A method for identifying and attributing ultra-short chain per- and polyfluoroalkyl substances

Abstract

The application discloses a kind of ultra-short chain perfluoro and polyfluoro alkyl compound identification and source attribution identification method, first obtain the multidimensional detection data of C1-C3 ultra-short chain PFAS in the sample to be measured, including chromatographic retention time, MS1 accurate mass and isotope pattern, MS2 fragment spectrum, at least two in ion migration collision cross section and delay column correction signal-to-noise ratio;Multidimensional data are input into pre-trained visual semantic embedding deep learning model, and the semantic embedding vector is extracted by the multi-modal feature fusion of fragment sparse adaptive mechanism, when the number of MS2 fragments is less than the preset threshold, the fragment dimension weight is automatically reduced and the retention time and collision cross section dimension weight are simultaneously improved;The semantic embedding vector is matched with source fingerprint database, and the identification confidence level and source attribution classification result are output.The application breaks through the technical bottleneck that ultra-short chain PFAS mass spectrum fragment is difficult to accurately identify under the condition of sparsity, realizes the joint determination of non-target screening and multi-source path attribution.

Description

Technical Field

[0001] This application belongs to the interdisciplinary field of environmental analytical chemistry and artificial intelligence, specifically involving a method for identifying and attributing pollution sources of ultrashort-chain perfluorinated compounds (PFAS, carbon chain length C1~C3) based on multimodal feature fusion deep learning. It is applicable to non-targeted screening, structural identification, and multi-source pathway attribution of ultrashort-chain PFAS in environmental water samples, sediments, and biological samples. Background Technology

[0002] Per- and polyfluoroalkyl substances (PFAS) are a class of synthetic persistent organic pollutants. With the continued tightening of global restrictions on long-chain PFAS (carbon chain length C8 and above), industry is gradually shifting towards short-chain and ultra-short-chain alternatives, leading to a sustained increase in emissions of ultra-short-chain PFAS with carbon chain lengths of C1 to C3. Ultra-short-chain PFAS mainly include trifluoroacetic acid (TFA, C2), perfluoropropionic acid (PFPrA, C3), trifluoromethanesulfonic acid (TFMS, C1), perfluoroethanesulfonic acid (PFEtS, C2), and perfluoropropanesulfonic acid (PFPrS, C3), etc. They have small molecular weights, high polarity, and high water solubility, exhibiting far greater migration and diffusion capabilities in the environment than their traditional long-chain counterparts. Global regulatory agencies urgently need identification methods for ultra-short-chain PFAS.

[0003] Existing PFAS identification technologies for ultrashort chains have the following unresolved technical shortcomings:

[0004] (a) Defects in chromatographic retention techniques

[0005] Traditional C18 reversed-phase chromatography columns exhibit extremely weak retention of C1-C3 ultrashort-chain PFAS. TFA and TFMS show almost no retention on C18 columns, co-eluting with the solvent front and failing to effectively separate from matrix interferences. Existing improvements include: using weak anion exchange columns (WAX columns) to achieve TFA determination, but the ion exchange separation mechanism still has limited selectivity for ultrashort-chain PFAS and is incompatible with the separation conditions of long-chain PFAS, requiring separate analysis; supercritical fluid chromatography can achieve low detection limits, but the equipment is not standard laboratory equipment, limiting its widespread adoption; polar reversed-phase chromatography has shown some improvement, but peak shape and reproducibility remain unstable.

[0006] Furthermore, existing technologies have reported the use of two-dimensional liquid chromatography-mass spectrometry (LC-MS) to cover 43 full-chain PFAS in a single analysis (CN121703327A), as well as methods using delayed column technology to delay system background signals (US20240118254A1), and non-targeted screening methods utilizing ion migration collision cross-sections to assist in false positive filtering. These technologies all represent improvements at the pretreatment / chromatographic separation level, focusing on how to separate ultrashort-chain PFAS from their medium- and long-chain counterparts chromatographically, or how to reduce background interference. They do not address the structural identification problem after separation, nor do they solve the technical challenge of achieving high-confidence identification through multimodal signal fusion under fragment sparse conditions. Therefore, the field lacks an efficient separation scheme with chromatographic conditions that can simultaneously cover C1-C3 ultrashort-chain PFAS and C4 and above medium- and long-chain PFAS in a single analysis, and even more so, lacks intelligent identification methods after separation.

[0007] (ii) Deficiencies in mass spectrometry identification technology

[0008] Mass spectrometry identification of ultrashort-chain PFAS faces a dual challenge: firstly, fragment information is extremely scarce; TFA can only generate 2 fragment ions, and TFMS and PFPrA can only generate 1-3 fragment ions, far fewer than the 3-8 fragment ions of C4 and above PFAS; secondly, TFA, as a common solvent residue and background contaminant in mass spectrometry systems, can still detect false positive signals in samples without TFA. Existing non-targeted screening methods (such as DeePFAS and spectral similarity molecular networks) rely on MS2 fragment abundance for structural annotation, which significantly reduces accuracy under fragment sparse conditions and lacks effective TFA background subtraction and false positive elimination strategies. Therefore, existing non-targeted PFAS identification methods rely on MS2 fragment abundance, fail in fragment sparse scenarios of ultrashort-chain PFAS, and lack TFA false positive elimination mechanisms.

[0009] (III) Deficiencies in Source Attribution Identification Technology

[0010] The pollution sources of ultra-short-chain PFAS are diverse and overlapping, including five typical source pathways: photochemical degradation of refrigerants / halogenated hydrocarbons (HFO-1234yf → atmospheric photolysis → TFA), degradation of fluorinated pesticides (trifluralin → soil degradation → TFA + PFPrA), direct industrial emissions (direct emission of TFMS and PFEtS from semiconductor process wastewater), degradation of shortened fluoropolymer chains (thermal degradation of PTFE / PVDF, etc. → short-chain and ultra-short-chain products), and metabolism of fluorinated drugs. Existing PFAS source identification technologies only target typical emission sources of long-chain PFAS such as aqueous film-forming foam (AFFF), landfills, and wastewater treatment plants, and have not established a multi-source pathway attribution system for ultra-short-chain PFAS. This has resulted in a lack of source fingerprint databases and source attribution classification methods for multi-source pathways of ultra-short-chain PFAS in this field.

[0011] (iv) Technological deficiencies in the intersection of AI and PFAS

[0012] Existing AI-driven PFAS technologies are all designed for C4 and above PFAS, and do not cover ultra-short chain scenarios. For example, the non-targeted identification method (CN118837428B) based on spectral similarity to generate molecular networks proposed by Shanghai Jiao Tong University relies on MS2 fragment similarity clustering, which has extremely low clustering resolution under fragment sparse conditions. The DeePFAS model proposed by a certain university encodes MS2 spectra based on convolution and Transformer architecture to achieve structural annotation, but it only targets complete MS2 spectra of C4 and above PFAS, and the encoder information input is insufficient under fragment sparse conditions. In previous research, the applicant proposed an AI-based watershed PFAS tracing and source determination method and system (CN121905342A). This patent adopts a visual semantic embedding deep learning (VSE-DLM) architecture, which integrates multi-source data of the watershed to achieve source determination of conventional PFAS with an accuracy of 94.8%. However, this fundamental technology is designed for conventional PFAS source tracing at the watershed scale, with input being spatiotemporal concentration data from multiple sources within the watershed. It is not specifically designed for identifying the structure of ultrashort-chain PFAS under fragment sparse conditions in a single sample. When faced with ultrashort-chain PFAS such as TFA and TFMS, which produce only 1-2 fragment ions, the MS2 spectral encoder in this architecture fails to function effectively due to insufficient input information. Furthermore, this approach does not address TFA false positive elimination or the construction of a dedicated source fingerprint database for ultrashort chains. In other words, existing AI-PFAS technologies in this field have not designed specific deep learning architectures and multimodal feature fusion strategies for fragment sparse scenarios of ultrashort-chain PFAS, nor have they achieved joint determination of identification and source attribution. Summary of the Invention

[0013] To address the shortcomings of existing technologies, this application proposes a method for the identification and source attribution of ultrashort-chain perfluorinated and polyfluoroalkyl compounds. It introduces a fragment sparsity adaptive mechanism, triple TFA false positive elimination, and a five-category ultrashort-chain-specific source fingerprint database, thereby extending the application scenario of the visual semantic embedding architecture from conventional PFAS watershed tracing to single-sample identification and source attribution of ultrashort-chain PFAS.

[0014] The technical solution adopted to achieve the above-mentioned objectives of this application is as follows:

[0015] A method for identifying and source attribution of ultrashort-chain perfluorinated and polyfluoroalkyl compounds includes the following steps: S1: acquiring multi-dimensional detection data of ultrashort-chain perfluorinated and polyfluoroalkyl compounds with carbon chain lengths of C1 to C3 in the sample to be tested, wherein the multi-dimensional detection data includes at least two of the following: anion exchange chromatography retention time, MS1 precise mass and isotope mode, MS2 fragment spectrum, ion migration collision cross section, and signal-to-noise ratio corrected for delayed column;

[0016] S2: Input the multi-dimensional detection data into a pre-trained visual semantic embedding deep learning model, and extract the semantic embedding vectors of ultrashort chain perfluorinated and polyfluoroalkyl compounds through multimodal feature fusion under fragment sparse conditions; the fragment sparse condition refers to the condition that the number of MS2 fragments is less than a preset threshold; the multimodal feature fusion includes a fragment sparse adaptive mechanism, that is, when the number of fragments is less than the preset threshold, the weight of the MS2 fragment dimension is automatically reduced and the weight of the chromatographic retention time dimension and the ion migration collision section dimension are simultaneously increased;

[0017] S3: Perform similarity matching between the semantic embedding vector and the semantic embedding vectors of each source type in the source fingerprint database of ultrashort chain perfluorinated and polyfluoroalkyl compounds, and output the identification confidence level and source classification result of ultrashort chain perfluorinated and polyfluoroalkyl compounds.

[0018] Furthermore, the preset threshold is 3 fragments; the fragment sparsity adaptive mechanism independently determines the number of fragments for each detected signal during the inference phase; the multi-dimensional detection data also includes... 13 C / 12 At least one of the C isotope ratio and the MS2 fragment mass difference, wherein the MS2 fragment mass difference is the mass difference of 49.9968 Da corresponding to the CF2 unit.

[0019] In step S1, the signal-to-noise ratio corrected by the delay column is obtained as follows: a delay column is connected in series between the injector and the analytical column of the liquid chromatography-mass spectrometry system, and the corrected signal-to-noise ratio SNR_corrected = (I_delay_sample - I_delay_background) / I_delay_background is calculated. SNR_corrected > 3 is determined to be a real sample signal; where I_delay_sample is the sample signal intensity with the delay column, and I_delay_background is the background signal intensity with the delay column.

[0020] S1 involves pretreatment of the sample to be tested. The pretreatment method includes: when the sample to be tested is a water sample, taking a water sample and adding... 13The C-labeled isotope internal standard mixed solution was adjusted to acidic pH, extracted and eluted by a weak anion exchange solid-phase extraction column, and then concentrated. When the sample to be tested was a sediment or biological sample, it was freeze-dried, ground, and then extracted by ultrasonication, followed by extraction and elution by a weak anion exchange solid-phase extraction column and then concentrated. The chromatographic detection conditions adopted a combination of weak anion exchange column and delay column.

[0021] The visual semantic embedding deep learning model includes: a visual encoder, which contains multiple sub-encoders that process detection data of different dimensions respectively, and the outputs of each sub-encoder are concatenated and then output as a visual feature embedding vector through a fusion layer.

[0022] The semantic encoder receives the source fingerprint semantic description vector and outputs the semantic feature embedding vector.

[0023] The cross-modal embedding space mapping module maps visual feature embedding vectors and semantic feature embedding vectors to a shared embedding space, and achieves visual-semantic alignment through an alignment loss function. The alignment loss function includes a combination of four terms: visual-semantic alignment loss, structural identification loss, source attribution classification loss, and physical guidance constraint loss.

[0024] The fragment sparsity adaptive mechanism adjusts the weights in the feature stitching layer of the visual encoder. When the number of fragments n is less than a preset threshold n0, a weighting coefficient w_MS2 = n / n0 is applied to the output feature vector of the MS2 fragment sub-encoder, an weighting coefficient w_RT = 1+(n0-n) / n0 is applied to the output feature vector of the chromatographic retention time sub-encoder, and an weighting coefficient w_CCS = 1+(n0-n) / n0 is applied to the output feature vector of the ion migration collision section sub-encoder. The weighted feature vectors are then stitched together and input into the fusion layer. The weighting coefficients are deterministic scalars, calculated from the number of fragments using a formula rather than being learnable parameters.

[0025] Furthermore, in the visual encoder, the MS2 fragment sub-encoder includes a convolutional layer. When the fragment sparse adaptive mechanism is triggered, the convolutional layer is bypassed, and the fully connected layer is used directly to process the MS2 fragment information.

[0026] The source attribution classification includes the following five source types:

[0027] a) A photochemical degradation source of refrigerant or halogenated hydrocarbon, characterized by trifluoroacetic acid as the main component and a trifluoroacetic acid to perfluoropropionic acid concentration ratio greater than 100.

[0028] b) A source of degradation of fluorinated pesticides, characterized by the coexistence of trifluoroacetic acid and perfluoropropionic acid with a concentration ratio of 1 to 10;

[0029] c) Direct industrial emission sources, characterized by perfluorosulfonic acid compounds as the main component and a concentration ratio of perfluorosulfonic acid to perfluorocarboxylic acid greater than 1;

[0030] d) A source of degradation for shortened fluoropolymer chains, characterized by a CF2-like distribution of C2-C6 perfluorocarboxylic acid homologues;

[0031] e) Sources of metabolic degradation of fluoride-containing drugs, characterized by trifluoroacetic acid being the main component in urban sewage and its concentration being positively correlated with population density;

[0032] The source fingerprint database of ultrashort-chain perfluorinated and polyfluoroalkyl compounds contains semantic description vectors for each source type. These semantic description vectors include compound composition features, concentration ratio features, matrix association features, seasonality features, co-detection marker features, spatial distribution features, degradation pathway features, and environmental migration behavior features.

[0033] The training process of the visual semantic embedding deep learning model introduces physical guidance constraints, which include:

[0034] The carbon-fluorine bond dissociation energy constraint limits the predicted fragment abundance distribution to no more than the allowable deviation range of the carbon-fluorine bond dissociation energy.

[0035] The water solubility-vapor pressure correlation constraint limits the predicted hydrophobicity of ultrashort-chain perfluorinated and polyfluorinated alkyl compounds to no more than 0.5.

[0036] The CF2 mass difference constraint restricts the distance between homologous objects in the embedded space to be consistent with the mass difference of the CF2 elements.

[0037] The S2 also includes trifluoroacetic acid false positive elimination, comprising the following three mechanisms:

[0038] 1) By using delayed column pre-separation, the system background trifluoroacetic acid signal is separated from the sample trifluoroacetic acid signal in terms of retention time;

[0039] 2) By deducting blank contributions from trifluoroacetic acid through system blank deduction;

[0040] 3) Through 13 C / 12 Cross-validation of C isotope ratios, when the sample contains trifluoroacetic acid 13 When the deviation between the C ratio and the natural abundance is less than the first threshold, it is determined to be a true sample signal. 13 A false positive is defined as a C ratio deviation from the system background trifluoroacetic acid that is less than the second threshold.

[0041] The confidence levels for identification include:

[0042] Level 1: Confirmation and identification, which must meet the following requirements: standard comparison, isotope internal standard quantification, retention time deviation less than 0.1 min, 13C isotope ratio verification, and MS2 fragment matching;

[0043] Level 2: Sparse fragment identification requires meeting the following criteria: multimodal fusion matching of the visual semantic embedding model, retention time deviation of less than 0.3 min, at least one MS2 fragment match, ion migration collision cross-section deviation of less than 2%, and passing false positive elimination verification.

[0044] Level 3: Provisional identification, requiring MS1 exact quality matching deviation to be less than 5 ppm and retention time deviation to be less than 0.5 min; the similarity matching is calculated using cosine similarity, and the confidence level of the source attribution classification result is determined based on the difference between the highest similarity and the second highest similarity.

[0045] Compared with the prior art, this application has the following advantages:

[0046] (1) High-confidence identification capability under fragment sparse conditions: Existing PFAS non-target identification methods rely on MS2 fragment abundance for structural annotation, and the accuracy drops below 60% when the number of fragments is less than 3. This application adopts a fragment sparse adaptive mechanism to automatically reduce the MS2 dimension weight and simultaneously increase the weight of chromatographic retention time and ion migration collision section dimensions when the number of fragments is 1 to 2, thereby achieving multimodal feature fusion compensation and maintaining the identification accuracy above 85%.

[0047] (2) Effective elimination of TFA false positives: The false positive detection rate of TFA in existing methods can reach 30%~50%. This application achieves this by delaying column separation, system blank subtraction, and... 13 C / 12 The combined effect of the three mechanisms of C isotope ratio cross-validation reduced the false positive probability of TFA to about 0.15%, significantly improving the reliability of the identification results.

[0048] (3) First realization of multi-source path attribution of ultra-short chain PFAS: Existing PFAS source identification technologies only target long chain PFAS emission sources such as AFFF, landfills, and sewage treatment plants. This application is the first to construct a classification system covering five types of source fingerprints, including refrigerant photolysis, degradation of fluorinated pesticides, industrial emissions, degradation of fluoropolymers, and metabolism of fluorinated drugs, filling the technological gap in the field of ultra-short chain PFAS source attribution identification.

[0049] (4) Single analysis covers the entire chain of PFAS: Existing methods require different chromatographic conditions to analyze ultrashort chain and medium-long chain PFAS separately. This application uses a combination of weak anion exchange column and delay column, which can cover C1~C14+ PFAS in a single 15-minute analysis, improving the analysis efficiency by 2 to 3 times.

[0050] (5) Physical constraints enhance model reliability: Traditional pure data-driven deep learning models are prone to overfitting when data is scarce. This application introduces three physical guidance constraints: carbon-fluorine bond dissociation energy, water solubility-vapor pressure correlation, and CF2 mass difference, so that the model prediction results always conform to the basic physicochemical laws of PFAS, ensuring the scientific rationality of the identification results. Attached Figure Description

[0051] Figure 1 A flowchart of the method for identifying and attributing ultrashort-chain perfluorinated and polyfluorinated alkyl compounds provided in this application;

[0052] Figure 2 This is a schematic diagram of the VSE-DLM model architecture provided in the embodiments of this application;

[0053] Figure 3 This is a schematic diagram illustrating the effect of PGNN physical guidance constraints in the embedding space in the embodiments of this application;

[0054] Figure 4 This is a flowchart illustrating the triple TFA false positive elimination mechanism in the embodiments of this application;

[0055] Figure 5 This is a schematic diagram of the ultrashort chain PFAS source fingerprint classification system in the embodiments of this application;

[0056] Figure 6 This is a logic diagram of the three-level confidence level determination system in the embodiments of this application;

[0057] Figure 7 This is a diagram showing the TFA source attribution results for seven sections of a typical river flowing into the sea in the Yangtze River Delta region, as described in Example 1 of this application. Detailed Implementation

[0058] The present application will now be described in detail with reference to specific embodiments.

[0059] The overall process of the identification and source attribution method for ultrashort-chain perfluorinated and polyfluoroalkyl compounds provided in this application is as follows: Figure 1 As shown, it includes the following steps:

[0060] S1 Data Acquisition Stage: Acquire multi-dimensional detection data of ultrashort chain PFAS of the sample to be tested;

[0061] S1.1 Sample Pretreatment

[0062] The samples to be tested can be environmental water samples (surface water, groundwater, drinking water, rainwater, industrial wastewater), sediments, or biological samples. The pretreatment steps are as follows:

[0063] (a) Water sample pretreatment

[0064] 1. Take a 500 mL water sample and add...13 C-labeled isotope internal standard mixed solution (containing C-labeled isotope internal standard mixed solution) 13 C2-TFA, 13 C4-PFBA, 13 C4-PFOA, 13 (C4-PFOS, etc.), internal standard concentration 10 ng / L.

[0065] 2. Adjust the pH of the water sample to 4.0±0.2 (using acetate buffer) to allow PFAS to exist in molecular form, thereby enhancing the adsorption efficiency of solid-phase extraction.

[0066] 3. Extraction was performed using a weak anion exchange solid-phase extraction column (WAX SPE cartridge, 6 mL / 150 mg):

[0067] Activation: 4 mL 0.1% ammonia water + methanol → 4 mL methanol → 4 mL pure water

[0068] Sample loading: Pass 500 mL of water sample through the WAX ​​column at a flow rate of 5–8 mL / min.

[0069] Washing: 4 mL 25 mM acetate buffer (pH 4) → 3 mL pure water, discard the effluent.

[0070] Elution: Collect the eluent using 4 mL of 0.1% ammonia-methanol solution.

[0071] 4. Concentrate the eluent to 0.5 mL by nitrogen blowing, add 0.5 mL of 2 mM ammonium acetate aqueous solution to make a final volume of 1 mL, and vortex to mix.

[0072] (b) Delay Column System Configuration

[0073] A delay column is installed in series between the injector and the column in the UPLC-MS / MS system to separate the system background signal of TFA from the true signal of the sample.

[0074] 1. Selection of delay column: Use a short column with the same packing material as the analytical column (such as WAX short column, 30 mm × 2.1 mm, 3 μm), or a dedicated PFAS delay column (such as Waters ACQUITY UPLC BEH C18 Delay Column, 50 mm × 2.1 mm, 1.7 μm).

[0075] 2. Delay column installation position: Installed between the mixer and the injector (traditional position), so that the system background TFA signal is retained on the delay column and arrives at the detector later than the sample TFA signal.

[0076] 3. By comparing the TFA elution times under conditions of no delay column and with delay column, the retention time offset Δt_delay of the system background TFA is confirmed and used for subsequent false positive elimination.

[0077] (c) Sediment / biological sample pretreatment

[0078] Sediment or biological samples need to be freeze-dried, ground, and weighed. 1 g of the dry sample is added to 10 mL of 0.1% ammonia and methanol and ultrasonically extracted for 30 min (repeated 3 times). The extracts are combined, and the subsequent WAX SPE steps are the same as for water samples.

[0079] S1.2 Instrument Testing Conditions

[0080] An ultra-high performance liquid chromatography-tandem high resolution mass spectrometry (UPLC-HRMS / MS) system was used, with the following specific conditions:

[0081] (a) Chromatographic conditions

[0082] 1. Analytical column: Weak anion exchange column (such as Waters ACQUITY UPLC BEH WAX Column, 100 mm × 2.1 mm, 1.7 μm), which has sufficient retention capacity for C1~C3 ultrashort chain PFAS and is compatible with the separation of C4+ PFAS.

[0083] 2. Mobile phases: Phase A: 2 mM ammonium acetate aqueous solution (pH≈6.8); Phase B: 2 mM ammonium acetate methanol solution

[0084] 3. Gradient elution program, as shown in Table 1, with flow rate: 0.3 mL / min; column temperature: 40°C; injection volume: 10 μL.

[0085] Table 1 Gradient elution program

[0086]

[0087] (b) Mass spectrometry conditions

[0088] 1. Mass spectrometry type: High-resolution mass spectrometry (QTOF or Orbitrap), resolution ≥30,000 FWHM.

[0089] 2. Ionization Mode: Electrospray Negative Ion Mode (ESI) - ).

[0090] 3. Scanning Mode:

[0091] Full Scan (MS1): m / z 50~1000, used to obtain accurate molecular weight and isotope patterns.

[0092] ddMS2 (Data depends on secondary scan): Top 5, collision energy 10 / 20 / 40 eV stepped, used to obtain MS2 fragment information.

[0093] AIF (All Ion Fragmentation Scan): m / z 50~500, collision energy 20 eV, used to supplement low-mass fragment information of ultrashort chain PFAS.

[0094] 4. Ion migration conditions (e.g., mass spectrometer equipped with IMS module): IMS drift tube nitrogen flow rate: 90 mL / min; IMS wave height: 40 V; obtain CCS (collision cross section value) for multimodal feature supplementation.

[0095] S1.3 Multi-dimensional detection data acquisition

[0096] For each ultrashort chain PFAS candidate signal, the following multimodal detection data were extracted:

[0097] Dimension 1—MS1 Precise Mass and Isotope Mode

[0098] (1) Precise molecular weight: such as TFA [MH] - =m / z112.9850 (mass deviation ≤5 ppm);

[0099] (2) Isotope pattern: 13 C1-TFA [MH] - =m / z 113.9884; 13 C2-TFA [MH] - =m / z114.9917;

[0100] (3) Calculation 13 C / 12 C isotope ratio R_iso=I ( 13 C1) / I ( 12 C), used to distinguish sample TFA from system background TFA (system background TFA). 13 The C ratio deviates from the natural abundance by 1.1%.

[0101] Dimension 2—Retention Time in Anion Exchange Chromatography

[0102] (1) Record the retention time t_R of each candidate signal on the WAX ​​column;

[0103] (2) Calculate the deviation between the retention time and the reference value in the literature: Δt_R = |t_R - t_R_ref|;

[0104] (3) Retention time characteristics serve as a key supplementary dimension when MS2 fragmentation is insufficient: different ultra-short chain PFAS (TFA vs TFMA vs PFPrA) have distinguishable retention time differences (approximately 0.3~1.5 min) on the WAX ​​column.

[0105] Dimension 3 – MS2 fragment spectrum, as shown in Table 2:

[0106] Table 2 MS2 fragment spectrum

[0107]

[0108] In Table 2, candidate signals with insufficient fragments (only 1~2 fragments) are marked as "fragment sparse" and enter the multimodal fusion processing stage S2.

[0109] Dimension 4 – Collision Cross Section (CCS) for Ion Migration, as shown in Table 3:

[0110] Table 3 Collision cross sections for ion migration

[0111]

[0112] Dimension 5 – Delayed Column Corrected TFA Signal-to-Noise Ratio

[0113] Calculate using the following formula: SNR_corrected = (I_delay_sample - I_delay_background) / I_delay_background;

[0114] When the calculated SNR_corrected > 3, it is determined to be a true TFA sample; otherwise, when SNR_corrected ≤ 3, it is determined to be a false positive in the system background.

[0115] S1.4 Data Preprocessing

[0116] Standardization preprocessing is performed on the extracted multi-dimensional raw data:

[0117] 1. MS1 Precise Mass: Converted to mass deviation in ppm, normalized to [0, 1];

[0118] 2. Retention time: Converted to a relative retention time index (with PFBA retention time as a reference), normalized to [0, 1];

[0119] 3. MS2 fragment spectrum: converted into fragment mass deviation vector and relative intensity vector;

[0120] 4. CCS value: Normalized to [0, 1] (with PFOS CCS as the upper limit).

[0121] 5. Isotope ratio: converted to 13 C / 12 C ratio deviation (deviation from natural abundance of 1.1%), normalized to [0, 1];

[0122] 6. SNR_corrected: Normalized to [0, 1];

[0123] The data from each dimension are combined into a multimodal feature vector F = {f_MS1, f_RT, f_MS2, f_CCS, f_iso, f_SNR}, with a total of 6 to 12 dimensions.

[0124] S2 Feature Extraction and Embedding Stage: The multi-dimensional detection data is input into the pre-trained Visual Semantic Embedding Deep Learning Model (VSE-DLM), and the semantic embedding vector of the ultra-short chain PFAS is extracted through multi-modal feature fusion under fragmented and sparse conditions.

[0125] S2.1 VSE-DLM Model Architecture

[0126] The Visual Semantic Embedding Deep Learning Model (VSE-DLM) used in this application comprises three core modules. The VSE-DLM model architecture is as follows: Figure 2 As shown.

[0127] S2.1.1 Visual Encoder (VE)

[0128] The visual encoder is responsible for extracting "visual" features (i.e., instrument detection features of chromatography-mass spectrometry-ion migration) from multimodal detection data. Its specific structure is as follows:

[0129] 1. Input layer: Receives 6~12-dimensional multimodal feature vectors F.

[0130] 2. Dimensity Sub-encoder:

[0131] oMS1 / isotope encoder: 1 fully connected layer (FC, 128-dimensional) → ReLU activation → BatchNorm → output h_MS1;

[0132] Retain-time sub-encoder: 1 fully connected layer (FC, 64-dimensional) → ReLU → output h_RT;

[0133] MS2 fragmented encoder: 1 convolutional layer (Conv1D, kernel=3, 64 filters) → ReLU → 1 FC layer (128 dimensions) → output h_MS2;

[0134] CCS sub-encoder: 1 fully connected layer (FC, 64-dimensional) → ReLU → output h_CCS;

[0135] TFA-SNR sub-encoder: 1 fully connected layer (FC, 64-dimensional) → ReLU → output h_SNR;

[0136] 3. Cross-dimensional fusion layer: Concatenate the outputs of each sub-encoder → Concat(h_MS1, h_RT, h_MS2, h_CCS, h_SNR) → 1-layer FC (256 dimensions) → ReLU → Dropout (0.3) → Output visual feature embedding vector v∈R 256 .

[0137] In this embodiment, a fragmentation sparsity adaptive mechanism is adopted: when the number of MS2 fragments is less than 3 (fragmentation sparse state), the adaptive mechanism is automatically triggered.

[0138] Bypass the Conv1D layer of the MS2 sub-encoder and directly use the FC layer to process only limited fragment information;

[0139] Reduce the output weights of the MS2 sub-encoder (weight_factor=n_fragments / 3);

[0140] Simultaneously increase the weights of the RT and CCS sub-encoders (weight_factor_RT=1+(3-n) / 3; weight_factor_CCS=1+(3-n) / 3);

[0141] The fewer the fragments, the greater the contribution weight of chromatographic retention and CCS dimension, thus achieving adaptive feature fusion of "fragment sparsity → non-fragment dimension compensation";

[0142] (i) Definition and judgment logic of the fragment quantity threshold n0=3

[0143] The selection of the fragment quantity threshold n0=3 is based on the following three criteria:

[0144] 1. Statistical basis for mass spectrometry fragmentation: In the non-targeted screening dataset of C4~C14 PFAS (≥500 compounds), the distribution of MS2 fragment number was statistically analyzed: the average fragment number for C4~C6 PFAS was 5.2±1.8 (n≥3, probability >95%), and the average fragment number for C1~C3 PFAS was 1.8±0.7 (n<3, probability >90%). The threshold n0=3 is precisely at the statistical dividing point between long-chain and ultra-short-chain fragment numbers.

[0145] 2. Basis for reliability of spectral matching: Traditional non-targeted identification methods for PFAS (such as DeePFAS and spectral similarity molecular networks) rely on MS2 fragment abundance spectra for structure annotation. Experiments show that when the number of fragments ≥ 3, the accuracy of Cosine similarity matching based on MS2 spectra is ≥ 92%; when the number of fragments < 3, the accuracy of Cosine similarity matching drops sharply to < 60% (because the spectral information is insufficient to distinguish isomers from noise fragments). The threshold n0=3 corresponds to the turning point of the spectral matching reliability from "reliable" to "unreliable".

[0146] 3. Basis for necessity of weight compensation: When the number of fragments ≥ 3, the feature vector output by the MS2 sub-encoder is h_MS2∈R 64 which already contains sufficient structure distinguishing information (carbon chain length, functional groups, neutral loss patterns, etc.), and no compensation is required from non-fragment dimensions; when the number of fragments < 3, h_MS2 only contains information of 1 to 2 fragment ions (TFA typically only contains CF3COO - one fragment, m / z 112.99), the structure distinguishing information is severely insufficient, and information compensation must be performed from dimensions such as RT (chromatographic behavior reflecting carbon chain length and functional groups) and CCS (reflecting the three-dimensional configuration of molecules).

[0147] Judgment logic: In the inference stage, for each detected signal, the model first extracts the number of MS2 fragment ions n_fragments from the ddMS 2 data (defined as the count of fragment ions with a signal-to-noise ratio SNR > 3 and a mass deviation < 5 ppm in the MS2 spectrum). If n_fragments < n0=3, it is judged as "fragment sparse state", and adaptive weight adjustment is triggered; if n_fragments ≥ n0=3, it is judged as "fragment sufficient state", and the weight of each dimension remains the default value of 1.0. This judgment is executed independently during the inference process of each detected signal, and different compounds may be in different fragment states (for example, TFA is in a fragment sparse state and PFPrA is in a fragment sufficient state in the same sample).

[0148] (ii) Specific implementation of weight adjustment at the feature concatenation layer

[0149] In this application, the fragment sparsity adaptive weight adjustment is implemented at the feature concatenation layer (Concat layer), not at the loss function layer. The specific implementation is as follows:

[0150] The 5 sub-encoders of the visual encoder respectively output feature vectors: h_MS1∈R 64 , h_RT∈R 64 , h_MS2∈R 64 , h_CCS∈R 64 , h_SNR∈R32 Before the fragmented sparse adaptive mechanism is activated, these 5 vectors are directly concatenated: default concatenation: v = Concat(h_MS1, h_RT, h_MS2, h_CCS, h_SNR) → v_raw∈R 288 →FC(256)→v∈R 256 ;

[0151] When the fragmented sparsity adaptive mechanism is activated, a dimension-level weighting coefficient w_d (d is the dimension index) is applied to the output vector of each sub-encoder before the Concat operation. The weighting coefficient is a scalar and acts on the entire sub-encoder output vector: Adaptive concatenation: v = Concat(w_MS1•h_MS1, w_RT•h_RT, w_MS2•h_MS2, w_CCS•h_CCS, w_SNR•h_SNR) → v_weighted∈R 288 →FC(256)→v∈R 256 .

[0152] The weighting coefficients are determined according to the rules in Table 4:

[0153] Table 4 Weighting Coefficient Rules

[0154]

[0155] Key design choice – the reason for applying weights to the Concat layer rather than the loss function layer:

[0156] Gradient flow consistency: If weight adjustment is only implemented in the loss function layer (such as applying penalty weights to the MS2 related classification probabilities in L_ident), the gradient signal is modulated by the weights during backpropagation, causing the gradient direction of the MS2 sub-encoder on fragmented sparse samples to be inconsistent with that on fragmented sufficient samples, which may destroy the pre-trained general PFAS feature embedding capability. However, when the weights are applied to the Concat layer (feature vector multiplied by scalar coefficients), the backpropagation gradients are naturally mixed when passing through the FC(256) fusion layer, and the gradient directions received by each sub-encoder are consistent, with only the amplitude being scaled proportionally without changing the gradient direction.

[0157] Consistency during inference: Both the training and inference phases are adaptively adjusted using the same Concat layer weight coefficients, eliminating the need for additional loss function modulation logic during inference and achieving end-to-end consistency.

[0158] (iii) Specific implementation differences between the training phase and the inference phase

[0159] Training phase:

[0160] During the transfer learning pre-training phase (C4~C14 PFAS dataset), all samples had a fragment count ≥3, weight coefficients were set to the default value of 1.0, and the fragment sparsity adaptive mechanism was not activated. The model learned the baseline ability of "multimodal feature fusion under sufficient fragmentation conditions".

[0161] In the fragmented sparse fine-tuning stage (C1~C3 PFAS dataset), the number of fragments for some samples is <3, and the weight coefficients are dynamically determined according to the above rules. At this time, the model continues to fine-tune the FC(256) fusion layer and CMEM mapping layer under the "fragmented sparse weight configuration", learning how to complete identification and source attribution in the feature space where MS2 information is deweighted and RT / CCS information is upweighted.

[0162] During the training phase, each batch may contain both fully fragmented samples (both with a weight of 1.0) and sparsely fragmented samples (MS2 downweighted, RT / CCS upweighted). The FC(256) fusion layer needs to process the feature vectors of the two weight configurations in the same batch. The influence of the difference in weight coefficients on the distribution of activation values ​​is eliminated by normalization through the BatchNorm layer.

[0163] During the training phase, the weight coefficients are deterministic scalars (calculated by the formula from n_fragments) and are not used as learnable parameters in gradient optimization. This is to ensure the interpretability and reproducibility of the weight adjustment rules and to prevent the model from bypassing the fragmentation sparsity compensation mechanism by learning the weight coefficients.

[0164] Reasoning stage:

[0165] For each detected signal, the model starts from ddMS. 2 Extract the number of MS2 fragments n_fragments from the data, and determine the fragment status according to the judgment logic.

[0166] If the fragmented and sparse state is (n<3), w_d is calculated according to the weighting rules above. Weighting coefficients are applied to the output vectors of each sub-encoder in the Concat layer, and the weighted visual feature embedding vector v∈R is output after passing through the FC(256) fusion layer. 256 .

[0167] If the fragmentation is sufficient (n≥3), the weight coefficients are all 1.0, and the output of the Concat layer is completely consistent with the default mode.

[0168] No additional threshold determination module or post-processing logic is required during the inference phase. The fragmented sparse adaptive mechanism is fully embedded in the feature fusion forward propagation path of the model, achieving adaptive adjustment with "zero additional inference overhead".

[0169] (iv) Specific processing of backpropagation gradient flow

[0170] During the fragmented sparse fine-tuning stage, when the weight coefficient w_d is applied to the Concat layer, the gradient is backpropagated. That is, the gradients received by each sub-encoder are scaled by weight coefficients:

[0171] MS2 sub-encoder: (When the fragmentation is sparse, the gradient magnitude is scaled by a weighting factor of n / 3. For example, when TFA n=1, the gradient magnitude is only 1 / 3 of the default value.)

[0172] RT sub-encoder: (When the fragmentation is sparse, the gradient magnitude is scaled by the weighting factor, such as when TFA n=1, the gradient magnitude is 5 / 3 of the default value).

[0173] CCS sub-encoder: (Same as RT sub-encoder);

[0174] The effect of this gradient scaling is that during the training process of fragmented sparse samples, the parameter update magnitude of the MS2 sub-encoder is automatically reduced, while the parameter update magnitude of the RT / CCS sub-encoder is automatically increased, which is consistent with the design intention of "fragmented sparsity → non-fragmented dimension compensation" at the parameter optimization level.

[0175] In this embodiment, since the weights of the MS1 / RT / CCS sub-encoders have been frozen during the pre-training phase (only the MS2 sub-encoder and CMEM module are fine-tuned), the gradient scaling actually affects: (1) the parameter update magnitude of the MS2 sub-encoder is scaled by the weight reduction coefficient on the fragmented sparse samples; (2) the parameter update of the FC(256) fusion layer is consistent for all samples; (3) the parameter update of the CMEM mapping layer is consistent for all samples.

[0176] S2.1.2 Semantic Encoder (SE)

[0177] The semantic encoder is responsible for extracting "semantic" features from the semantic information of the ultra-short chain PFAS source fingerprint:

[0178] Input layer: Receives the source fingerprint semantic description vector S (8-dimensional, see the source fingerprint database definition described above).

[0179] Two fully connected layers: FC (128) → ReLU → BatchNorm → FC (256) → ReLU → Dropout (0.2);

[0180] Output semantic feature embedding vector s∈R 256 .

[0181] S2.1.3 Cross-Modal Embedding Mapping (CMEM) Module

[0182] The cross-modal embedding space mapping module maps visual embedding vector v and semantic embedding vector s to a shared embedding space, achieving visual-semantic alignment and matching:

[0183] 1. Shared embedded space dimension: d=128;

[0184] 2. Visual side mapping: v→FC(128)→L2 normalization→v'∈R 128 v' is the normalized visual embedding vector;

[0185] 3. Semantic side mapping: s→FC(128)→L2 normalization→s'∈R 128 s' is the normalized semantic embedding vector;

[0186] 4. The alignment loss function is as follows:

[0187] L_total = α1 • L_align + α2 • L_ident + α3 • L_source + α4 • L_phys, where α1 = 0.4, α2 = 0.3, α3 = 0.2, α4 = 0.1 (adjustable), L_align is the visual-semantic alignment loss, L_ident is the structural identification loss, L_source is the source attribution classification loss, and L_phys is the physical guidance constraint loss.

[0188] Visual-semantic alignment loss L_align

[0189] L_align=Σ i max(0, margin-cos(v') i ,s' i + ) + cos(v' i ,s' i - ), where s' i + For the correct source fingerprint semantic embedding to match, s' i - For non-matching source fingerprint semantic embedding, margin=0.2.

[0190] Structural identification loss L_ident

[0191] The shared embedding vector v' output by the CMEM module i ∈R 128 Based on this, the connection structure identification and classification head is established:

[0192] z_ident, i = W_ident · v' i +b_ident (W_ident ∈ R^(J×128), b_ident ∈R^J);

[0193] y'_ident, i = softmax(z_ident, i) (y'_ident, i ∈ R^J); where J = 6 is the number of compound categories: TFA, PFPrA, TFMS, PFEtS, PFPrS, and non-PFAS interfering compounds.

[0194] The structural identification loss uses multi-class cross-entropy:

[0195] , where N is the number of samples in the batch; y_ident,ij ∈ {0,1} is the true unique thermal label of the i-th sample on the j-th compound; y'_ident,ij is the j-th predicted probability output by the structure identification classification head.

[0196] Source attribution classification loss L_source

[0197] Similarly, in v' i Based on the source attribution classification header:

[0198] z_source, i = W_source · v' i + b_source (W_source ∈ R^(K×128), b_source ∈ R^K);

[0199] y'_source, i = softmax(z_source, i) (y'_source, i ∈ R^K), where K = 5 is the number of source types: S1 refrigerant photolysis source, S2 fluorinated pesticide degradation source, S3 direct industrial emission source, S4 fluoropolymer degradation source, S5 fluorinated drug metabolism source.

[0200] The source attribution classification loss uses multi-class cross-entropy:

[0201] , where y_source,ik ∈ {0,1} is the true unique hot label of the i-th sample on the k-th source; y'_source,ik is the predicted probability of the k-th class output by the source attribution classification head.

[0202] S2.2 Transfer Learning Training Strategy

[0203] Due to the limited size of the ultra-short chain PFAS (C1~C3) dataset, this application adopts a transfer learning strategy:

[0204] S2.2.1. Pre-training stage: Use the complete C4~C14 PFAS dataset (including rich MS2 fragments) to pre-train the visual encoder and semantic encoder of VSE-DLM to establish a general PFAS feature embedding capability.

[0205] Pre-training data: Combined data of ≥500 PFAS compounds × ≥20 emission source types;

[0206] Pre-training objective: L_align + L_ident.

[0207] S2.2.2. Fragmentation and Sparsity Fine-tuning Stage: Freeze the weights of the MS1 / RT / CCS sub-encoders of the visual encoder, and fine-tune only the MS2 sub-encoder and the cross-modal mapping module to adapt to the fragmentation and sparsity scenario of the ultra-short chain PFAS.

[0208] Fine-tuning data: Limited datasets of C1~C3 PFAS (≥50 samples × ≥10 emission source types);

[0209] Fine-tuning target: L_total;

[0210] S2.2.3. Fragmented Sparse Adaptive Activation: After fine-tuning, the fragmented sparse adaptive mechanism is automatically activated without additional training.

[0211] S2.3 Physically Guided Constraints (PGNN Mechanism)

[0212] This step introduces Physically Guided Neural Network (PGNN) constraints during model training. The effect of PGNN physics-guided constraints in the embedding space is as follows: Figure 3 As shown.

[0213] Figure 3 In the middle, constraint 1, the carbon-fluorine bond dissociation energy constraint, is as follows:

[0214] L_CF_bond=Σ i max(0, |E_predicted(i) - E_reference(CF)| - tolerance), where E_reference(CF) ≈ 485 kJ / mol and tolerance = 20 kJ / mol. This constraint ensures that the model does not predict fragmentation modes that violate the stability of carbon-fluorine bonds.

[0215] Constraint 2: The water solubility-vapor pressure correlation constraint is as follows:

[0216] L_water_sol=Σ i max(0, log_Kow_predicted(i) - log_Kow_threshold), where log_Kow_threshold=0.5, this constraint ensures that the model does not attribute ultrashort chain PFAS to emission sources characterized by hydrophobic migration.

[0217] Constraint 3, CF2 quality difference constraint, is as follows:

[0218] L_CF2_spacing=Σ max(0,|Δm_embedding(i,j)-49.9968|-0.001), where Δm(CF2)=49.9968 Da, this constraint ensures that the distance between PFCA and PFSA homologs in the embedding space is consistent with the mass difference of CF2.

[0219] S2.4 TFA False Positive Elimination Mechanism

[0220] This application incorporates a triple TFA false positive elimination mechanism in the S2 stage, such as Figure 4 As shown:

[0221] S2.4.1 First Stage – Delayed Column Separation

[0222] By using a delayed column pre-separation, the retention time delay Δt_delay of the system background TFA (approximately 2-5 min) separates it from the sample TFA signal. Only the signal with t_R within the sample TFA window (t_R_ref ± 0.1 min) is retained for subsequent processing.

[0223] S2.4.2 Second Step – System Blank Deduction

[0224] For every 10 samples, one system blank (pure water after a complete pretreatment process) is inserted, and the blank TFA signal intensity I_blank is recorded. The sample TFA signal is reduced by the blank contribution: I_corrected = I_sample - I_blank. If I_corrected < LOD, it is considered a false positive and is discarded.

[0225] S2.4.3 Third layer – Cross-validation of isotope ratios

[0226] Calculate the TFA of the sample 13 C / 12 C is the ratio R_iso_sample. When |R_iso_sample-1.1%| < 0.3%, it is judged as a true TFA sample; when |R_iso_sample-R_iso_background| < 0.1%, it is judged as a false positive.

[0227] P_false_positive = P(delay_fail) × P(blank_fail) × P(isotope_fail) ≈ 0.05 × 0.10 × 0.03 = 0.0015, meaning the false positive probability drops to approximately 0.15%.

[0228] S3 Source Attribution Determination Stage: The semantic embedding vector is matched with the ultra-short chain PFAS source fingerprint database, and the identification confidence level and source attribution classification result of the ultra-short chain PFAS are output.

[0229] S3.1 Ultra-short PFAS Source Fingerprint Database

[0230] The ultra-short chain PFAS source fingerprint database constructed in this application contains the following:

[0231] (a) The source classification system is shown in Table 5, and it is divided into five categories:

[0232] Table 5 Source Classification System

[0233]

[0234] (b) The source fingerprint semantic description vector is shown in Table 6, which has 8 dimensions:

[0235] Table 6 Source Fingerprint Semantic Description Vector Table

[0236]

[0237] (c) Database construction method

[0238] 1. Literature Data Mining: Collect PFAS detection data of published ultrashort chains worldwide (≥200 articles) and classify and label them according to source type.

[0239] 2. Accumulation of experimental data: Samples were collected from known emission sources to obtain multi-dimensional detection data.

[0240] 3. Continuous database updates: New data is added to the database every quarter, and the source fingerprint semantic encoder is retrained.

[0241] The ultra-short chain PFAS source fingerprint classification system constructed in this step is illustrated as follows: Figure 5 As shown, there are a total of five source categories and eight dimensions.

[0242] S3.2 Source Attribution Determination Process

[0243] S3.2.1 Perform cosine similarity calculation between the visual embedding vector v' output from stage S2 and the semantic embedding vectors s'1~s'5 of each source type in the source fingerprint database:

[0244] sim(v', s') k =cos(v', s') k ) = (v' • s' k ) / (||v'||•||s' k ||).

[0245] S3.2.2. Select the source type with the highest similarity as the attribution result:

[0246] source_predicted=argmax k sim(v', s') k ).

[0247] S3.2.3. Calculate the attribution confidence score:

[0248] confidence_source=max k sim(v', s') k -second_max k sim(v', s') k The confidence level is determined as follows: confidence_source > 0.3 is considered high confidence; 0.1~0.3 is considered medium confidence; and < 0.1 is considered low confidence (requires manual verification).

[0249] S3.3 Confidence level classification, as shown in Table 7:

[0250] Table 7 Confidence Level Classification Table

[0251]

[0252] In this step, the logic for determining the three-level confidence level classification system is as follows: Figure 6 As shown, the automatic determination of confidence level is implemented by the model output layer: when p_ident > 0.95 and meets the Level 1 condition → Level 1, when p_ident > 0.80 and meets the Level 2 condition → Level 2, when p_ident > 0.60 and meets the Level 3 condition → Level 3, when p_ident < 0.60 → not identified (excluded).

[0253] The technical solution of this application will be explained in detail below with reference to specific implementation cases.

[0254] Example 1: Identification and source attribution of ultrashort chain PFAS in surface water of typical rivers flowing into the sea in the Yangtze River Delta

[0255] Sample Information: Surface water samples from seven sections of typical rivers flowing into the sea in the Yangtze River Delta, from upstream to downstream: upstream mountainous section, upstream reservoir inlet section, midstream densely farmland section, midstream and downstream mixed agricultural and light industrial section, downstream mixed urban and industrial section, downstream core urban area section, and estuary section. Sampling time: June 2026 (high water season). 500 mL water samples were collected from each section, and three parallel samples were collected.

[0256] Pretreatment: Following the S1 stage water sample pretreatment procedure, 500 mL of water sample was extracted with WAX SPE, and the delay column system was configured with a short WAX column (30 mm × 2.1 mm). A delay column blank control was also configured for each cross section.

[0257] Instrument analysis: UPLC-HRMS / MS (Waters ACQUITY UPLC + Xevo G2-XS QTOF), WAX analytical column (100 mm × 2.1 mm), gradient elution for 15 min, ESI - Mode, Full Scan + ddMS 2 +AIF. IMS module turned on (nitrogen gas in drift tube 90 mL / min).

[0258] Multi-dimensional data acquisition—TFA detection data is shown in Table 8:

[0259] Table 8. Multi-dimensional Data Acquisition – TFA Detection Data Table

[0260] Multi-dimensional data acquisition—Table 9 shows the PFAS detection data of other ultra-short chains:

[0261] Table 9. Multi-dimensional Data Acquisition – PFAS Detection Data for Other Ultra-short Chains

[0262]

[0263] VSE-DLM processing – Fragmented sparse adaptive activation:

[0264] The MS2 fragment count of the TFA detected signals from all 7 sections was 2 (<3), triggering the fragment sparsity adaptive mechanism. The weights were adjusted as follows: MS2 weight = 2 / 3 = 0.67; RT weight = 1 + 1 / 3 = 1.33; CCS weight = 1 + 1 / 3 = 1.33. The visual encoder output v∈R for each section... 256 After CMEM cross-modal mapping, v'∈R 128 .

[0265] VSE-DLM processing – Source attribution matching results are shown in Table 10:

[0266] Table 10 VSE-DLM Processing – Source Attribution Matching Results

[0267]

[0268] VSE-DLM processing – Confidence level determination is shown in Table 11:

[0269] Table 11 VSE-DLM Processing – Confidence Determination Table

[0270]

[0271] In Table 11, the TFA of all 7 sections met the Level 2 criteria for suspected fragment sparse identification: p_ident > 0.80 + fragment number 2 + RT deviation < 0.3 min + 13 C-ratio verification passed + SNR_corrected > 3 + triple false positive elimination passed.

[0272] The conclusions and analysis are as follows:

[0273] In this embodiment, TFA (all with Level 2 fragment sparse suspected identification confidence) was successfully identified in the surface water of 7 sections of typical rivers flowing into the sea in the Yangtze River Delta. The identification accuracy was maintained at ≥85% even with only 2 fragments, thanks to the fragment sparse adaptive mechanism.

[0274] The TFA source attribution results for seven sections of typical rivers flowing into the sea in the Yangtze River Delta are as follows: Figure 7 As shown, Figure 7 The source attribution results show a clear spatial differentiation pattern: the upstream mountain section and the upstream reservoir inlet section are both S1 refrigerant photolysis sources (TFA / PFPrA are 202 and 100, respectively, and HFO-1234yf residues were detected); the midstream densely farmland section and the mid-to-lower reaches mixed agricultural-light industrial zone section are both S2 fluorinated pesticide degradation sources (TFA / PFPrA are 2.4 and 3.0, respectively, and pesticide markers such as trifluoromethylaniline and trifluralin degradation products were detected); the downstream mixed urban and industrial zone section and the downstream core urban area section are both S5 fluorinated drug metabolism sources (TFA / PFPrA are 100 and 141, respectively, and fluorinated drug metabolites such as fluconazole were detected); the estuary section is an S1 refrigerant photolysis source (TFA / PFPrA=200, and HFO-1234yf residues were detected), which may be influenced by the combined effects of tidal mixing in the estuary and the transport of S1 sources from the upstream.

[0275] The TFA concentration in the downstream urban and industrial mixed zone reached as high as 780 ng / L, but the model still attributed it to the S5 drug metabolism source rather than the S3 industrial emission source, indicating that population density and drug use were the dominant contributors to TFA during this monitoring period. The S3 industrial emission source did not become the single dominant source in this 7-section monitoring, which may be related to the fact that the sampling section was not directly located at the emission outlet of the industrial park.

[0276] Appendix Figure 7 The confidence levels of the annotations are consistent with the model output: the confidence level of the upstream mountain section is medium (0.13), and the confidence level of the downstream core urban area section is high (0.27); the confidence levels of the upstream reservoir inlet section, the midstream dense farmland section, the mid-to-lower reaches mixed agricultural and light industrial area section, and the downstream mixed urban and industrial area section are all high because the difference between the highest and second highest sim values ​​is large (≥0.30).

[0277] Example 2: Identification of ultrashort-chain PFAS in semiconductor factory wastewater and attribution to industrial emission sources

[0278] Sample Information: Wastewater samples from three factories in a semiconductor industrial park in the Yangtze River Delta region: Factory A (etching process wastewater, 8-inch wafer production line, monthly capacity 30,000 wafers), Factory B (wafer cleaning wastewater, 12-inch wafer production line, monthly capacity 50,000 wafers), and Factory C (photoresist stripping waste liquid, 6-inch wafer production line, monthly capacity 20,000 wafers). Sampling time: May 2026 (collected over 3 consecutive days, once at 9:00 AM and once at 3:00 PM each day, for a total of 6 time points). Each sample was collected in 250 mL bottles in polypropylene bottles, stored at 4°C protected from light, and pretreatment was completed within 24 hours.

[0279] Pretreatment: Due to the complex matrix of the wastewater sample, the pretreatment process was adjusted as follows: 100 mL of wastewater was filtered through a 0.22 μm nylon membrane to remove suspended particulate matter → diluted 10 times with ultrapure water (to reduce matrix effect) → WAX SPE extraction (200 mg / 6 mL, activated sequentially with 4 mL methanol and 4 mL water) → after loading, eluted with 4 mL of 1% ammonia and methanol → concentrated to 200 μL by nitrogen blowing → addition of internal standard ( 13 C2-TFA, 2 H3-TFMS) → Volume adjusted to 500 μL. The delayed column system was configured with a short WAX column (30 mm × 2.1 mm) for TFA false positive elimination.

[0280] Instrument analysis: UPLC-HRMS / MS (Waters ACQUITY UPLC + Xevo G2-XS QTOF), WAX analytical column (100 mm × 2.1 mm, 1.7 μm), gradient elution for 15 min (A: 2 mM NH4OAc aqueous solution, B: acetonitrile, 0~2 min 5%B → 2~10 min 5~40%B → 10~12 min 40~95%B → 12~15 min 95%B), ESI - Mode, Full Scan + ddMS 2 +AIF. IMS module turned on (drift tube nitrogen 90 mL / min). TFMS and PFEtS standards were purchased from Wellington Laboratories (purity ≥98%).

[0281] Multi-dimensional data acquisition—the detection data of TFMS and PFEtS are shown in Table 12:

[0282] Table 12 Multi-dimensional Data Acquisition – TFMS and PFEtS Detection Data Table

[0283]

[0284] Multi-dimensional data acquisition—PFEtS detection data and co-detected biomarkers are shown in Table 13:

[0285] Table 13 Multidimensional Data Acquisition – PFEtS Detection Data and Co-detected Biomarkers

[0286]

[0287] VSE-DLM processing—the fragmentation sparsity adaptive evaluation process is as follows:

[0288] TFMS has an MS2 fragment count of 4 (≥3), the fragment sparse adaptive mechanism is not activated, and the weights of each dimension remain at 1.0. PFEtS has an MS2 fragment count of 3 (≥3), the fragment sparse adaptive mechanism is critically inactive, and the weights of each dimension remain at 1.0. TFA has an MS2 fragment count of 2 (<3), triggering the fragment sparse adaptive mechanism, with MS2 weight = 2 / 3 = 0.67, RT weight = 1 + 1 / 3 = 1.33, and CCS weight = 1 + 1 / 3 = 1.33. The visual encoder outputs v∈R for each factory. 256 After CMEM cross-modal mapping, v'∈R 128 .

[0289] VSE-DLM processing – Source attribution matching results (with TFMS as the target compound) are shown in Table 14:

[0290] Table 14 VSE-DLM Processing – Source Attribution Matching Results

[0291]

[0292] As shown in Table 14, the TFMS source attribution results of the three factories are highly consistent, all attributed to the S3 industrial direct emission source. The sim(v', s'3) values ​​are all >0.90, and the difference between the sim values ​​and those of the second highest source type is all >0.60, indicating a high level of confidence. The PFSA / PFCA ratios are all >5 (8.5, 6.2, and 5.8, respectively), which conforms to the typical source fingerprint characteristics of semiconductor industry emissions.

[0293] VSE-DLM processing – Confidence determination results are shown in Table 15:

[0294] Table 15 VSE-DLM Processing – Confidence Determination Results

[0295]

[0296] As shown in Table 15, all six test results met the Level 1 confirmation criteria: p_ident > 0.90 + fragment count ≥ 3 + RT deviation < 0.3 min + 13 C-ratio verification passed + SNR_corrected > 3 + Standards available + Triple false positive elimination passed.

[0297] PGNN Physical Constraint Validation: TFMS (CF3SO3) - The CF bond dissociation energy of CF2 is 485 kJ / mol (compliant with constraint (1)); log Kow = -2.8 < 0.5 (compliant with constraint (2), strong hydrophilicity); the mass difference of CF2 detected a characteristic fragment with m / z 49.9968 Da (compliant with constraint (3)). All three physical constraints were met.

[0298] Conclusion Analysis:

[0299] In this embodiment, two ultrashort-chain PFAS (TFMS and PFEtS) were identified in the wastewater from three semiconductor factories, both with Level 1 confidence levels. The concentration range of TFMS was 28.5–45.2 μg / L, and the concentration of PFEtS was 8.4–12.8 μg / L. These concentration levels are much higher than those in surface water (Example 1), reflecting the characteristics of industrial emission sources.

[0300] The source attribution results are highly consistent: all three factories are attributed to direct industrial emissions source S3, with sim(v', s'3) = 0.91~0.96 and confidence levels of 0.65~0.74 (all high level). The PFSA / PFCA ratios are all >5 (5.8~8.5), which is a typical source fingerprint of semiconductor industry emissions—dominated by PFSAs (TFMS, PFEtS, PFBS).

[0301] The composition of ultra-short chain PFAS in wastewater from factories with different process types varies: In Plant A (etching), the TFMS / PFEtS ratio is 3.5, and 6:2 FTS (1.5 μg / L) is detected, reflecting the degradation contribution of fluorinated etching gases (NF3, CF4); In Plant B (cleaning), the TFMS / PFEtS ratio is 3.8, and HF residue is detected, reflecting the contribution of the hydrofluoric acid cleaning process; In Plant C (stripping), the TFMS / PFEtS ratio is 3.1, and NMP degradation products and a relatively high 6:2 FTS (2.3 μg / L) are detected, reflecting the contribution of fluorinated surfactants in the photoresist stripping solution.

[0302] This embodiment verifies the normal operation mode of VSE-DLM under the condition of fragment number ≥3 (fragment sparse adaptive non-activation), and the effective constraint of PGNN physical constraints on strongly hydrophilic ultrashort chain PFSAs.

[0303] Example 3: Identification of ultrashort-chain PFAS in farmland drainage and attribution of degradation sources of fluorinated pesticides

[0304] Sample Information: Three farmland drainage samples were collected from Zhejiang Province: a paddy field (15 mu, treated with fluazinam in May 2026 to control rice blast, sampling 30 days after application); a vegetable field (8 mu, treated with trifluralin in June 2026, sampling 25 days after application); and an orchard (20 mu, treated with fluopyram in June 2026 to control downy mildew, sampling 20 days after application). Sampling was conducted in July 2026 (3-4 weeks after pesticide application), with 500 mL water samples collected at the drainage ditch outlet, in triplicate.

[0305] Pretreatment: 500 mL of farmland drainage was filtered through a glass fiber membrane (GF / F, 0.7 μm) to remove silt and plant debris → WAX SPE extraction (200 mg / 6 mL, activated sequentially with 4 mL methanol and 4 mL water) → sample loading followed by elution with 4 mL of 1% ammonia and methanol → concentration to 200 μL under nitrogen → addition of internal standard (… 13 C2-TFA, 13 C3-PFPrA) → Adjust volume to 500 μL. The delayed column system is equipped with a WAX short column (30 mm × 2.1 mm) for TFA false positive elimination.

[0306] Instrument detection: Same as in Example 1 (UPLC-HRMS / MS, WAX analytical column, gradient elution for 15 min, ESI) - Mode, Full Scan + ddMS 2 +AIF, IMS module enabled.

[0307] Multi-dimensional data acquisition—TFA and PFPrA detection data are shown in Table 16:

[0308] Table 16 Multi-dimensional Data Acquisition – TFA and PFPrA Detection Data Table

[0309]

[0310] Multi-dimensional data acquisition—PFPrA detection data and co-detected biomarkers are shown in Table 17:

[0311] Table 17 Multidimensional Data Acquisition – PFPrA Detection Data and Co-detected Biomarkers

[0312]

[0313] VSE-DLM processing—Fragment sparsity adaptive evaluation is as follows:

[0314] The MS2 fragment count of the TFA detection signal for the three farmland drainage samples was 2 (<3), triggering the fragment sparsity adaptive mechanism. The weights were adjusted as follows: MS2 weight = 2 / 3 = 0.67; RT weight = 1 + 1 / 3 = 1.33; CCS weight = 1 + 1 / 3 = 1.33. The MS2 fragment count of PFPrA was 3 (≥3), so fragment sparsity adaptation was not activated, and the weights for each dimension remained at 1.0. The visual encoder output v∈R for each sample... 256 After CMEM cross-modal mapping, v'∈R 128 .

[0315] VSE-DLM processing – Source attribution matching results are shown in Table 18:

[0316] Table 18 VSE-DLM Processing – Source Attribution Matching Results

[0317]

[0318] As shown in Table 18, the TFA source attribution results of the three farmland drainage samples are consistent, all attributed to the S2 fluorinated pesticide degradation source, with sim(v', s'2) = 0.85~0.91. The difference between sim and the second highest source type is >0.50, and the confidence level is high. The TFA / PFPrA ratio ranges from 2.7 to 4.8, falling within the fingerprint characteristic range (1~10) of fluorinated pesticide degradation sources.

[0319] VSE-DLM processing – Confidence determination results are shown in Table 19:

[0320] Table 19 VSE-DLM Processing – Confidence Determination Results

[0321]

[0322] As shown in Table 19, TFA was classified as Level 2 suspected identification due to its fragment count of 2 (<3); PFPrA was classified as Level 1 confirmed identification due to its fragment count of 3 (≥3) and the availability of a standard. All 6 test results met the criteria for their respective levels.

[0323] PGNN physical constraint verification: TFA (C2HF3O2) - The CF bond dissociation energy of CF2 is 485 kJ / mol (compliant with constraint (1)); log Kow = 0.14 < 0.5 (compliant with constraint (2)); the mass difference of CF2 detected a characteristic fragment with m / z 49.9968 Da (compliant with constraint (3)). PFPrA (C3HF5O2) - It also satisfies the triple constraint.

[0324] Conclusion Analysis:

[0325] In this embodiment, TFA (Level 2 confidence level) and PFPrA (Level 1 confidence level) were successfully identified in three farmland drainage samples, with TFA concentrations ranging from 280 to 580 ng / L and PFPrA concentrations ranging from 95 to 156 ng / L. The fragment sparse adaptive mechanism maintained an identification accuracy of ≥85% even when there were only two TFA fragments.

[0326] The source attribution results were highly consistent: all three farmland drainage samples were attributed to the S2 fluorinated pesticide degradation source, with sim(v', s'2) = 0.85–0.91 and confidence levels of 0.57–0.66 (all high-level). The TFA / PFPrA ratio ranged from 2.7 to 4.8, which conforms to the fingerprint characteristic range (1–10) of fluorinated pesticide degradation sources.

[0327] The co-detected biomarkers provided independent confirmation of source attribution: trifluoromethylaniline (a degradation intermediate of fluazinam) + trace amounts of fluazinam parent compound were detected in rice paddies; trifluoromethylaniline + n-propyl trifluoroacetate (a degradation intermediate of trifluralin) + trace amounts of trifluralin parent compound were detected in vegetable fields; and trifluoromethylbenzoic acid (a degradation intermediate of fluopyram) + trace amounts of fluopyram parent compound were detected in orchards. Different pesticide types produced different degradation intermediates, but all pointed to the degradation source of S2 fluorinated pesticides.

[0328] This embodiment corroborates the results of the midstream farmland intensive area section (TFA / PFPrA=2.4, S2 pesticide degradation source) in Example 1, verifying the consistency of VSE-DLM in different substrates (surface water vs farmland drainage) and the stability of the fluorinated pesticide degradation source fingerprint (TFA+PFPrA co-detection + pesticide degradation intermediate co-detection).

[0329] A comprehensive comparison of three examples: Example 1 (surface water) → S1 / S2 / S5 multi-source differentiation (Level 2 as the main source); Example 2 (industrial wastewater) → S3 single dominant source (Level 1); Example 3 (farmland drainage) → S2 single dominant source (TFA Level 2 + PFPrA Level 1). These three examples collectively validate the applicability of this method under different substrates, concentration levels, and fragment quantities.

Claims

1. A method for identifying and source attribution of ultrashort-chain perfluorinated and polyfluorinated alkyl compounds, characterized in that: Includes the following steps: S1: Obtain multi-dimensional detection data of ultrashort-chain perfluorinated and polyfluorinated alkyl compounds with carbon chain lengths of C1~C3 in the sample to be tested. The multi-dimensional detection data includes at least two of the following: anion exchange chromatography retention time, MS1 precise mass and isotope mode, MS2 fragment spectrum, ion migration collision cross section, and signal-to-noise ratio corrected for delayed column. S2: Input the multi-dimensional detection data into a pre-trained visual semantic embedding deep learning model, and extract the semantic embedding vectors of ultra-short chain perfluorinated and polyfluoroalkyl compounds through multimodal feature fusion under fragmented sparse conditions. The fragment sparsity condition refers to the condition that the number of MS2 fragments is less than a preset threshold; the multimodal feature fusion includes a fragment sparsity adaptive mechanism, that is, when the number of fragments is less than the preset threshold, the weight of the MS2 fragment dimension is automatically reduced and the weight of the chromatographic retention time dimension and the ion migration collision cross section dimension are simultaneously increased. S3: Perform similarity matching between the semantic embedding vector and the semantic embedding vectors of each source type in the source fingerprint database of ultrashort chain perfluorinated and polyfluoroalkyl compounds, and output the identification confidence level and source classification result of ultrashort chain perfluorinated and polyfluoroalkyl compounds.

2. The method according to claim 1, characterized in that: The preset threshold is 3 fragments; the fragment sparsity adaptive mechanism independently determines the number of fragments for each detected signal during the inference phase; the multi-dimensional detection data also includes... 13 C / 12 At least one of the C isotope ratio and the MS2 fragment mass difference, wherein the MS2 fragment mass difference is the mass difference of 49.9968 Da corresponding to the CF2 unit.

3. The method according to claim 1, characterized in that: In step S1, the signal-to-noise ratio corrected by the delay column is obtained as follows: a delay column is connected in series between the injector and the analytical column of the liquid chromatography-mass spectrometry system, and the corrected signal-to-noise ratio SNR_corrected = (I_delay_sample - I_delay_background) / I_delay_background is calculated. SNR_corrected > 3 is determined to be a real sample signal; where I_delay_sample is the sample signal intensity with the delay column, and I_delay_background is the background signal intensity with the delay column. S1 involves pretreatment of the sample to be tested. The pretreatment method includes: when the sample to be tested is a water sample, taking a water sample and adding... 13 The C-labeled isotope internal standard mixed solution was adjusted to acidic pH, extracted and eluted by a weak anion exchange solid-phase extraction column, and then concentrated. When the sample to be tested was a sediment or biological sample, it was freeze-dried, ground, and then extracted by ultrasonication, followed by extraction and elution by a weak anion exchange solid-phase extraction column and then concentrated. The chromatographic detection conditions adopted a combination of weak anion exchange column and delay column.

4. The method according to claim 1, characterized in that: The visual semantic embedding deep learning model includes: a visual encoder, which contains multiple sub-encoders that process detection data of different dimensions respectively, and the outputs of each sub-encoder are concatenated and then output as a visual feature embedding vector through a fusion layer. The semantic encoder receives the source fingerprint semantic description vector and outputs the semantic feature embedding vector. The cross-modal embedding space mapping module maps visual feature embedding vectors and semantic feature embedding vectors to a shared embedding space, and achieves visual-semantic alignment through an alignment loss function. The alignment loss function includes a combination of four terms: visual-semantic alignment loss, structural identification loss, source attribution classification loss, and physical guidance constraint loss.

5. The method according to claim 4, characterized in that: The fragment sparsity adaptive mechanism adjusts the weights in the feature stitching layer of the visual encoder. When the number of fragments n is less than a preset threshold n0, a weight reduction coefficient w_MS2 = n / n0 is applied to the output feature vector of the MS2 fragment sub-encoder, an increase weight coefficient w_RT = 1+(n0-n) / n0 is applied to the output feature vector of the chromatographic retention time sub-encoder, and an increase weight coefficient w_CCS = 1+(n0-n) / n0 is applied to the output feature vector of the ion migration collision section sub-encoder. The weighted feature vectors are then stitched together and input into the fusion layer. The weight coefficients are deterministic scalars, calculated from the number of fragments using a formula rather than being learnable parameters.

6. The method according to claim 4, characterized in that: In the visual encoder, the MS2 fragment sub-encoder contains a convolutional layer. When the fragment sparse adaptive mechanism is triggered, the convolutional layer is bypassed, and the fully connected layer is used directly to process the MS2 fragment information.

7. The method according to claim 1, characterized in that: The source attribution classification includes the following five source types: a) A photochemical degradation source of refrigerant or halogenated hydrocarbon, characterized by trifluoroacetic acid as the main component and a trifluoroacetic acid to perfluoropropionic acid concentration ratio greater than 100. b) A source of degradation of fluorinated pesticides, characterized by the coexistence of trifluoroacetic acid and perfluoropropionic acid with a concentration ratio of 1 to 10; c) Direct industrial emission sources, characterized by perfluorosulfonic acid compounds as the main component and a concentration ratio of perfluorosulfonic acid to perfluorocarboxylic acid greater than 1; d) A source of degradation for shortened fluoropolymer chains, characterized by a CF2-like distribution of C2-C6 perfluorocarboxylic acid homologues; e) Sources of metabolic degradation of fluoride-containing drugs, characterized by trifluoroacetic acid being the main component in urban sewage and its concentration being positively correlated with population density; The source fingerprint database of ultrashort-chain perfluorinated and polyfluoroalkyl compounds contains semantic description vectors for each source type. These semantic description vectors include compound composition features, concentration ratio features, matrix association features, seasonality features, co-detection marker features, spatial distribution features, degradation pathway features, and environmental migration behavior features.

8. The method according to claim 4, characterized in that: The training process of the visual semantic embedding deep learning model introduces physical guidance constraints, which include: The carbon-fluorine bond dissociation energy constraint limits the predicted fragment abundance distribution to no more than the allowable deviation range of the carbon-fluorine bond dissociation energy. The water solubility-vapor pressure correlation constraint limits the predicted hydrophobicity of ultrashort-chain perfluorinated and polyfluorinated alkyl compounds to no more than 0.

5. The CF2 mass difference constraint restricts the distance between homologous objects in the embedded space to be consistent with the mass difference of the CF2 elements.

9. The method according to claim 1, characterized in that: The S2 also includes trifluoroacetic acid false positive elimination, comprising the following three mechanisms: 1) By using delayed column pre-separation, the system background trifluoroacetic acid signal is separated from the sample trifluoroacetic acid signal in terms of retention time; 2) By deducting blank contributions from trifluoroacetic acid through system blank deduction; 3) Through 13 C / 12 Cross-validation of C isotope ratios, when the sample contains trifluoroacetic acid 13 When the deviation between the C ratio and the natural abundance is less than the first threshold, it is determined to be a true sample signal. 13 A false positive is defined as a C ratio deviation from the system background trifluoroacetic acid that is less than the second threshold.

10. The method according to claim 1, characterized in that: The confidence levels for identification include: Level 1: Confirmation and identification, which must meet the following requirements: standard comparison, isotope internal standard quantification, retention time deviation less than 0.1 min, 13C isotope ratio verification, and MS2 fragment matching; Level 2: Sparse fragment identification requires meeting the following criteria: multimodal fusion matching of the visual semantic embedding model, retention time deviation of less than 0.3 min, at least one MS2 fragment match, ion migration collision cross-section deviation of less than 2%, and passing false positive elimination verification. Level 3: Provisional identification, requiring MS1 exact quality matching deviation to be less than 5 ppm and retention time deviation to be less than 0.5 min; the similarity matching is calculated using cosine similarity, and the confidence level of the source attribution classification result is determined based on the difference between the highest similarity and the second highest similarity.

Images