodiments corresponding to a molecular prediction device for the gas-getter performance of liquid impregnating agents. It should be noted that this application describes the embodiments of the device from two perspectives: one based on functional modules and the other based on hardware.

[0082] Figure 4 A block diagram of a molecular prediction device for the getter properties of a liquid impregnating agent, as provided in another embodiment of this application, is shown below. Figure 4 As shown, it includes: The modeling dataset construction module 11 is used to obtain the molecular descriptors of the impregnating agent molecular samples and measure the signed inhalation response values ​​of the impregnating agent molecular samples, and construct the corresponding modeling dataset. The partitioning module 12 is used to perform multiple hierarchical random partitioning of the modeling dataset according to the sign of the signed inspiratory response values. In the training set obtained by each partition, data preprocessing, feature selection and hyperparameter search are performed respectively to obtain the preprocessing rules, molecular descriptor set and at least one set of model parameter combinations for that partition, and apply them to the corresponding validation set. The prediction value fusion module 13 is used to obtain the out-of-fold prediction value by cross-validating based on the training set of each partition and each set of model parameters using at least two different types of machine learning models as base learners, and then obtain the initial prediction value of the training set by stacking and fusing the meta learners. The candidate model construction module 14 is used to correct the initial prediction value based on the initial prediction value and the corresponding signed inspiratory response value, combined with the confidence factor of the applicable domain of the molecular structure, to obtain the target prediction value, and to obtain the candidate model corresponding to each division and each combination of model parameters. The representative model determination module 15 is used to perform external validation on each candidate model using the corresponding validation set, obtain evaluation indicators, and determine and store the representative model according to preset rules. The processing and sorting module 16 is used to obtain the molecular descriptors of the predicted candidate molecules, perform the same preprocessing as the molecular descriptor set according to the preprocessing rules corresponding to the representative model, call the representative model to determine the corresponding asspirability prediction value, and output the priority sorting result according to each asspirability prediction value.

[0083] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0084] Figure 5 A structural diagram of an electronic device provided in another embodiment of this application, such as... Figure 5 As shown, the electronic device includes: a memory 20 for storing computer programs; The processor 21 is configured to execute a computer program to implement the steps of the molecular prediction method for the gas absorption performance of the liquid impregnating agent as described in the above embodiments.

[0085] The electronic devices provided in this embodiment may include, but are not limited to, smartphones, tablets, laptops, or desktop computers.

[0086] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one of the following hardware forms: Digital Signal Processor (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0087] The memory 20 may include one or more computer-readable storage media, which may be non-transitory. The memory 20 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 20 is used to store at least the following computer program 201, which, after being loaded and executed by the processor 21, is capable of implementing the relevant steps of the molecular prediction method for the gas-getter performance of the liquid impregnating agent disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 20 may also include an operating system 202 and data 203, and the storage method may be temporary or permanent storage. The operating system 202 may include Windows, Unix, Linux, etc.

[0088] In some embodiments, the electronic device may further include a display screen 22, an input / output interface 23, a communication interface 24, a power supply 25, and a communication bus 26.

[0089] Those skilled in the art will understand that Figure 5 The structures shown do not constitute a limitation on electronic devices and may include more or fewer components than those shown.

[0090] The electronic device provided in this application includes a memory and a processor. When the processor executes the program stored in the memory, it can implement the molecular prediction method for the gas absorption performance of the liquid impregnating agent described above and has the same beneficial effects.

[0091] Finally, this application also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps described in the above method embodiments.

[0092] It is understood that if the methods in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0093] The foregoing provides a detailed description of a molecular prediction method, apparatus, and medium for the gas absorption performance of a liquid impregnating agent. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0094] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A molecular prediction method for the gas-getter properties of liquid impregnating agents, characterized in that, include: Obtain the molecular descriptor of the impregnating agent molecular sample and measure the signed inspiratory response value of the impregnating agent molecular sample, and construct the corresponding modeling dataset; The modeling dataset is randomly divided into multiple stratified partitions according to the sign of the signed inspiratory response values. Data preprocessing, feature filtering, and hyperparameter search are performed in the training set obtained from each partition to obtain the preprocessing rules, molecular descriptor set, and at least one set of model parameter combinations for that partition, and then applied to the corresponding validation set. Using at least two different types of machine learning models as base learners, cross-validation is performed based on the training set partitioned each time and each set of model parameter combinations to obtain out-of-fold prediction values, and the initial prediction values ​​of the training set are obtained by stacking and fusing meta learners. Based on the initial predicted value and the corresponding signed inspiratory response value, the target predicted value is obtained by correcting the initial predicted value in combination with the confidence factor of the applicable domain of the molecular structure, and the candidate models corresponding to each division and each combination of model parameters are obtained respectively. The candidate models are externally validated using the corresponding validation set, evaluation indicators are obtained, and representative models are determined and stored according to preset rules. Obtain molecular descriptors of predicted candidate molecules, perform the same preprocessing as the molecular descriptor set according to the preprocessing rules corresponding to the representative model, call the representative model to determine the corresponding asspirability prediction value, and output the priority ranking result according to each asspirability prediction value.

2. The molecular prediction method for the gas-getter properties of liquid impregnating agents according to claim 1, characterized in that, The process of obtaining molecular descriptors for impregnating agent molecular samples and measuring the signed inspiratory response values ​​of the impregnating agent molecular samples, and constructing the corresponding modeling dataset, includes: Obtain the molecular structural formula or simplified molecular linear input specification of the impregnating agent molecular sample, use descriptor calculation software to determine the corresponding molecular descriptor, and form a molecular descriptor matrix based on the molecular descriptor; The labeled inhalation response values ​​of the impregnating agent molecule sample are measured and recorded, wherein a labeled inhalation response value greater than zero indicates inhalation, and a labeled inhalation response value less than zero indicates gas evolution. The molecular descriptor matrix is ​​cleaned to remove descriptors whose quality does not meet the preset conditions, so as to obtain a modelable descriptor matrix. The modelable descriptor matrix is ​​associated with the signed inspiratory response values ​​to construct the modeling dataset.

3. The molecular prediction method for the gas-getter properties of liquid impregnating agents according to claim 2, characterized in that, The step of cleaning the molecular descriptor matrix to remove descriptors whose quality does not meet preset conditions, in order to obtain a modelable descriptor matrix, includes: Descriptors with a missing rate higher than a first preset threshold are removed; Descriptors with variance below a second preset threshold are removed; For any two descriptors with a correlation coefficient higher than the third preset threshold, the one with a lower correlation to the marked inspiratory response value is removed.

4. The molecular prediction method for the gas-getter properties of liquid impregnating agents according to claim 2, characterized in that, The process of performing data preprocessing, feature selection, and hyperparameter search within each training set partition includes: The modelable descriptor matrix in the training set is imputed by missing median values ​​and standardized to obtain the preprocessing rules for this partitioning. The standardized modelable descriptor matrix is ​​subjected to feature filtering using a recursive feature elimination method to obtain the molecular descriptor set for this partitioning. Hyperparameter search is performed based on the set of molecular descriptors to obtain the model parameter combination for this partition.

5. The molecular prediction method for the gas-getter properties of liquid impregnating agents according to claim 1, characterized in that, The step of using at least two different types of machine learning models as base learners, and performing cross-validation based on the training set partitioned at each time and each set of model parameter combinations to obtain out-of-bounds prediction values ​​includes: The training set obtained from each partition is divided into multiple subsets; One subset is used as the validation subset and the remaining subsets are used as the training subsets to train each base learner and obtain the prediction value corresponding to the validation subset. By traversing each subset, the out-of-division prediction value of each base learner for all samples in the training set is obtained, wherein the out-of-division prediction value of each training sample comes from the base learner that does not include that sample.

6. The molecular prediction method for the gas-getter properties of liquid impregnating agents according to claim 5, characterized in that, The initial predicted value of the training set for this partition is obtained by stacking and fusing the meta-learners, including: The out-of-place prediction values ​​of each base learner are combined into a two-level input matrix; The meta-learner is trained using the second-level input matrix as input and the corresponding signed inhalation response value as output to obtain the initial prediction value.

7. The molecular prediction method for the gas-getter properties of liquid impregnating agents according to claim 1, characterized in that, The step of correcting the initial predicted value based on the initial predicted value and the corresponding signed inspiratory response value, combined with the confidence factor of the applicable domain of the molecular structure, to obtain the target predicted value includes: The residual is calculated based on the initial predicted value of the training set and the corresponding signed inspiratory response value, and a residual correction model is established with the molecular descriptor set as input and the residual as output. The applicable domain confidence factor of the molecular structure is determined based on the structural space of the training set; The initial predicted value is corrected according to the applicable domain confidence factor and the residual correction model to obtain the target predicted value; The expression for the target predicted value is: ； in, Let be the target predicted value for the i-th sample; The initial predicted value for the i-th sample; This is the residual shrinkage coefficient; The applicable domain confidence factor for the i-th sample; The residual correction model is for the i-th sample.

8. The molecular prediction method for the gas absorption properties of liquid impregnating agents according to any one of claims 1-7, characterized in that, Acquiring the evaluation indicators and determining and storing the representative model according to the preset rules includes: The evaluation metrics corresponding to each candidate model are obtained, including: coefficient of determination, root mean square error, mean absolute error, cross-validation prediction coefficient, external prediction coefficient, consistency correlation coefficient, and direction discrimination accuracy. After normalizing each evaluation index according to a preset weight, the weighted sum is obtained to get the corresponding evaluation result. Based on the evaluation results, the candidate model that meets the comprehensive performance requirements is selected as the representative model.

9. An electronic device, characterized in that, Includes memory used to store computer programs; A processor, configured to execute the computer program to implement the steps of the molecular prediction method for the gas absorption properties of a liquid impregnating agent as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the molecular prediction method for the gas absorption properties of a liquid impregnating agent as described in any one of claims 1 to 8.

Images