A method for nanoplastic analysis using AIOF-induced confined localization and machine learning-assisted SERS

Abstract

The present invention discloses a method for AIOF-induced confined positioning and machine learning-assisted SERS for nanoplastic analysis, which belongs to the field of nanoplastic trace detection, high-precision quantification and classification technology, and includes the following steps: step 1, preparing In2O3 / FTO; step 2, preparing an AIOF substrate through a silver mirror reaction; step 3, preparing a nanoplastic dispersion and collecting a SERS spectrum; step 4, characterizing the performance of the AIOF substrate; step 5, preprocessing the spectral data; step 6, classifying nanoplastics dispersed in different environmental samples through a machine learning model based on the spectral data preprocessed in step 5. The present invention adopts the above-mentioned method for AIOF-induced confined positioning and machine learning-assisted SERS for nanoplastic analysis. The AIOF nanoarray SERS substrate obtained by a unique preparation method has a multi-layer nanowall array structure with interconnected holes, has a larger surface area, can provide more adsorption sites and hotspots, and thus produces a stronger SERS signal.

Description

Technical Field

[0001] The present invention relates to the technical field of nanoplastic trace detection, high-precision quantification and classification, and in particular to a method for using AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis. Background Art

[0002] With the mass production and widespread use of plastics, their accumulation in the environment is increasing. Microplastics are currently widely defined as solid plastic particles ≤5 mm in size composed of polymers, functional additives and other chemicals, and nanoplastics (<1 μm) are considered to be an extension of microplastics. Micro- and nanoplastics in the environment are ingested by animals and enter their bodies. After accumulating in various tissues, they are widely transferred through the food chain and pose a threat to human health. However, the lack of methods to quickly and non-destructively identify and quantify these plastics in complex biological matrices has hindered research on the accumulation of micro- and nanoplastics in animals.

[0003] Current SERS technology still faces many challenges in achieving high-sensitivity detection of nanoplastics in complex samples. While coffee rings, which enrich micro- and nanoplastics, can be easily obtained by simply allowing droplets to dry naturally on a flat SERS substrate, achieving high SERS detection sensitivity, the random deposition of nanoparticles during the natural drying process of the colloidal droplets results in the coffee rings formed on the solid substrate having an uneven shape and position, making the distribution of the enriched micro- and nanoplastics quite uneven and difficult to detect with the naked eye or under a low-magnification microscope at low concentrations. Furthermore, actual samples contain different types of plastics, along with Raman peaks from other impurities. This greatly increases the complexity of the Raman spectrum, making it extremely difficult to quickly distinguish the types of plastics in the sample and accurately quantify them. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for AIOF-induced confined positioning and machine learning-assisted SERS for nanoplastic analysis. The AIOF nanoarray SERS substrate obtained by a unique preparation method has a multi-layer nanowall array structure with interconnected holes, has a larger surface area, can provide more adsorption sites and hotspots, thereby generating a stronger SERS signal. The confinement effect brought by the grooves thereon induces the deposition of micro-nanoplastics at designated positions to a certain extent, provides a positioning strategy for SERS detection of low-concentration nanoplastics, reduces the interference of other impurities on SERS detection in complex environments, and avoids to a certain extent the problem that the position of coffee rings formed by low-concentration plastics after drying is difficult to locate using portable Raman.

[0005] To achieve the above objectives, the present invention provides a method for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, comprising the following steps:

[0006] Step 1: Preparation of In2O3 / FTO;

[0007] Step 2: Prepare the AIOF substrate through silver mirror reaction;

[0008] Step 3: preparing a nanoplastic dispersion and collecting a SERS spectrum;

[0009] Step 4: Characterization of AIOF substrate performance;

[0010] Step 5: Preprocess the spectral data;

[0011] Step 6: Based on the spectral data preprocessed in step 5, classify the nanoplastics dispersed in different environmental samples through a machine learning model.

[0012] Preferably, the specific operation of step 1 is: placing the cleaned FTO substrate with the conductive surface facing down into a polytetrafluoroethylene reactor filled with InCl3·4H2O and C2H5NS ethanol solution for reaction, and a golden yellow film grows on the FTO surface, taking it out and repeatedly rinsing it with ethanol and distilled water, and then placing it in an oven to dry;

[0013] The product is then annealed to obtain In2O3 / FTO, which is then taken out and cut into small pieces for scoring before use.

[0014] Preferably, in step 1, the amount of InCl3·4H2O is 5-10mM, the amount of C2H5NS is 10-30mM, the amount of ethanol solvent is 40mL, the reaction time is 10-15h, and the reaction temperature is 150-200°C;

[0015] The annealing parameters were set as follows: annealing at 400-500°C for 3-5h at a heating rate of 2-3°C / min.

[0016] Preferably, the specific operation of step 2 is: placing the In2O3 / FTO obtained in step 1 into a clean beaker, first adding silver ammonia solution, and then adding glucose solution, taking out the product after the reaction and rinsing it, and finally obtaining an Ag / In2O3 / FTO substrate with a surface uniformly loaded with high-density silver nanoparticles, that is, an AIOF substrate.

[0017] Preferably, in step 2, the concentration of the silver ammonia solution is 0.1-0.2 M, the dosage is 1-2 mL, the concentration of glucose is 0.1-0.5 M, the dosage is 0.1-1.0 mL, and the reaction is maintained for 10-15 min.

[0018] Preferably, the specific operation of step three is: suspending PET solid particles in a sodium dodecyl sulfate solution, ultrasonically treating the PET solid particles to obtain a PET monodisperse mother liquor, and preparing PS monodisperse mother liquors and PMMA monodisperse mother liquors of different particle sizes; subsequently, adding different volumes of the PET monodisperse mother liquors, PS monodisperse mother liquors, and PMMA monodisperse mother liquors to different environmental samples to prepare nanoplastic dispersions of different concentrations;

[0019] The above nanoplastic dispersion was dropped onto the AIOF substrate and dried before collecting the SERS signal.

[0020] Preferably, in step three, the amount of PET solid particles used is 20-30 mg, the solution is 10-15 mL of pure water containing 2% sodium lauryl sulfate, and the ultrasonic time is 30-45 min.

[0021] Preferably, the specific operation of step four is: using a Gemini 500SEM to characterize the SEM image of the AIOF substrate, using a rotating target X-ray diffractometer equipped with Cu Kα radiation to record the crystal structure, and collecting data in the range of 20° to 80°, using a JEM 2100F to perform TEM, energy dispersive X-ray spectroscopy and selected area electron diffraction characterization, the UV-visible absorption spectrum is measured in a TU-1950 dual-beam UV-visible spectrophotometer, and the Raman mapping is measured by a micro-confocal Raman spectrometer equipped with a 532 nm laser.

[0022] Preferably, the specific operation of step 5 is: the initial range of the SERS spectrum is 600-1800 cm -1 , cut to 700-1700cm -1 , background subtraction was performed on different nanoplastic spectra using BWSpec software. The peak intensities showing variations in different nanoplastic spectra were normalized to a standard scale between 0 and 1 using Min-Max to eliminate the impact of intensity differences on the model. Principal component analysis and t-distributed random neighbor embedding methods were used to project high-dimensional data onto a two-dimensional plane. Data visualization of nanoplastic categories scattered in different environmental samples was performed in principal component analysis and t-distributed random neighbor embedding scatter plots.

[0023] Preferably, the specific operation of step six is: using KNN, GDBT, CNN and Transformer algorithms to classify nanoplastics dispersed in different environmental samples, and revealing the relationship between the SERS characteristic peak intensity and the concentration of nanoplastics in the sample by linear fitting of the data. For each concentration, three different measurement points are taken to calculate the average value and standard deviation of the intensity. The model fitting is performed using the coefficient of determination R 2To evaluate, analyze and predict the exponential data, and draw the fitting curve by taking the logarithm of the concentration and the corresponding characteristic peak intensity;

[0024] The confusion matrix graphically describes the performance of the machine learning model in distinguishing nanoplastics in different environmental samples. The prediction performance is evaluated using a series of metrics, including accuracy, recall, precision, and F1 value:

[0025]

[0026] where tp, fp, tn, and fn represent the number of true positive, false positive, true negative, and false negative predictions, respectively;

[0027] For external validation, the established machine learning model was used to classify nanoplastics in different environmental samples to determine the classification accuracy.

[0028] Therefore, the present invention adopts the above-mentioned AIOF-induced confined localization and machine learning-assisted SERS method for nanoplastic analysis, which has the following beneficial effects:

[0029] (1) Efficient substrate preparation and performance advantages: The AIOF nanoarray SERS substrate obtained by a unique preparation method has a multi-layer nanowall array structure with interconnected holes, which has a larger surface area and can provide more adsorption sites and hotspots, thereby generating a stronger SERS signal. The confinement effect brought by the scratches on it can, to a certain extent, induce the deposition of micro-nanoplastics at designated locations, providing a positioning strategy for SERS detection of low-concentration nanoplastics. While reducing the interference of other impurities on SERS detection in complex environments, it also circumvents the problem that the location of coffee rings formed by low-concentration nanoplastics after drying is difficult to locate using portable Raman spectroscopy.

[0030] (2) Accurate classification capability in multiple scenarios: The four classification models developed based on machine learning can accurately classify different types of nanoplastics in multiple environments after scientific preprocessing and dimensionality reduction of Raman spectral data.

[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of an embodiment of a method for using AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis according to the present invention;

[0033] Figure 2This is a diagram of the morphology, structure, and properties of an AIOF substrate in accordance with an embodiment of a method for using AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis.

[0034] Figure 3 This is a performance characterization diagram of an AIOF substrate according to an embodiment of a method for using AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis according to the present invention;

[0035] Figure 4 This is a flow chart of data preprocessing, data dimensionality reduction, and ML classification model for an embodiment of a method for AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis according to the present invention;

[0036] Figure 5 This is an embodiment of the method for using AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis of nanoplastics in different environments using a machine learning model;

[0037] Figure 6 This is a quantitative analysis diagram of nanoplastics dispersed in different environmental samples according to an embodiment of the method for nanoplastic analysis using AIOF-induced confined localization and machine learning-assisted SERS;

[0038] Figure 7 The present invention provides a method for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, which includes a Raman spectrum of 4-MBA on a blank substrate and a SERS spectrum of 4-MBA adsorbed on an AIOF substrate.

[0039] Figure 8 The present invention is a method for AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis. The silver plating time is 11 minutes. The amount of InCl3·4H2O is 10 -8 M) Relationship between the SERS performance;

[0040] Figure 9 The present invention is a method for the analysis of nanoplastics by AIOF-induced confined localization and machine learning-assisted SERS. The silver plating time is different from that of 4-MBA (10 -8 M) the relationship between the SERS intensity;

[0041] Figure 10 The present invention is a method for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, which shows SERS spectra of nanoplastics with different particle sizes in pure water with AIOF as the substrate;

[0042] Figure 11 This is an evaluation of the nanoplastic classification performance of four models under different environments in an embodiment of the method for nanoplastic analysis using AIOF-induced confined localization and machine learning-assisted SERS, with indicators including accuracy, recall, precision, and F1 score.

[0043] Figure 12 The decision boundaries of different environments determined by KNN in an embodiment of a method for nanoplastic analysis using AIOF-induced confined localization and machine learning-assisted SERS of the present invention are as follows;

[0044] Figure 13 The present invention discloses a method for using AIOF-induced confined localization and machine learning-assisted SERS for nanoplastic analysis in different environments and shows the limit of detection (LOD) and actual detection sensitivity (True). DETAILED DESCRIPTION

[0045] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0046] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0047] The present invention provides a method for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, comprising the following steps:

[0048] Step 1: Prepare In2O3 / FTO: Place the clean FTO substrate, conductive side down, into a polytetrafluoroethylene reactor filled with InCl3·4H2O and C2H5NS ethanol solution. A golden film grows on the FTO surface. Remove the substrate and rinse it repeatedly with ethanol and distilled water, then place it in an oven overnight.

[0049] The product is then annealed to obtain In2O3 / FTO, which is then taken out and cut into small pieces for scoring before use.

[0050] The dosage of InCl3·4H2O is 5-10mM, the dosage of C2H5NS is 10-30mM, the dosage of ethanol solvent is 40mL, the reaction time is 10-15h, and the reaction temperature is 150-200℃;

[0051] The annealing parameters were set as follows: annealing at 400-500°C for 3-5h at a heating rate of 2-3°C / min.

[0052] Step 2: Prepare an AIOF substrate through the silver mirror reaction: Place the In2O3 / FTO obtained in step 1 into a clean beaker, first add silver ammonia solution, then add glucose solution. After the reaction, take out the product and rinse it clean to finally obtain an Ag / In2O3 / FTO substrate with a high-density silver nanoparticles uniformly loaded on the surface, i.e., an AIOF substrate.

[0053] The concentration of silver ammonia solution is 0.1-0.2M, the dosage is 1-2mL, the concentration of glucose is 0.1-0.5M, the dosage is 0.1-1.0mL, and the reaction is maintained for 10-15min.

[0054] Step 3: Prepare nanoplastic samples and acquire SERS spectra: PET solid particles are suspended in a sodium dodecyl sulfate solution and ultrasonically treated to obtain a PET monodisperse mother liquor. PS monodisperse mother liquors and PMMA monodisperse mother liquors of different particle sizes are also prepared. Subsequently, different volumes of the PET monodisperse mother liquors, PS monodisperse mother liquors, and PMMA monodisperse mother liquors are added to different environmental samples to prepare nanoplastic dispersions of different concentrations.

[0055] The above nanoplastic dispersion was dropped onto the AIOF substrate and dried before collecting the SERS signal.

[0056] The dosage of PET solid particles is 20-30 mg, the dispersion liquid is 10-15 mL of pure water containing 2% sodium lauryl sulfate, and the ultrasonic time is 30-45 minutes.

[0057] Step 4. Characterization of the performance of the AIOF substrate: The AIOF substrate was characterized by SEM images using a Gemini 500SEM, and the crystal structure was recorded using a rotating target X-ray diffractometer equipped with Cu Kα radiation, and data were collected in the range of 20° to 80°. TEM, energy dispersive X-ray spectroscopy, and selected area electron diffraction were performed using a JEM 2100F. The UV-visible absorption spectrum was measured in a TU-1950 dual-beam UV-visible spectrophotometer, and Raman mapping was determined by a micro-confocal Raman spectrometer equipped with a 532 nm laser. The above steps are as follows: Figure 1 As shown in (a) in .

[0058] Step 5: Preprocessing of spectral data: The initial range of SERS spectrum is 600-1800 cm -1 , cut to 700-1700cm -1, background subtraction was performed on different nanoplastic spectra using BWSpec software. The peak intensities showing variations in different nanoplastic spectra were normalized to a standard scale between 0 and 1 using Min-Max to eliminate the influence of intensity differences on the model. Principal component analysis and t-distributed random neighbor embedding methods were used to project high-dimensional data onto a two-dimensional plane. Data of different nanoplastic categories were visualized in scatter plots of principal component analysis and t-distributed random neighbor embedding methods.

[0059] Step 6: Classify the nanoplastic samples using a machine learning model based on the spectral data preprocessed in step 5: Use KNN, GDBT, CNN, and Transformer algorithms to classify nanoplastics dispersed in different environmental samples, and perform linear fitting on the data to reveal the relationship between the SERS characteristic peak intensity and the concentration of the nanoplastic sample. For each concentration, take three different measurement points to calculate the average and standard deviation of its intensity. The model fit is calculated using the coefficient of determination R 2 To evaluate, analyze and predict the exponential data, and draw the fitting curve by taking the logarithm of the concentration and the corresponding characteristic peak intensity;

[0060] The confusion matrix graphically describes the performance of the machine learning model in distinguishing nanoplastics in different environmental samples. The prediction performance is evaluated using a series of metrics, including accuracy, recall, precision, and F1 value:

[0061]

[0062] where tp, fp, tn, and fn represent the number of true positive, false positive, true negative, and false negative predictions, respectively;

[0063] For external validation, the established machine learning model was used to classify all nanoplastics added to pure water, Yangtze River water, and fish samples to determine the classification accuracy, e.g. Figure 1 As shown in (b) in .

[0064] Example 1

[0065] The present invention provides a method for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, comprising the following steps:

[0066] Step 1: Preparation of In2O3 / FTO: A clean FTO substrate, conductive surface facing downward, was placed in a Teflon reactor containing an ethanolic solution of InCl3·4H2O and C2H5NS. After 14 hours of reaction at 160°C, a golden film grew on the FTO surface. The film was removed and repeatedly rinsed with ethanol and distilled water before being placed in a 160°C oven overnight. Subsequently, the product was annealed at 450°C at a heating rate of 2°C / min for 4 hours to obtain In2O3 / FTO. Finally, the annealed In2O3 / FTO was removed and cut into 5×5 mm pieces, which were then scored and used for later use.

[0067] Step 2: Prepare the AIOF substrate through the silver mirror reaction: Place the prepared glass slide flat in a clean beaker, add 1 mL of 0.2 M silver ammonia solution, and then add 0.5 mL of 0.3 M glucose solution. Maintain the reaction for 11 minutes, then remove and rinse it clean. Finally, an Ag / In2O3 / FTO substrate with a uniform surface load of high-density silver nanoparticles, i.e., an AIOF substrate, can be obtained.

[0068] The four steps of AIOF production include growing In2S3 on the FTO surface (ISF), thermal annealing of In2S3 on the FTO surface to generate In2O3 (IOF), scratching the IOF surface, and in-situ growth of Ag nanoparticles on the IOF to generate Ag / In2O3 / FTO (AIOF). The In2S3 grown on the FTO in the first step is a three-dimensional structure composed of a multi-layer nanowall array structure with interconnected holes. The In2O3 obtained after thermal annealing effectively retains the three-dimensional structure of In2S3, such as Figure 2 Compared with the IOF substrate without silver coating, SEM analysis after in situ silver coating showed that the entire surface of the three-dimensional (3D) network architecture was uniformly decorated with densely packed Ag nanoparticles (Ag NPs) with an average particle size of 97 nm, as shown in Figure 2(a). Figure 2 (b) and Figure 2 As shown in (d) in the figure, the TEM image further proves that Ag NPs are uniformly attached to all surfaces of In2O3. Energy dispersive X-ray (EDX) analysis of the elemental composition and distribution of the silver-plated AIOF shows that it is composed of In, O and Ag, as shown in Figure 2 As shown in (c) in the figure, the XRD pattern directly confirms that In2S3 (47.99°) is completely transformed into In2O3 (30.58°) after thermal annealing, and the subsequently grown Ag NPs are elemental silver (38.12°, 44.28°, 64.43°). Figure 2As shown in (f) in the figure. The STEM image shows a clear boundary at the interface between Ag and In2O3, with lattice spacings of 0.146 and 0.115 nm, corresponding to the (220) crystal plane of Ag and the (742) crystal plane of In2O3, respectively. Figure 2 As shown in (e) in the figure. In2O3 has light absorption ability in the range of 200-350nm. In order to improve the SERS performance of the AIOF substrate under 785nm laser, the AIOF obtained by in situ growth of AgNPs on the IOF has a strong LSPR effect in the range of 400-800nm, as shown in Figure 2 As shown in (g), this is attributed to the localized surface plasmon resonance of Ag nanoparticles.

[0069] In order to better understand the enhancement mechanism of the AIOF substrate, FDTD simulations were performed on Ag NPs and Ag / In2O3. The FDTD results showed that from the Z-axis direction, a strong hotspot was generated in the nanogap between the Ag NPs; the interface between the Ag NPs and In2O3 also produced a weak enhancement, and by comparing the electric field intensity in the hotspot, it was found that the electric field intensity in the hotspot on the Ag / In2O3 was higher than that in the hotspot of the single Ag NPs under the same conditions, which was attributed to the synergistic enhancement between Ag and In2O3, as shown in Figure 2. Figure 2 As shown in (h) and (i).

[0070] Since the structure of IOF determines the arrangement state of Ag NPs subsequently grown on it, the number of hot spots formed per unit volume, and the adsorption effect on material molecules, the amount of InCl3·4H2O used in the growth of In2S3 was investigated under fixed reaction time, temperature, and raw material ratio. Figure 8 (a) shows the difference between the amount of InCl3·4H2O and 4-MBA (10 -8 The relationship between the SERS performance of 1074 cm -1 The histogram of the peak intensity shows that the SERS performance of the AIOF substrate is the best when the amount of InCl3·4H2O is 8mM. Figure 8 As shown in (b) in the figure. The SEM image shows that when the amount of InCl3·4H2O is 8mM, the In2O3 on the IOF surface presents the highest roughness and the most pore structures for connecting different layers, as shown in Figure 8 This provides more Ag NPs growth sites, thereby increasing the number of hot spots formed per unit volume, and also provides more adsorption sites for the probe. Figure 3 (a)-(b) and Figure 9(a)-(b) show the relationship between the silver plating time and the SERS performance of 4-MBA (10-8M) at the optimal InCl3·4H2O dosage (8mM). The AIOF substrate has the best SERS performance when the silver plating time is 11 minutes.

[0071] The SEM images show that before the silver plating time is 11 minutes, the accumulation of Ag nanoparticles on the entire surface of the three-dimensional network architecture gradually becomes denser and the nano-gap gradually decreases with the increase of time. When the silver plating time is 12 minutes, many silver dendrites grow on the surface, which seriously weakens the LSPR effect of Ag NPs. Figure 9 As shown in (c)-(f) in the figure, this optimized AIOF substrate was used for subsequent detection of probe molecules and micro-nanoplastics in different environmental samples.

[0072] Step 3. Prepare nanoplastic dispersion and collect SERS spectra: First, 25 mg of PET solid particles were suspended in 10 mL of 2% sodium dodecyl sulfate (SDS) solution and ultrasonically treated for 30 minutes to obtain 10 mL of PET monodisperse mother liquor (2.5 mg / mL). PS monodisperse mother liquors and PMMA monodisperse mother liquors (2.5 mg / mL) of different particle sizes were purchased directly. Subsequently, nanoplastic dispersions in different environments were obtained by mixing the mother liquors and real samples. The above nanoplastic dispersions were dropped on an AIOF substrate and dried before collecting SERS signals.

[0073] SERS spectra were acquired using a portable Raman spectrometer (i-Raman Plus, B&W Tek, USA) using a 50× microscope objective lens, operating with an excitation laser wavelength of 785 nm, an integration time of 3 s, and a spectral range of 600–1800 cm . -1 , all spectral data were collected using the AIOF substrate as the SERS substrate.

[0074] Step 4: Characterization of AIOF substrate performance: SEM images were taken using a Gemini 500SEM equipped with Cu Kα radiation. The crystal structure was recorded using a rotating target X-ray diffractometer (Rigaku D / MAX2500VL / PC), collecting data from 20° to 80°. TEM, energy-dispersive X-ray spectroscopy (EDS), and selected-area electron diffraction were performed using a JEM 2100F instrument. UV-visible absorption spectra were measured using a TU-1950 dual-beam UV-visible spectrophotometer. Raman mapping was determined using a confocal Raman microscope (LabRam HR Evolution) equipped with a 532 nm laser.

[0075] Step 5: Preprocessing of spectral data: The initial range of SERS spectrum is 600-1800 cm -1 , cut to 700-1700cm -1 ,like Figure 4 As shown in (a). Background subtraction was performed on the spectra of different nanoplastics using BWSpec software. The peak intensities that showed variations in the spectra of different nanoplastics were normalized to a standard scale between 0 and 1 using Min-Max to eliminate the impact of intensity differences on the model. To reduce dimensionality, principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) methods were used to project the high-dimensional data into a two-dimensional (2D) plane. Data visualization of three different nanoplastic categories (PS, PMMA, PET) was performed in PCA and t-SNE scatter plots. These illustrations were drawn based on two principal components.

[0076] Figure 4 (b) in the figure shows the preprocessing process, where I is the baseline correction diagram, ⅠⅠ is the smoothing diagram, and ⅢⅠ is the normalization diagram. When using Raman spectroscopy data for machine learning, data preprocessing steps such as baseline correction, smoothing, and normalization are critical to ensuring model performance. These operations can effectively eliminate noise interference, unify data scales, and highlight the actual chemical information, thereby significantly improving the model's accuracy, robustness, and generalization ability.

[0077] The first step is baseline correction. BWSpec software is used to fit the obtained Raman data and subtract the background signal to restore the true peak height and peak shape.

[0078] The purpose of subsequent smoothing is to reduce noise. Savitzky-Golay filter is used for spectral smoothing because it can reduce noise while retaining the peak shape. This can reduce the interference of noise on the machine learning model and make it easier to learn the real features in the spectral data.

[0079] Finally, normalization is required. Under different experimental conditions, such as laser power or sample concentration, the intensity of the Raman signal can vary significantly. Using Min-Max normalization, the spectral data is scaled to a range of 0-1, eliminating the influence of intensity differences on the model.

[0080] After performing the above preprocessing on the Raman spectral data, it is possible to effectively prevent certain features from dominating the training process due to their large size. After this preprocessing, the Raman spectral data were then subjected to dimensionality reduction using principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), respectively. Both PCA and t-SNE scatter plots demonstrate the potential for accurate differentiation among different types of plastic samples, despite some overlap. Figure 4(c) shows that t-SNE dimensionality reduction produces better clustered data than PCA. This is because t-SNE, as a nonlinear dimensionality reduction method, focuses on preserving the local structure between data points and is particularly good at visualizing high-dimensional data such as Raman spectra, which contain a lot of characteristic information.

[0081] Step 6. Classify the nanoplastic samples through machine learning models based on the spectral data preprocessed in step 5: Among them, the four machine learning models are: two traditional machine learning models (KNN and GDBT) and two deep learning models (CNN and Transformer). After preprocessing the original spectra, their classification capabilities in different scenarios are evaluated. The confined area positioning SERS on the AIOF nanoarray and the customized Raman spectroscopy processing method based on machine learning are used to achieve accurate classification and trace measurement of nanoplastics in complex conditions in multiple scenarios.

[0082] The KNN, GDBT, CNN, and Transformer models were all trained using only preprocessed data. Nanoplastic samples were classified using the KNN, GDBT, CNN, and Transformer models. To ensure comparability and consistency, all models used the same training and test sets. Visualization of the decision boundaries of the KNN dimensionality reduction components highlighted the distinct regions corresponding to each nanoplastic type, highlighting the model's ability to distinguish between PS, PMMA, and PET plastic samples. This demonstrates the stability of the AIOF substrate in complex environments and the model's applicability in real-world applications.

[0083] K-Nearest Neighbors (KNN): Optimizes key parameters using a grid search, evaluates the number of neighborhood samples in the interval [3, 5], and tests the performance of the "uniform" (uniformly weighted) and "distance" (inversely proportional to distance) weighting strategies.

[0084] Gradient Boosted Decision Tree (GBDT): Bayesian optimization was used for hyperparameter tuning. Key parameters adjusted included the number of base learners (100 decision trees), the maximum depth of a single tree ([3, 5, 10]), the minimum number of samples for node splitting ([1, 2, 5]), the minimum number of samples for leaf nodes ([1, 3, 5]), and the maximum depth of the tree structure ([3, 5]).

[0085] Convolutional Neural Network (CNN): This model builds a one-dimensional convolutional architecture with deep feature extraction capabilities. The network comprises three feature extraction modules (each consisting of Conv1D, MaxPooling1D, and BatchNorm). By stacking convolutional layers, the network gradually expands the receptive field to capture global features while effectively avoiding interference from local features. The classifier uses a three-layer fully connected structure, with a sigmoid activation function at the end to achieve classification decisions.

[0086] Transformer network: A hybrid architecture is constructed to target high-dimensional feature data. The first layer uses one-dimensional convolution for local feature extraction and dimensionality compression. Long-range dependencies are then captured through three Transformer encoding layers (including a multi-head self-attention mechanism and a feedforward network). Finally, features are integrated through one-dimensional convolutional layers and fully connected layers, and the sigmoid activation function is also used at the end to complete the classification task.

[0087] Quantitative analysis of nanoplastics:

[0088] In order to reveal the relationship between the SERS characteristic peak intensity and the concentration of nanoplastics, a linear fit was performed on the data. For each concentration, three different measurement points were taken to calculate the average value and standard deviation of the intensity. The coefficient of determination (R 2 ) was evaluated, and the exponential data were analyzed and predicted by taking the logarithm of the concentration and the corresponding characteristic peak intensity, and then drawing the fitting curve.

[0089] The confusion matrix is ​​used to graphically describe the performance of the machine learning model in distinguishing nanoplastics. The prediction performance is evaluated using a series of metrics, including accuracy, recall, precision, and F1 value:

[0090]

[0091] where tp, fp, tn, and fn represent the number of true positive, false positive, true negative, and false negative predictions, respectively.

[0092] For external validation, the established machine learning model was used to classify nanoplastics dispersed in different environmental samples to determine the classification accuracy.

[0093] In order to quantify the contribution of the AIOF substrate to the enhancement effect, its enhancement factor EF was calculated according to the following formula:

[0094] EF=(I SERS / I BULK )×(N BULK / N SERS );

[0095] Among them, I SERSis the intensity of the SERS spectrum, I BULK is the intensity of ordinary Raman spectrum, N SERS is the average number of molecules in the laser spot excited by SERS, N BULK is the average number of molecules in the laser spot for ordinary Raman excitation.

[0096] Example 2

[0097] In order to demonstrate the applicability of this method for detecting nanoplastics under natural environmental water and biological conditions, taking the Yangtze River ecology as an example, Yangtze River water and bream widely distributed in the Yangtze River basin were selected as models to evaluate the precise quantitative ability of the AIOF substrate, and three types of nanoplastics (PS, PMMA, PET) were introduced into pure water, Yangtze River water and fish meat.

[0098] For pure water samples: PS (20-400 nm), PMMA and PET solutions (W-PS20, W-PS100, W-PS200, W-PS400, W-PMMA and W-PET) of different concentrations were obtained by directly diluting the respective mother solutions with ultrapure water.

[0099] For the Yangtze River water samples: PS200 and PMMA solutions (Y-PS, Y-PMMA) of different concentrations were obtained by directly diluting the mother liquor with Yangtze River water.

[0100] For biological samples: Step 1: Obtain about 20g of fish tissue from bream obtained from the Yangtze River basin and grind it into fish paste; Step 2: Take 2g of fish paste and add different masses of PS2000 and PET plastics to prepare mixed samples with different mass gradients; Step 3: Add the above mixed samples to 20mL of 15% (w / w) tetramethylammonium hydroxide and ultrasonicate at room temperature for 60min to fully digest the mixed samples; Step 4: Centrifuge at 2000rpm for 2min to completely precipitate the protein-loaded nanoplastic with a small amount of protein; Step 5: After removing the supernatant, add 100μL of ultrapure water to redisperse the remaining trace protein and extracted nanoplastics in the precipitate (B-PS, B-PET).

[0101] For the detection of probe molecules, the AIOF substrate was immersed in 200 μL of 4-MBA solution of different concentrations for 30 minutes and then taken out to collect SERS signals; for the detection of nanoplastics in water samples, 60 μL of nanoplastic dispersion droplets were dried on the substrate surface and then SERS signals were collected; for the detection of fish samples, 60 μL of redispersed droplets of different mass concentrations were taken and dried on the AIOF substrate and then SERS signals were collected.

[0102] First, the SERS performance of the prepared AIOF substrate was tested after being soaked in the classic Raman reporter dye 4-mercaptobenzoic acid (4-MBA). The Raman spectra of 4-MBA gradient concentrations were obtained using a portable Raman spectrometer. The measurement conditions were: laser wavelength 785nm, exposure time 3s. The AIOF substrate was soaked in different concentrations (10 -8 M to 10 -12 M) in 4-MBA solution for 30 min, the lowest concentration detected was 10 -12 M, such as Figure 3 Then, select 1074cm -1 The Raman intensity at was used to generate a standard curve, which showed good linear correlation at various 4-MBA concentrations, as shown in Figure 4. Figure 3 As shown in (e), the detection limit (LOD) of AIOF substrate for 4-MBA is 2.4×10 -13 .

[0103] like Figure 7 As shown in Figure 2, the enhancement factor (EF) of the substrate after 4-MBA solution was dropped and dried on the AIOF substrate was calculated, and the EF was 5.045×10 10 In addition, the uniformity of SERS signal is a key factor in determining the reliability of SERS substrate. Taking W-PS200 in the actual sample as a model, at 1002 cm -1 The main peak position of the 20×20μm containing 400 measurement points 2 The Raman mapping within the range shows the uniformity of the Raman signal within a small range, e.g. Figure 3 As shown in (g); at 1002cm -1 The main peak position of the 5×5mm 2 The scratches on the substrate show the uniformity of the Raman signal over a large range within the scratches on the substrate, with an RSD of 7.03%. Figure 3 As shown in (f) in .

[0104] Since the specific intensity in the SERS spectrum is proportional to the analyte concentration, accurate and sensitive quantitative analysis can be achieved by detecting the intensity of the characteristic peak. -1 、PMMA:812cm -1 、PET:1616cm -1 ) as a feature, the concentration gradient of the collected nanoplastics exceeding the LOD and the intensity of the corresponding features were logarithmically processed and then linearly fitted to achieve high-precision quantification.

[0105] In order to verify the detection ability of the AIOF substrate for different types of nanoplastics commonly found in the environment, W-PS20, W-PS100, W-PS200, W-PS400, W-PMMA and W-PET dispersed in pure water were detected. Figure 3 (c) shows the characteristic peaks of each nanoplastic identified in the SERS spectrum: PS at 1002 cm -1 (in-plane breathing vibration of benzene ring), 1030 cm -1 and 1600cm -1 : Symmetrical and asymmetric stretching vibrations of benzene ring (CC bond vibration); PMMA at 811cm -1 (C-O-C bond stretching vibration) and 1452 cm -1 (CH bond bending vibration), 1730cm -1 (Stretching vibration of C=O bond in ester group); PET at 1100cm -1 (bending vibration of O-CH2 in ethylene glycol unit), 1616 cm -1 (CC bond stretching vibration), 1280cm -1 (stretching vibration of CO bond in ester group) and 1730cm -1 (Stretching vibration of C=O bond in ester group). The detection sensitivity of the substrate for PS with the same particle size and different types of plastics is 500ppb (W-PS20), 250ppb (W-PS100), 1ppm (W-PS400), 1ppm (W-PS200), 2.5ppm (W-PMMA) and 5ppm (W-PET), respectively, and has a good linear relationship, such as Figure 3 (h)-(i) and Figure 10 The significantly lower detection limit of PS can be attributed to its intrinsically higher Raman activity, especially compared with PMMA and PET.

[0106] The present invention collected 800 Raman spectral data of W-PS200, W-PMMA and W-PET nanoplastics with different concentrations dispersed in pure water through AIOF substrate, and their concentrations were all higher than their respective LODs. Figure 4 (d) Training of the four models. Subsequently, Raman data of different concentrations dispersed in pure water (94 for W-PET, 94 for W-PMMA, and 103 for W-PS) dispersed in Yangtze River water (110 for Y-PMMA and 116 for Y-PS) and dispersed in fish samples (89 for B-PET and 110 for B-PS) were collected using an AIOF substrate. After data preprocessing, these data served as test sets for the four models in the three environments. The concentrations of these data were all above the LOD.

[0107] Figure 6 The LOD range of nanoplastics in all environments ( Figure 12 ) showed a good linear relationship, and the SERS spectra of microplastics in these actual environments were highly consistent with their intrinsic Raman spectra. There was no additional Raman characteristic peak in the Yangtze River water sample, but the 757cm -1 and 950cm -1 The Raman characteristic peaks at the positions are attributed to incompletely separated fish digestion products. These substances will not interfere with the reading and identification of the intensity of the characteristic peaks of microplastics.

[0108] In common pure water samples, the Raman spectra of W-PMMA and W-PET on the AIOF substrate are highly consistent with their intrinsic Raman spectra ( Figure 6 (a), (c)), and the detection sensitivity reached a very high level (W-PMMA: 2.5ppm, W-PET: 5ppm) ( Figure 6 (b), (d)). The Raman patterns of Y-PS and Y-PMMA in the Yangtze River water sample with similar environment have no obvious changes ( Figure 6 (e), (g)), the detection sensitivity (Y-PS: 1ppm, Y-PMMA: 10ppm) was not significantly reduced ( Figure 6 (f) and (h) in the figure. This is attributed to the fact that there are no large pollutants in the Yangtze River water that can easily cover the substrate, so there is no strong interference with the signal of the dried nanoplastics.

[0109] However, the detection sensitivity of AIOF substrate for B-PS and B-PET in biological samples was significantly reduced (B-PS: 10 μg / g; B-PET: 50 μg / g) ( Figure 6 (j) and (l) in the figure, and a new wavelength of 757 cm was added in both B-PS and B-PET samples. -1 and 950cm -1 The Raman characteristic peak at the position ( Figure 6 This is attributed to the fact that even after alkaline digestion and centrifugal separation steps, a small amount of fish digestion products will remain in the biological samples mixed with microplastics, which will adhere to the AIOF substrate during the drying process, thereby hindering the nanoplastics from entering the hot spot area of ​​the substrate to a certain extent, resulting in a decrease in the SERS signal.

[0110] We further evaluated two traditional machine learning models (KNN and GDBT) and two deep learning models (CNN and Transformer) ( Figure 4(d) in the figure). KNN and GDBT are common supervised learning multi-classification models. Considering the complexity of Raman spectral data in real samples, especially fish samples, two deep learning models, CNN and Transformer, were developed to classify plastic samples. For water environments such as pure water and Yangtze River water, the difference between traditional machine learning models and deep learning is not significant. However, for the classification of nanoplastics in biological samples, the deep learning Transformer (96%) and CNN (99%) models significantly outperformed the traditional machine learning KNN (52%) and GDBT (62%) models in terms of accuracy. Figure 5 (a) and Figure 10 This may be due to the fact that when testing plastics in bream meat, the digestion products of fish meat cannot be completely separated from microplastics, thus introducing additional Raman characteristic peaks that are not in the training set, such as Figure 5 As shown in (b) in .

[0111] Figure 5 (c) shows that W-PS, W-PMMA, and W-PET clusters dispersed in pure water are located in three different regions, among which only the W-PS and W-PMMA clusters have a slight overlap, as shown in the confusion matrix diagram. The KNN model has a 1.94% probability of misidentifying PS as PMMA in pure water, but the recognition accuracy of PMMA is 100%, as shown in the figure. Figure 5 As shown in (e) in , this is consistent with the KNN decision boundary diagram showing that only a small number of points representing PS enter the PMMA area, as shown in Figure 5 As shown in (d) in the figure, GDBT has a 0.97% probability of identifying W-PS as W-PMMA and a 1.06% probability of identifying W-PMMA as W-PS. Figure 5 As shown in (f), the recognition accuracy of the two models for W-PET is 100%, which is consistent with the results; the Y-PS and Y-PMMA clusters dispersed in the Yangtze River water are in the same area as the clusters of similar nanoplastics dispersed in pure water, and even completely overlap and are completely independent, which corresponds to Figure 4 The confusion matrix diagrams (i) and (j) show that the KNN and GDBT classification accuracy for nanoplastics in the Yangtze River water is 100%, as shown in Figure 2. Figure 11As shown in (a); B-PS and B-PET dispersed in biological samples are clustered in completely different areas from similar nanoplastics dispersed in pure water. The reason for this phenomenon may be that the fish sample is doped with the Raman characteristic peaks of high-intensity proteins after fish digestion. Among them, the more dispersed B-PET scatter points are located in the W-PS, W-PMMA, and W-PET clusters dispersed in pure water. This is manifested in the confusion matrix diagram where the KNN and GDBT models incorrectly identify PET as PS and PMMA in large quantities, as shown in Figure 2. Figure 4 As shown in (m) and (n) in .

[0112] The Transformer and CNN deep learning models have a 100% recognition accuracy for nanoplastics dispersed in pure water and Yangtze River water, and 96% and 99% accuracy for PS and PET dispersed in fish samples, respectively. The confusion matrix of the Transformer classification accuracy is shown in the figure below. Figure 5 The confusion matrix diagram of CNN classification accuracy is shown in (g), (k), and (o). Figure 5 This is shown in (h), (l), and (p) in the figure. This is attributed to the CNN convolutional layer's excellent ability to extract local features. By stacking convolutional layers, the receptive field is increased, resulting in better extraction of characteristic peaks. Therefore, with only a small amount of pure water data, it is possible to achieve near-100% accuracy in detecting nanoplastics in Yangtze River water and fish samples.

[0113] The Transformer has strong global feature extraction capabilities and can also accurately identify characteristic peaks. However, its attention mechanism calculates similarity for every peak in the data, causing the presence of mixed peaks in fish samples to affect the model's nanoplastic classification performance. This also demonstrates that the developed CNN and Transformer deep learning models, after training with only a small amount of Raman data from micro-nanoplastics dispersed in pure water, can accurately identify nanoplastics in a variety of real-world samples and have broad applicability.

[0114] Therefore, the present invention adopts the above-mentioned AIOF-induced confined positioning and machine learning-assisted SERS method for nanoplastic analysis. The AIOF nanoarray SERS substrate obtained by a unique preparation method has a multi-layer nanowall array structure with interconnected holes, has a larger surface area, and can provide more adsorption sites and hotspots, thereby generating a stronger SERS signal.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, characterized by: The following steps are involved: Step 1: Preparation of In2O3 / FTO; The specific operation of step 1 is as follows: the cleaned FTO substrate is placed with the conductive surface facing down in a polytetrafluoroethylene reactor filled with InCl3·4H2O and C2H5NS ethanol solution for reaction. A golden film grows on the FTO surface, which is then removed and repeatedly rinsed with ethanol and distilled water, and then placed in an oven to dry; The product is then annealed to obtain In2O3 / FTO, which is then taken out and cut into small pieces for scoring before use; In step 1, the amount of InCl3·4H2O is 5-10mM, the amount of C2H5NS is 10-30mM, the amount of ethanol solvent is 40mL, the reaction time is 10-15h, and the reaction temperature is 150-200℃; The annealing parameters were set as follows: annealing at 400-500 °C for 3-5 h at a heating rate of 2-3 °C / min; Step 2: Prepare the AIOF substrate through silver mirror reaction; The specific operation of step 2 is as follows: put the In2O3 / FTO obtained in step 1 into a clean beaker, add silver ammonia solution first, and then add glucose solution. After the reaction, take out the product and rinse it clean, finally obtaining an Ag / In2O3 / FTO substrate with a high density of silver nanoparticles uniformly loaded on the surface, i.e., an AIOF substrate; In step 2, the concentration of silver ammonia solution is 0.1-0.2M, the dosage is 1-2mL, the concentration of glucose is 0.1-0.5M, the dosage is 0.1-1.0mL, and the reaction is maintained for 10-15min; Step 3: preparing a nanoplastic dispersion and collecting a SERS spectrum; Step 4: Characterization of AIOF substrate performance; Step 5: Preprocess the spectral data; Step 6: Based on the spectral data preprocessed in step 5, classify the nanoplastics dispersed in different environmental samples through a machine learning model.

2. The method for nanoplastic analysis using AIOF-induced confined localization and machine learning-assisted SERS according to claim 1, characterized in that: The specific operation of step three is as follows: PET solid particles are suspended in a sodium dodecyl sulfate solution, which is ultrasonically treated to obtain a PET monodisperse mother liquor, and PS monodisperse mother liquors and PMMA monodisperse mother liquors of different particle sizes are prepared; subsequently, different volumes of the PET monodisperse mother liquor, PS monodisperse mother liquor, and PMMA monodisperse mother liquor are added to different environmental samples to prepare nanoplastic dispersions of different concentrations; The above nanoplastic dispersion was dropped onto the AIOF substrate and dried before collecting the SERS signal.

3. The method for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS according to claim 2, characterized in that: In step 3, the amount of PET solid particles used is 20-30 mg, the solution is 10-15 mL of pure water containing 2% sodium lauryl sulfate, and the ultrasonic time is 30-45 min.

4. The method of claim 1 for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, characterized in that: The specific operations of step 4 are as follows: the AIOF substrate is characterized by SEM images using a Gemini 500 SEM, the crystal structure is recorded using a rotating target X-ray diffractometer equipped with Cu Kα radiation, and data are collected in the range of 20° to 80°, TEM, energy dispersive X-ray spectroscopy and selected area electron diffraction are performed using a JEM 2100F, the UV-visible absorption spectrum is measured in a TU-1950 dual-beam UV-visible spectrophotometer, and the Raman mapping is determined by a micro-confocal Raman spectrometer equipped with a 532 nm laser.

5. The method of claim 1 for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, characterized in that: The specific operation of step 5 is: the initial range of SERS spectrum is 600-1800cm -1 , cut off to 700-1700cm -1 , background subtraction was performed on different nanoplastic spectra using BWSpec software. The peak intensities showing variations in different nanoplastic spectra were normalized to a standard scale between 0 and 1 using Min-Max to eliminate the impact of intensity differences on the model. Principal component analysis and t-distributed random neighbor embedding methods were used to project high-dimensional data onto a two-dimensional plane. Data visualization of nanoplastic categories scattered in different environmental samples was performed in principal component analysis and t-distributed random neighbor embedding scatter plots.

6. The method of claim 1 for analyzing nanoplastics using AIOF-induced confined localization and machine learning-assisted SERS, characterized in that: The specific operation of step six is ​​to use KNN, GDBT, CNN and Transformer algorithms to classify nanoplastics dispersed in different environmental samples, and reveal the relationship between the SERS characteristic peak intensity and the concentration of nanoplastics in the sample by linear fitting the data. For each concentration, three different measurement points are taken to calculate the average and standard deviation of the intensity. The model fitting is calculated using the coefficient of determination R 2 To evaluate, analyze and predict the exponential data, and draw the fitting curve by taking the logarithm of the concentration and the corresponding characteristic peak intensity; The performance of the machine learning model in distinguishing nanoplastics in different environmental samples was described graphically through the confusion matrix. The prediction performance was evaluated using a series of metrics, including accuracy, recall, precision, and F 1 value: ； ； ； ； in and denote the number of true positive, false positive, true negative, and false negative predictions, respectively; For external validation, the established machine learning model was used to classify nanoplastics in different environmental samples to determine the classification accuracy.

Images