An electrostatic formation nanowire semiconductor VOCs identification method based on automatic learning and ensemble learning

Abstract

This invention discloses a method for identifying VOCs in electrostatically formed nanowire semiconductors based on automatic learning and ensemble learning. The identification method includes the following steps: acquiring a first dataset or a second dataset, wherein the first dataset includes the back-gate threshold voltage offset, back-gate subthreshold swing offset, and back-gate source-leakage current offset of the VOCs to be detected before and after electrostatically forming nanowires; the second dataset includes the junction-gate threshold voltage offset, junction-gate subthreshold swing offset, and junction-gate source-leakage current offset of the VOCs to be detected before and after electrostatically forming nanowires; processing the first dataset using a pre-trained CatBoost model, Stacking model, or Blending model to obtain the types of VOCs to be detected; or, processing the second dataset using a pre-trained CatBoost model, Stacking model, or Blending model to obtain the types of VOCs to be detected.

Description

Technical Field

[0001] This invention belongs to the field of volatile organic compound (VOCs) detection technology, specifically relating to a method for identifying VOCs in electrostatically formed nanowire semiconductors based on automatic learning and ensemble learning. Background Technology

[0002] Volatile organic compounds (VOCs) are widely generated in the industrial production processes of petrochemicals, pharmaceuticals, printed circuit boards, and building materials. They are also significant sources of fine particulate matter (PM2.5) and ozone, contributing to air pollution. They are also known to have harmful effects on the health and growth of humans, animals, and plants. Recently, it has been reported that different VOCs present in human exhaled breath can be used for the early diagnosis of cancer and other diseases. Accurate detection of VOCs is particularly important for air quality control, ecological restoration, and human health monitoring. Currently, most VOC detection devices are bulky and expensive. With the development of the Internet of Things (IoT) era, sensors need to be miniaturized, low-power, and compatible with large-scale integration.

[0003] Electrostatically formed nanowire (EFN) gas and VOCs sensors are robust, ultra-miniaturized sensors based on traditional CMOS processes, offering very low power consumption and compatibility with VLSI manufacturing. They are multi-gate field-effect transistors (FETs) with several degrees of freedom (such as the surrounding back gate and junction gate) capable of extracting parameters (such as threshold voltage, subthreshold swing, etc.) as input features for machine learning implementations.

[0004] Existing research has demonstrated the existence of fingerprint features and achieved some classification success. Currently, natural connectivity is used for feature fusion to combine two sets of experimental features. However, this method merges unrelated feature combinations from two sets into a single feature representation, leading to overfitting and inflated classification accuracy. Furthermore, the physical meaning of natural connectivity is not clear enough to interpolate missing values, further hindering improvements in classification accuracy. Summary of the Invention

[0005] To address the problems existing in the prior art, the purpose of this invention is to provide a method for identifying electrostatically formed nanowire semiconductor VOCs based on automatic learning and ensemble learning. This invention can achieve a better classification accuracy.

[0006] The technical solution adopted in this invention is as follows:

[0007] A method for identifying VOCs in electrostatically formed nanowire semiconductors based on automatic learning and ensemble learning includes the following process:

[0008] Obtain either dataset ΔBG (as the first dataset) or dataset ΔJG (as the second dataset), where dataset ΔBG includes the back gate threshold voltage shift (ΔV) of the VOCs to be detected before and after electrostatic formation of nanowires. th (BG)), back gate subthreshold swing offset (ΔSS(BG)), and back gate source-drain current offset (ΔI) on (BG)), the dataset ΔJG includes the junction gate threshold voltage shift (ΔV) of the VOCs to be detected before and after the formation of nanowires via electrostatics. th (JG)), gate subthreshold swing offset (ΔSS(JG)), and gate source-drain current offset (ΔI) on (JG));

[0009] The dataset ΔBG can be processed using a pre-trained CatBoost, Stacking, or Blending model to obtain the types of VOCs to be detected; alternatively, the dataset ΔJG can be processed using a pre-trained CatBoost, Stacking, or Blending model to obtain the types of VOCs to be detected.

[0010] Preferably, the CatBoost model is trained using the training set data corresponding to the ΔBG dataset or the training set data corresponding to the ΔJG dataset. The CatBoost model trained using the training set data corresponding to the ΔBG dataset is used to process the ΔBG dataset, and the CatBoost model trained using the training set data corresponding to the ΔJG dataset is used to process the ΔJG dataset.

[0011] The Stacking model is trained using ΔBG data or ΔJG data. The Stacking model trained on ΔBG data is used to process the ΔBG dataset, and the Stacking model trained on ΔJG data is used to process the ΔJG dataset.

[0012] The Blending model is trained using ΔBG or ΔJG data. The Stacking model trained on ΔBG data is used to process the ΔBG dataset, and the Blending model trained on ΔJG data is used to process the ΔJG dataset.

[0013] Preferred approach: Use a pre-trained Stacking model to process the dataset ΔBG;

[0014] The dataset ΔJG is processed using a pre-trained Blending model.

[0015] Preferably, the Stacking model is obtained by row integration of the CatBoost model, the LightGBM model, and the Extra Trees model;

[0016] The blending model is obtained by ensemble of the CatBoost model, LightGBM model, and Extra Trees model.

[0017] Preferred method: The process of acquiring the dataset ΔBG includes: measuring the junction gate voltage V before and after the VOCs to be identified are electrostatically formed on a SiO2 thin film on nanowires. JG Source-drain voltage V DS Back gate voltage V BG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage V DS Back gate voltage V BG and measuring drain current I DS Calculate the back gate threshold voltage shift (ΔV) of the VOCs to be identified before and after electrostatically forming a SiO2 thin film on the nanowires. th (BG)), back gate subthreshold swing offset (ΔSS(BG)), and back gate source-drain current offset (ΔI) on (BG));

[0018] The process of obtaining the dataset ΔJG includes: measuring the back gate voltage V before and after the VOCs to be identified are electrostatically formed on a SiO2 thin film on nanowires. BG Source-drain voltage V DS Junction gate voltage V JG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage V DS Back gate voltage V BG and measuring drain current I DS Calculate the junction gate threshold voltage shift (ΔV) of the VOCs to be identified before and after electrostatically forming a SiO2 thin film on the nanowires. th (JG)), gate subthreshold swing offset (ΔSS(JG)), and gate source-drain current offset (ΔI) on (JG)).

[0019] Preferably, the VOCs include one or a mixture of several of hexanol, butanol, propanol, ethanol, acetone and acetic acid.

[0020] This invention also provides an electrostatically formed nanowire semiconductor VOCs identification system based on automatic learning and ensemble learning, comprising:

[0021] Pre-computation module: Used to calculate either dataset ΔBG or dataset ΔJG, where dataset ΔBG includes the back gate threshold voltage shift (ΔV) of the VOCs to be detected before and after electrostatic formation of nanowires. th (BG)), back gate subthreshold swing offset (ΔSS(BG)), and back gate source-drain current offset (ΔI) on (BG)), the dataset ΔJG includes the junction gate threshold voltage shift (ΔV) of the VOCs to be detected before and after the formation of nanowires via electrostatics. th (JG)), gate subthreshold swing offset (ΔSS(JG)), and gate source-drain current offset (ΔI) on (JG));

[0022] The identification module is used to process the dataset ΔBG using a pre-trained CatBoost model, Stacking model, or Blending model to obtain the types of VOCs to be detected; or, it can process the dataset ΔJG using a pre-trained CatBoost model, Stacking model, or Blending model to obtain the types of VOCs to be detected.

[0023] The present invention also provides an electronic device, comprising:

[0024] One or more processors;

[0025] A storage device on which one or more programs are stored;

[0026] When the one or more programs are executed by the one or more processors, the one or more processors implement the electrostatically formed nanowire semiconductor VOCs identification method based on automatic learning and ensemble learning as described above.

[0027] The present invention also provides a storage medium, characterized in that it stores a computer program thereon, wherein the computer program, when executed by a processor, implements the electrostatically formed nanowire semiconductor VOCs identification method based on automatic learning and ensemble learning as described above.

[0028] The present invention has the following beneficial effects:

[0029] This invention designs classification features based on the physical detection principle of electrostatically formed nanowires, thereby achieving high classification accuracy. Specifically, based on the detection process, this invention selects features (i.e., dataset ΔBG or dataset ΔJG) from the physical parameters of the two directions in which the electrostatically formed nanowires are formed. Therefore, both model training and computation are significantly superior to existing technologies that use six data points—back-gate threshold voltage offset, back-gate subthreshold swing offset, back-gate source-drain current offset, junction-gate threshold voltage offset, junction-gate subthreshold swing offset, and junction-gate source-drain current offset—as input parameters for model training and computation. Furthermore, this invention allows datasets ΔBG and ΔJG to be used independently to train the classification model, eliminating various problems caused by natural feature connections. Using the method of this invention, a high classification accuracy (~96%) and comprehensive evaluation index are obtained. Attached Figure Description

[0030] Figure 1 Schematic diagram of bias configuration for electrostatically formed nanowire semiconductor VOCs sensor;

[0031] Figure 2 This is a schematic diagram illustrating the construction of independent feature sets for model training and VOCs identification when electrostatic nanowires adsorb VOCs in two different directions, as used in this invention.

[0032] Figure 3 The flowchart illustrates how the classification model trained in this invention is based on automatic learning and ensemble learning to complete VOCs model recognition.

[0033] Figure 4(a) shows the confusion matrix of the CatBoost classifier in the ΔBG dataset of this invention; Figure 4(b) shows the confusion matrix of the CatBoost classifier in the ΔJG dataset of this invention.

[0034] Figure 5(a) shows the prediction error of the CatBoost classifier in the ΔBG dataset of this invention; Figure 5(b) shows the prediction error of the CatBoost classifier in the ΔJG dataset of this invention.

[0035] Figure 6(a) shows the CatBoost classification report of the ΔBG dataset of this invention; Figure 6(b) shows the CatBoost classification report of the ΔJG dataset of this invention.

[0036] Figure 7(a) shows the classification error of the ensemble classifier of the present invention on the ΔBG dataset; Figure 7(b) shows the classification error of the ensemble classifier of the present invention on the ΔJG dataset;

[0037] Figure 8(a) shows the global feature importance distribution of the ΔBG dataset of this invention; Figure 8(b) shows the global feature importance distribution of the ΔJG dataset of this invention.

[0038] Figure 9(a) shows the distribution of local feature importance for each category in the ΔBG dataset of this invention; Figure 9(b) shows the distribution of local feature importance for each category in the ΔJG dataset of this invention. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings.

[0040] With the development of automatic learning technology, a series of open-source platforms have emerged, such as PyCaret, MLJAR, EvalML, FLAML, TPOT, H2O, and LightAutoML. This invention uses PyCaret to perform multi-classification of target analytes, obtaining the recognition accuracy and comprehensive evaluation index of different algorithms. Utilizing the calculation of feature parameters and the principle of VOCs detection, this invention constructs two independent datasets (i.e., dataset ΔBG or dataset ΔJG), and then uses each dataset for classification. The importance of classification features is analyzed, and their role in classification under different bias voltages (EFN formation direction) is analyzed from a physical perspective.

[0041] This invention proposes an algorithm framework with clear physical meaning, high classification accuracy, and automatic model training. The specific scheme is as follows.

[0042] The present invention provides a scheme for model establishment, training, and evaluation in an electrostatically formed nanowire semiconductor VOCs identification method based on automatic learning and ensemble learning, comprising the following steps:

[0043] Step 1, Experimental Subjects:

[0044] To achieve concentration-independent selectivity for the target VOC, 1000 ppm butanol, 1700 ppm propanol, 1800 ppm ethanol, 3000 ppm acetone, and 8800 ppm acetic acid were used to obtain the same response for all six molecules (~0.95). The volatile organic compound category labels—hexanol, butanol, propanol, ethanol, acetone, and acetic acid—are numbered sequentially from 0 to 5.

[0045] Step 2, Feature Selection:

[0046] Before and after the VOCs to be identified pass through the SiO2 thin film on the electrostatically formed nanowires, for different gate voltages V JG Source-drain voltage V DS The back gate voltage V is changed throughout the process. BG Measure the drain current I DS Then, the back gate threshold voltage (ΔV) of VOCs before and after passing through EFN is calculated. th (BG)), back gate subthreshold swing (ΔSS(BG)), and back gate source-drain current (ΔI) onThe change in (BG) is denoted as ΔBG. In another different set of experiments, the VOCs to be identified were analyzed before and after passing through the EFN at different V... BG and V DS Below, the entire process of changing V JG Measurement I DS Calculate the gate threshold voltage (ΔV) of the junction before and after VOCs pass through EFN. th (JG)), gate subthreshold swing (ΔSS(JG)), and gate source-drain current (ΔI) on The change in (JG) is denoted as ΔJG. These extracted parameters are used as the fingerprint characteristics of the target analyte. Back gate voltage V BG Gate voltage V JG Source-drain voltage V DS See configuration Figure 1 To make EFN symmetrical, take V JG =V JG1 =V JG2 The two sets of parameters correspond to the two formation directions of EFN, such as... Figure 2 As shown, datasets ΔBG and ΔJG are independent but have the same class label.

[0047] Step 3, Algorithm Architecture Design:

[0048] Volatile organic compound (VOC) classification is a multi-class classification problem. Ensemble learning, especially stacked algorithm frameworks, is better suited for multi-class classification and achieves higher accuracy. Ensemble learning can use multiple weak learning models to build a strong learning model. Determining which algorithm is best suited for classifying the data in this invention requires individual validation, and hyperparameter tuning is very time-consuming. Therefore, the overall algorithm architecture employs a strategy combining automatic learning and ensemble learning; the flowchart is shown below. Figure 3 .

[0049] Step 4, Outlier Check:

[0050] Dataset inspection aims to identify and correct identifiable errors in data files, including checking data consistency, handling invalid values, missing values, and outliers. This invention focuses on outlier issues; outliers at boundaries should be preserved.

[0051] Step 5, Interpolation:

[0052] In machine learning, due to the time-consuming nature of experiments, datasets are often expanded to obtain better results. Interpolation is a common method for expanding tabular data; therefore, this invention uses linear interpolation for data expansion.

[0053] Step 6, Classification model selection and optimization:

[0054] The classification models to be screened included algorithms such as K-Nearest Neighbors, Support Vector Machines, Decision Trees, Random Forests, XGBoost, CatBoost, and LightGBM. These models were compared using their F1 scores. Four models with an F1 score > 0.9 were selected: CatBoost, XGBoost, LightGBM, and Extra Trees. The hyperparameters of these models were fine-tuned to obtain the optimized models, which were then sorted again by F1 score to determine the best CatBoost model (see Table 1). Table 1 shows the hyperparameter-optimized models (F1 > 0.9) for ΔBG and ΔJG datasets. Table 1 shows that the highest F1 scores in both the ΔBG and ΔJG datasets were achieved by the CatBoost model. The average F1 scores for the ΔBG and ΔJG datasets were 0.9525 ± 0.0150 and 0.9602 ± 0.0124, respectively, with corresponding average accuracies of 95.27 ± 1.48% and 96.03 ± 1.23%. This indicates that the CatBoost model performs well on both the ΔBG and ΔJG datasets.

[0055] Table 1

[0056]

[0057]

[0058] Step 7, Model Integration:

[0059] For ΔBG and ΔJG datasets, single-model ensembles were performed. Four models with an F1 score greater than 0.9 were selected for ensemble learning, using ensemble methods including Bagging, Boosting, Blending, and Stacking. The first two ensemble methods are applicable to all four individual models. Blending and Stacking require multiple models; due to the time-consuming nature of XGBoost, this invention selected CatBoost, LightGBM, and Extra Trees models for ensemble learning. The ensemble learning models for the ΔBG and ΔJG datasets are shown in Table 2. The results indicate that the optimal ensemble model for the ΔBG dataset is the Stacking model, while the optimal ensemble model for the ΔJG dataset is the Blending model. For the ΔBG experiment, Stacking significantly improved accuracy, achieving 0.9571±0.0173 and an F1 score of 0.9569±0.0175. For the ΔJG experiment, the Blending ensemble model showed substantial improvement, achieving both accuracy and an F1 score of 0.9681±0.0097. Model ensemble not only improves accuracy and other metrics, but also reduces standard deviation.

[0060] Table 2

[0061]

[0062]

[0063] Step 8, Model Evaluation:

[0064] CatBoost classification was evaluated using confusion matrices (see Figures 4(a) and 4(b)), prediction errors (see Figures 5(a) and 5(b)) and classification reports (see Figures 6(a) and 6(b)). Confusion matrices were used to analyze the prediction results of the ΔBG and ΔJG experiments. In the confusion matrix, the sum of each row represents the number of samples in the actual class, and the sum of each column represents the number of samples predicted to belong to that class. The key role of the confusion matrix is ​​to summarize the number of correct and incorrect predictions by class. It is this feature importance analysis for each class that overcomes the limitations of holistic analysis.

[0065] Figures 5(a) and 5(b) illustrate the class prediction error calculated for each class using the Catboost classifier. The horizontal axis and legend correspond to the actual classes labeled 0 through 5 (in Figures 5(a) and 5(b), the numbers above each bar represent the labels from bottom to top, for example, in Figure 5(a), the first bar is labeled 05, indicating that the dark area below the bar corresponds to class 5 (i.e., acetic acid), and the dark area above the bar corresponds to class 0 (i.e., hexanol), and so on for the other bars). The vertical axis represents the number of predicted classified samples. The sequence of numbers at the top of the bars is the legend sequence from bottom to top. Figures 6(a) and 6(b) further illustrate the evaluation metrics for each classification result, such as precision, recall, and the F1 score calculated based on the confusion matrix of the two datasets. Since this invention uses the F1 score to optimize the model, relatively high precision and recall are required. As shown in Figures 5(a) and 6(a), ΔBG achieves the best classification results for labels 2 (propanol) and 5 (acetic acid), while labels 3 (ethanol) and 4 (acetone) are most prone to misclassification. As shown in Figures 5(b) and 6(b), ΔJG achieves the best classification results for labels 5 (acetic acid), 4 (acetone), and 2 (propanol), while labels 0 (hexanol) and 1 (butanol) are most frequently misclassified.

[0066] After model ensemble, this invention uses the best model to analyze the prediction results; that is, the ΔBG dataset is predicted using a stacked ensemble model, while the ΔJG dataset is predicted using a hybrid ensemble model. The intuitive results of the class prediction errors of the ensemble classifiers are shown in Figures 7(a) and 7(b). The coordinates and legends in Figures 7(a) and 7(b) are the same as those in Figures 5(a) and 5(b). The main color of each bar represents a prediction that is consistent with reality, while other colors represent misclassifications of that class. Compared to Figures 5(a) and 5(b), it can be observed that the classification results in Figures 7(a) and 7(b) are superior to those in Figures 5(a) and 5(b).

[0067] Step 9, Feature Importance Evaluation:

[0068] The contributions of features are helpful in studying the working mechanism of sensors, experimental planning, and improving classifier performance. SHAP analysis was used to plot the global feature importance distribution (Fig. 8(a) and Fig. 8(b)) and the local feature importance (Fig. 9(a) and Fig. 9(b)). Global feature importance, in a physical sense, reflects the influence of three features in the EFN formation process. Local feature importance shows the classification contribution of each feature for a specific VOC. This is helpful in studying the adsorption mechanism.

[0069] The contributions of features contribute to the study of sensor operating mechanisms, experimental planning, and improvement of classifier performance. Feature selection in decision trees provides some interpretability. The interpretability of model features includes global interpretability and local interpretability. The corresponding analytical tools are feature importance and Shapley Additional Interpretation (SHAP), used to plot the global feature importance distribution in Figure 8(a) and Figure 8(b) and the local feature importance in Figure 9(a) and Figure 9(b), respectively.

[0070] Figures 8(a) and 8(b) show the contribution of each feature in the classifiers for the ΔBG and ΔJG datasets, respectively. Figure 8(a) shows that the feature importance order for the adjusted CatBoost model classification on the ΔBG dataset is ΔSS(BG), ΔVth(BG), and ΔIon(BG). Physically, this ranking represents the influence of these three features on volatile organic compound (VOC) classification during the vertical formation of EFNs. Figure 8(b) shows that the feature importance order for the adjusted CatBoost model classification on the ΔJG dataset is ΔVth(JG), ΔSS, and ΔIon(BG). Physically, this order corresponds to the influence of the three features on the horizontal formation of EFNs. Furthermore, the ΔSS feature was applied for the first time for VOCs classification, and its overall performance was excellent, which was not emphasized in previous classifications. Specifically, in the ΔBG dataset (EFN vertical strata), ΔSS ranks first in importance among the three features, while in the ΔJG dataset (EFN horizontal strata), ΔSS ranks second in importance.

[0071] More specifically, Figures 9(a) and 9(b) show the contribution of each feature for each VOC. As can be seen from Figure 9(a), on the ΔBG dataset (EFN vertical formation), ΔSS contributes the most to features with class labels 0 (hexanol), 1 (butanol), 3 (ethanol), and 4 (acetone), while ΔVth(BG) contributes the most to features with class labels 2 (propanol) and 5 (acetic acid). Similarly, as can be seen from Figure 9(b), on the ΔJG dataset (EFN horizontal formation), ΔVth(JG) contributes the most to features with class labels 1 (butanol), 4 (acetone), and 5 (acetic acid), ΔSS(JG) contributes the most to features with class labels 0 (hexanol) and 2 (propanol), and ΔIon(JG) contributes the most to features with class label 3 (ethanol). This further demonstrates the importance of multiple degrees of freedom associated with multi-gate EFNs, which can extract different parameters that contribute differently to the selective gas identification through machine learning.

[0072] After the classification model is trained and tested, it can be used to classify the VOCs to be detected. The specific process is as follows:

[0073] Step 1, Collect data:

[0074] Before and after the VOCs to be identified are electrostatically formed onto a SiO2 thin film on nanowires, the junction gate voltage V is measured. JG Source-drain voltage V DS Back gate voltage V BG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage VDS Back gate voltage V BG and measuring drain current I DS Calculate the back gate threshold voltage shift (ΔV) of the VOCs to be identified before and after electrostatically forming a SiO2 thin film on the nanowires. th (BG)), back gate subthreshold swing offset (ΔSS(BG)), and back gate source-drain current offset (ΔI) on (BG)), thus obtaining the dataset ΔBG;

[0075] Alternatively, the back gate voltage V can be measured before and after the VOCs to be identified are electrostatically formed onto a SiO2 thin film on nanowires. BG Source-drain voltage V DS Junction gate voltage V JG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage V DS Back gate voltage V BG and measuring drain current I DS Calculate the junction gate threshold voltage shift (ΔV) of the VOCs to be identified before and after electrostatically forming a SiO2 thin film on the nanowires. th (JG)), gate subthreshold swing offset (ΔSS(JG)), and gate source-drain current offset (ΔI) on (JG)), thus obtaining the dataset ΔJG;

[0076] When actually collecting data, you only need to obtain one of the datasets ΔBG or ΔJG.

[0077] Step 2, categorized calculation:

[0078] The dataset ΔBG is processed using a pre-trained CatBoost model, Stacking model, or Blending model (in this case, the dataset ΔBG is obtained in step 1) to obtain the types of VOCs to be detected.

[0079] Alternatively, the dataset ΔJG can be processed using a pre-trained CatBoost, Stacking, or Blending model (in which case the dataset ΔJG was obtained in step 1) to obtain the types of VOCs to be detected.

[0080] In summary, this invention uses VOCs with the same responsivity as the identification target and electrostatic formation of nanowire semiconductor physical adsorption as the basic principle. It uses the physical parameters of the semiconductor response before and after VOCs access as the identification fingerprint. Through feature calculation, dataset construction, dataset checking, dataset balancing and expansion, algorithm framework design, selection of automatic learning framework, algorithm analysis and comparison, model optimization, model ensemble, and model evaluation, and includes feature importance analysis, model fusion, and feature fusion steps, it achieves model selection, optimization, and fusion to obtain the best classification accuracy. It has the following advantages: First, the feature set construction uses the physical parameters of VOCs before and after passing through the EFN with two different bias voltages during the detection process as features, with clear physical meaning and improved accuracy, reaching 0.9603±0.0123 for a single model. Second, automatic learning selects the best training model, greatly improving efficiency and reducing human intervention. Third, combined with ensemble learning, it further improves classification efficiency by 0.9681±0.0097. Fourth, it can analyze the classification feature importance of different directions of EFN formation from both the overall VOCs and individual VOCs. The different carbon chain lengths, functional groups, and charge distributions of VOCs with different molecular structures are helpful for studying the adsorption mechanism of FEN in VOC molecules.

Claims

1. A method for identifying VOCs in electrostatically formed nanowire semiconductors based on automatic learning and ensemble learning, characterized in that, The process includes the following: Acquire either a first dataset or a second dataset. The first dataset includes the back-gate threshold voltage shift, back-gate subthreshold swing shift, and back-gate source-drain current shift of the VOCs to be detected before and after electrostatic formation of nanowires. The second dataset includes the junction gate threshold voltage shift, junction gate subthreshold swing shift, and junction gate source-drain current shift of the VOCs to be detected before and after electrostatic formation of nanowires. Specifically, the process of acquiring the first dataset includes: measuring the junction gate voltage V before and after the VOCs to be identified form a SiO2 thin film on the nanowires via electrostatic formation. JG Source-drain voltage V DS Back gate voltage V BG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage V DS Back gate voltage V BG and measuring drain current I DS The back-gate threshold voltage shift, back-gate subthreshold swing shift, and back-gate source-drain current shift of the VOCs to be identified before and after the SiO2 film is formed on the nanowires by electrostatic deposition are calculated. The first dataset is denoted as ΔBG. The process of obtaining the second dataset includes: measuring the back-gate voltage V before and after the SiO2 film is formed on the nanowires by electrostatic deposition of the VOCs to be identified. BG Source-drain voltage V DS Junction gate voltage V JG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage V DS Back gate voltage V BG and measuring drain current I DS The junction gate threshold voltage shift, junction gate subthreshold swing shift, and junction gate source-drain current shift of the VOCs to be identified before and after the SiO2 film is formed on the nanowire by electrostatics are calculated. The second dataset is denoted as ΔJG. Classification model selection and optimization: The classification models to be selected include K-nearest neighbors, support vector machines, decision trees, random forests, XGBoost models, CatBoost models, and LightGBM models. These classification models are compared by F1 score, and four classification models with F1 > 0.9 are selected: CatBoost model, XGBoost model, LightGBM model, and Extra Trees model. The hyperparameters of these models are fine-tuned to obtain the optimized model, and then sorted again by F1 score to obtain the best model, CatBoost model. Model Ensemble: Single-model ensembles were performed for ΔBG and ΔJG datasets respectively. Four models with F1 > 0.9 were selected for ensemble learning. Ensemble methods included Bagging, Boosting, Blending, and Stacking. Bagging and Boosting were applicable to all four individual models. Blending and Stacking required multiple models, so CatBoost, LightGBM, and Extra Trees models were selected for ensemble. The optimal ensemble model for the ΔBG dataset was the Stacking model, and for the ΔJG dataset, the optimal ensemble model was the Blending model. The training is performed using either the training set data corresponding to the first dataset or the training set data corresponding to the second dataset. The first dataset is processed using the CatBoost model trained with the training set data corresponding to the first dataset, and the second dataset is processed using the CatBoost model trained with the training set data corresponding to the second dataset. The training is performed using either the training set data corresponding to the first dataset or the training set data corresponding to the second dataset. The Stacking model trained using the training set data corresponding to the first dataset is used to process the first dataset, and the Stacking model trained using the training set data corresponding to the second dataset is used to process the second dataset. The first dataset is trained using either the training set data corresponding to the first dataset or the training set data corresponding to the second dataset. The first dataset is processed using the blending model trained with the training set data corresponding to the first dataset, and the second dataset is processed using the blending model trained with the training set data corresponding to the second dataset. The first dataset is processed using a pre-trained CatBoost, Stacking, or Blending model to obtain the types of VOCs to be detected; alternatively, the second dataset is processed using a pre-trained CatBoost, Stacking, or Blending model to obtain the types of VOCs to be detected.

2. The method for identifying electrostatically formed nanowire semiconductor VOCs based on automatic learning and ensemble learning according to claim 1, characterized in that: The first dataset is processed using the pre-trained Stacking model; The second dataset is processed using the pre-trained Blending model.

3. The method for identifying electrostatically formed nanowire semiconductor VOCs based on automatic learning and ensemble learning according to claim 1, characterized in that, The Stacking model is obtained by integrating the CatBoost model, LightGBM model and ExtraTrees model. The blending model is obtained by ensemble of the CatBoost model, LightGBM model, and Extra Trees model.

4. The method for identifying VOCs in electrostatically formed nanowire semiconductors based on automatic learning and integrated learning according to claim 1, wherein the VOCs include one or a mixture of several of hexanol, butanol, propanol, ethanol, acetone and acetic acid.

5. A VOCs identification system for electrostatically formed nanowire semiconductors based on automatic learning and ensemble learning, characterized in that, include: Pre-calculation module: used to calculate either a first dataset or a second dataset. The first dataset includes the back-gate threshold voltage shift, back-gate subthreshold swing shift, and back-gate source-drain current shift of the VOCs to be detected before and after electrostatic formation of nanowires. The second dataset includes the junction gate threshold voltage, junction gate subthreshold swing, and junction gate source-drain current of the VOCs to be detected before and after electrostatic formation of nanowires. Specifically, the process of obtaining the first dataset includes: measuring the junction gate voltage V before and after the VOCs to be identified form a SiO2 thin film on the nanowires via electrostatic formation. JG Source-drain voltage V DS Back gate voltage V BG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage V DS Back gate voltage V BG and measuring drain current I DS The back-gate threshold voltage shift, back-gate subthreshold swing shift, and back-gate source-drain current shift of the VOCs to be identified before and after the SiO2 film is formed on the nanowires by electrostatic deposition are calculated. The first dataset is denoted as ΔBG. The process of obtaining the second dataset includes: measuring the back-gate voltage V before and after the SiO2 film is formed on the nanowires by electrostatic deposition of the VOCs to be identified. BG Source-drain voltage V DS Junction gate voltage V JG and drain current I DS Then, the junction-gate voltage V is obtained through measurement. JG Source-drain voltage V DS Back gate voltage V BG and measuring drain current I DS The junction gate threshold voltage shift, junction gate subthreshold swing shift, and junction gate source-drain current shift of the VOCs to be identified before and after the SiO2 film is formed on the nanowire by electrostatics are calculated. The second dataset is denoted as ΔJG. The identification module is used for classifying and optimizing models. The classification models to be screened include K-nearest neighbors, support vector machines, decision trees, random forests, XGBoost models, CatBoost models, and LightGBM models. These models are compared by their F1 scores, and four models with F1 > 0.9 are selected: CatBoost model, XGBoost model, LightGBM model, and Extra Trees model. The hyperparameters of these models are then fine-tuned to obtain the optimal model, and finally sorted by F1 score to obtain the best model, CatBoost model. And for model ensembles: ΔBG and ΔJG were each ensembled individually. Four models with F1 > 0.9 were selected for ensemble learning. Ensemble methods included Bagging, Boosting, Blending, and Stacking. Bagging and Boosting were applicable to all four individual models. Blending and Stacking required multiple models, so CatBoost, LightGBM, and Extra Trees were selected for ensemble. The optimal ensemble model for the ΔBG dataset was the Stacking model, and for the ΔJG dataset, the optimal ensemble model was the Blending model. And for model training: Training is performed using either the training set data corresponding to the first dataset or the training set data corresponding to the second dataset; the CatBoost model trained using the training set data corresponding to the first dataset is used to process the first dataset, and the CatBoost model trained using the training set data corresponding to the second dataset is used to process the second dataset; training is performed using either the training set data corresponding to the first dataset or the training set data corresponding to the second dataset; the Stacking model trained using the training set data corresponding to the first dataset is used to process the first dataset, and the Stacking model trained using the training set data corresponding to the second dataset is used to process the second dataset; training is performed using either the training set data corresponding to the first dataset or the training set data corresponding to the second dataset; the Blending model trained using the training set data corresponding to the first dataset is used to process the first dataset, and the Blending model trained using the training set data corresponding to the second dataset is used to process the second dataset; And it can be used to process the first dataset using a pre-trained CatBoost model, Stacking model, or Blending model to obtain the types of VOCs to be detected; or, to process the second dataset using a pre-trained CatBoost model, Stacking model, or Blending model to obtain the types of VOCs to be detected.

6. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the electrostatically formed nanowire semiconductor VOCs identification method based on automatic learning and ensemble learning as described in any one of claims 1 to 4.

7. A storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the electrostatically formed nanowire semiconductor VOCs identification method based on automatic learning and ensemble learning as described in any one of claims 1 to 4.

Images