Chemical Fingerprints of New vs Weathered Plastics: A Machine Learning Approach

Abstract

Regulatory initiatives to mitigate plastic pollution face challenges in identifying microplastics and their sources. Recent advances in computational fingerprinting offer opportunities for the forensic investigation of plastic pollution sources. Here, we present an open-source computational workflow integrating multi-instrumental, nontargeted (organic and inorganic) data from three distinct mass spectrometry methods to evaluate two approaches for potential plastic pollution source tracking. The workflow includes a novel data imputation, normalization, and feature selection strategy with a Random Forest classifier applied to data sets of 42 pristine (store-bought) and 21 weathered (environmentally collected) plastics from U.S. retail products and California beaches. To develop a workflow for source identification, we first, evaluated an “Universal Fingerprint” approach, which tested whether chemical profiles from pristine, store-bought plastic products could be used to identify weathered environmental counterparts based on their chemical additive patterns. This approach yielded low classification accuracy due to high geographic and manufacturer-specific formulation variability for plastic products that the random forest model could not sufficiently resolve. Second, we evaluated a “Suspect–Source Comparison” approach that simulated a scenario where the chemical fingerprint of an environmentally sampled plastic was matched to the fingerprint of a suspect source. Under this paradigm, the workflow achieved a high classification accuracy across all platforms (up to 1.00 MCC for ICP-MS/MS and 0.79 for HPLC-QToF-MS). Additionally, we found that the different analytical techniques functioned best as complementary data sets to distinguish product categories, while combining all raw data sets prior to classification introduced excessive variability that hindered the classification. Our results demonstrate the feasibility of using relative chemical similarity to connect environmental plastics to potential local pollution sources. The presented workflow establishes a methodological foundation for localized source tracking, environmental litigation, and targeted regulatory monitoring.

1. Introduction

Plastic emissions are a global pollution problem with mounting evidence of ecological impacts and risks to human health. − Ecotoxicological research has identified thousands of potentially toxic chemicals associated with plastics in use globally, making plastics a complex mixture pollution issue. − Environmental weathering and adsorption of pollutants on environmental plastics ,, further increase the complex chemical profiles of plastic pollution, which hinders the development of effective regulation and remediation of plastic pollution. − Moreover, current routine monitoring programs only screen for a limited number of plastic-associated chemicals from lists of compounds that have been deemed problematic. − This can lead to the omission of relevant environmental contaminants from regulatory monitoring. As a result, analytical and data analysis methods are needed to capture the chemical complexity of plastic pollution, characterize plastics and their associated chemicals, support the comparison of environmental plastics against candidate source products, and ultimately inform and enable source attribution, reduction, and effective remediation.

Increasing interest in plastics-associated chemicals combined with greater accessibility to nontargeted analyses provides a unique opportunity to study the chemical complexity of plastics and identify potential “chemical fingerprints” for specific plastic sources or products. Yet, extracting plastic chemical fingerprints and source tracking is challenged by the large data sets that these analyses produceoften consisting of thousands of chemical features and feature patterns. ,, Computational fingerprinting and machine learning (ML) approaches offer a solution for this problem. They have already been successfully used as tools to extract representative chemical fingerprints for source attribution of contaminants of emerging concerns (not associated with plastics). − However, these approaches have only been used for data sets produced analyzed by single analytical techniques or individual chemical classes. Given the complexity of plastic chemical mixtures, a class-by-class or technique-specific analytical approach cannot capture the full diversity of plastic chemical fingerprints. Moreover, opportunities to use such analytical data to track dominant local plastic pollution sources or specific plastic polluters have rarely been explored. , Clearly, there is a need for an integrated, multi-instrument ML workflow that leverages nontargeted data to derive robust plastic chemical fingerprints for the characterization and classification of plastics based on their chemical composition.

This study presents a computational workflow that includes feature selection, comparison of the influence of data filtering, data imputation, and normalization, and the application of a Random Forest classifier on the combination of multiple large analytical data sets including automatic thermal desorption-gas chromatography–mass spectrometry (ATD-GC-MS), high-performance liquid chromatography quadrupole-time-of-flight mass spectrometry (HPLC-QToF-MS), and inductively coupled plasma-tandem mass spectrometry (ICP-MS/MS). The workflow was developed with the goal to systematically evaluate the feasibility of plastic source tracking using two distinct approaches: 1) through the matching of representative chemical fingerprints from environmentally sampled plastic items to equivalent store-bought plastic items and 2) through the matching of representative chemical fingerprints of environmentally sampled plastics to specific local suspect polluters or pollution sources.

The specific objectives of this study were to enable the comparison of environmental plastic pollution against candidate source products by a) identifying representative chemical fingerprints from the nontargeted chemical profiles of plastic items, b) establishing the boundary conditions required to reliably link environmentally sampled plastics to their suspected source origins, and c) investigating the feasibility of utilizing multiple analytical techniques as complementary diagnostic data sets.

2. Materials and Methods

2.1. Sample Collection and Preparation

In this study, a total of 63 plastic items were investigated (Table S1). Of these, 21 were environmentally sampled items (sample code USE) collected based on data from a 30-year beach survey (1988–2021) in California, USA, with sampling locations including Point Reyes National Seashore, Los Angeles Harbor, and the Gyre Education Sample Area (approximately 35° N, 136° W) (Table S1). The environmentally sampled plastics were among the most common items found in this survey. The remaining plastics were store-bought, with 42 purchased in the same areas in California where the environmental samples were collected (sample code USSB). These store-bought items were purchased with the aim of acquiring (closely) matching items to those collected from the environment wherever possible to enable more relevant comparisons between the sample sets.

All samples were cut into 1–2 mm (diameter) microplastic-sized fragments using ceramic scissors that were cleaned with acetone. The microplastic fragments were analyzed using three techniques: ATD-GC-MS, HPLC-QToF-MS, and ICP-MS/MS. For ATD-GC-MS analysis, 76 individual measurements were obtained in accordance with Markes international application note 161. For HPLC-QToF-MS analysis, 504 individual measurements were collected using a previously published leaching and nontarget analysis method. For ICP-MS/MS analysis, a previously validated microwave-assisted acid digestion (MWAD) method was applied to collect 91 individual measurements with three replicates of each plastic product. Across the three analytical platforms, each plastic category was represented by a minimum of 3 to 5 physical replicates per item, with the exact number varying by platform depending on material availability (Supplementary Table S1). Detailed descriptions of sample handling procedures, quality controls, and instrumental analysis are presented in Supporting Information Section S1 and S2. No dedicated surface cleaning step (e.g., rinsing or sonication) was applied to the environmental samples prior to cutting. This was a deliberate choice, as surface-sorbed contaminants may carry chemical signatures of the environmental conditions that a plastic has been exposed to. For the HPLC-QToF-MS analysis, a multipoint leaching protocol (SI Section S1.1.2) with leachate samples collected on days 0, 2, 15, and 34 allowed for an evaluation of leaching kinetics, providing insight into whether detected compounds were predominantly surface-associated or embedded within the bulk polymer. For all three analytical techniques, the computational workflow included blank subtraction and feature filtering that retained only features shared between store-bought and environmental plastics, thereby minimizing the influence of surface-sorbed environmental contaminants on the classification model.

It is important to note that the environmentally sampled plastics and their store-bought counterparts were not strictly paired samples. However, great care was taken during sampling to ensure that store-bought counterparts were as close to identical as possible to the environmentally sampled items, including matching the exact manufacturers and purchasing products near the sampling sites. In addition, a selection of samples was analyzed using laser-direct infrared spectroscopy to confirm that the selected store-bought samples matched the polymer type of the environmental equivalents (Supplementary Table S3). This approach does not eliminate the possibility of formulation differences between environmentally sampled plastics and their store-bought counterparts but was designed to minimize their influence while preserving the use of real-world environmental samples and maintaining the ecological relevance of the study design.

2.2. Chemical Analysis

The ATD-GC-MS analysis was performed using a Markes International TD-100xr, with a toluene-D₈ internal standard (IS) used for all analyses. The MS instrument scan parameters (for ATD-GC-MS) were configured with a mass range of 34 to 350 Da. For HPLC-QToF-MS analysis, all samples were analyzed using an Agilent 1260 Infinity II HPLC system with an Agilent 6546 QToF-MS (Santa Clara, California, USA) using a method that was previously reported in Fries and Sühring (2023). Upon data collection, nontargeted analysis was performed on the sample data using the “Find by Molecular Feature” mining algorithm within the Compound Discovery workflow of Agilent’s MassHunter Qualitative Analysis (v10.0) software (Santa Clara, California, USA). The raw data for each sample was then exported into a data table with Excel format for further filtering and analysis. For ICP-MS/MS analysis, samples were analyzed using an Agilent 8800 ICP-MS/MS (Agilent Technologies, Japan) coupled to an ESI SC-4 DX FAST autosampler (Elemental Scientific, USA). Throughout the measurements, an in-house QC solution (Inorganic Ventures, USA) was measured at regular intervals. Further information concerning the sample preparation and instrument settings can be found in previous publications. − Multielement data were processed using MassHunter 4.4 (Agilent Technologies, Japan) and exported for further evaluation in a custom Excel workflow. Further detailed descriptions of software settings and QA/QC protocols for the chemical analyses were presented in Supporting Information Sections S1.2 and S2.

2.3. Computational Fingerprinting Workflow

The data processing, statistical analyses (univariate and multivariate), training of machine learning classifiers, and comparison of feature selection approaches were performed using R with RStudio (version 2023.06.0 + 421). All scripts are available at https://github.com/huymanhnguyen0811/Microplastic-Fingerprinting. The workflow (Figure ) included data processing, statistical analyses, training machine-learning classifiers, and comparison of feature selection approaches. Details of each step of the workflow are explained in the following sections.

1.

Flowchart illustrating the steps used in the computational workflow including sample collection and instrumental analysis (purple), chemical fingerprinting workflow, with a source-tracking assessment developed herein (blue and green). Steps are numbered (1–5) and in blue. Details of each step are in green. Critical “go/no go” checks are in red.

2.3.1. Data Processing

2.3.1.1. Peak Alignment and Blank Subtraction

To identify recurring chemical features across different samples, the detected peaks were aligned based on their mass-to-charge (m/z) and matching retention time windows. Appropriate retention time windows were identified by inspection of the retention times of internal standards, i.e., a solution of isotope-labeled standards spiked into all samples prior to analysis, covering a range or properties and polarities. The raw data included 8,385 features detected with ATD-GC-MS, 764,921 features detected using HPLC-QToF-MS, and 47 trace elements identified using ICP-MS/MS. After peak alignment, there were 2,918 features for ATD-GC-MS analysis and 187,005 features for HPLC-QToF-MS analysis (all ICP-MS/MS data was retained). Further details on choosing appropriate RT windows for peak alignment can be found in Supporting Information Section S3.1.

Following alignment, a blank-subtracted signal response for each feature was calculated by subtracting the signal in the blank from its signal in the corresponding samples. Features with blank-corrected signal responses smaller than three times the standard deviation of their signal responses in the blank samples were removed from the data set. Following blank correction, 1,221 features remained for ATD-GC-MS analysis, 181,347 features for HPLC-QToF-MS analysis, and 47 features (element-gas mode combinations) for ICP-MS/MS analysis.

2.3.1.2. Feature Reduction

Seven analytical data combinations were established from the three analytical data sources ATD-GC-MS, HPLC-QToF-MS, and ICP-MS/MS: individual data sets from ATD-GC-MS (“GC”), HPLC-QToF-MS (“HPLC”), and ICP-MS/MS (“ICP”), as well as combinations of the raw data from ATD-GC-MS and HPLC-QToF-MS (“GC-HPLC”), ATD-GC-MS and ICP-MS/MS (“GC-ICP”), HPLC-QToF-MS and ICP-MS/MS (“HPLC-ICP”), and data from all three analytical techniques (“GC-HPLC-ICP”). Combined data sets were built by vertically stacking samples into a single matrix, with missing platform features zero-filled. Because this created a block structure where each sample was confined to a single platform, no item-level cross-platform data was truly integrated. We therefore treated these combined matrices as strictly exploratory and excluded them from our main classification findings. Since the analytical techniques enable the detection of different chemistries, features unique to one analytical technique appeared as missing values (MVs) in data sets from the other analytical techniques. To ensure that these MVs were missing at random, for each of the seven data combinations, the number of chemical features and percentage of MVs before and after feature filtering were calculated. Pearson correlation coefficients were calculated for the following pairs: (i) m/z versus number of MVs, (ii) retention time versus number of MVs, and (iii) mean response versus number of MVs. Near-zero correlations suggested that MVs were randomly distributed across the data set and indicated the true absence of those chemical features.

Following an examination of the missingness type of MVs, three feature reduction approaches were applied on all seven data combinations: the “regular 80% rule”, “modified 80% rule” and an in-house feature filtering. The “regular 80% rule” is an approach where chemical features with MVs in more than 20% are removed from the data. The “modified 80% rule” excluded chemical features from data sets that had more than 20% of MVs in each plastic category. In addition, an in-house feature filtering method was tested to remove features that

• 1.had ≥ 90% MVs, unless the features were present in 90% of a specific plastic category. These features were retained in the data as potential diagnostic marker compounds for a specific type of plastic product,
• 2.appeared in only one sample,
• 3.were present in only one plastic category, and
• 4.had fewer than three data points across at least two plastic categories.

To facilitate category classification of environmental plastics from their store-bought counterparts, only features that existed in both store-bought and environmental data were selected for further analysis. The resulting number of retained features after feature filtering are presented in Supplementary Table S4. The selection criteria were designed to retain chemically resilient features that persist through environmental degradation, thereby providing consistent diagnostic markers for linking weathered plastics to their store-bought counterparts. While compounds unique to either new or weathered plastics, such as environmental transformation products or additives lost during degradation, could offer valuable insights into weathering pathways and emerging contaminant discovery, their inclusion introduced a substantial increase in missing values and computational cost that ultimately reduced classification accuracy. Therefore, these features were excluded from the classification workflow but represent a promising avenue for future research on plastic weathering chemistry. Additionally, for ICP-MS/MS data, elements such as Na, Ba, Mg, K, and Ca were excluded from the analysis. These elements are ubiquitous in both environmental and commercial materials and thus offer low specificity for distinguishing between plastic categories. After feature reduction, the number of features in the individual data sets were 51 (“GC”), 380 (“HPLC”), 47 (“ICP”).

2.3.2. Classification of Environmental Plastic Samples into Plastic Categories Using Machine Learning

To evaluate classification performance and ensure the generalizability of this workflow across the analytical data combinations, base model validation was structured using a strict leave-one-item-out grouped k-fold cross-validation (LOIO–CV) scheme. , Under this scheme, every physical replicate of an item (including both cut pieces and instrumental injections, identified by a shared plastic item identifier) was strictly confined to a single fold. This ensured that no replicate of a held-out object appeared in the training set during baseline evaluations, with hyperparameter tuning performed inside a grouped inner loop. The number of folds, k, was dictated by the minimum class frequency within the partitioned data set. Final per-object predictions were derived via majority vote across each plastic item’s replicates, and classification performance was reported at the plastic item level (e.g., plastic cups, cigarette lighters, balloons, etc.) using the Matthews correlation coefficient (MCC). For the primary internal classification task, all items (store-bought and environmental) were pooled and partitioned using this LOIO–CV scheme.

Source tracking was assessed as a distinct task from within-data set classification to evaluate how effectively the workflow could link environmentally sampled plastics back to their store-bought counterpart categories (Table S1). Rather than treating data splitting as a statistical validation exercise, the matched data set (items possessing both pristine store-bought and weathered environmental samples) was strategically partitioned into specific train/test configurations to model two distinct forensic approaches. Test Case 1 (The Universal Fingerprint Approach): This evaluated whether generalized chemical fingerprints learned from pristine store-bought items could autonomously identify weathered environmental counterparts without a comparison point. This was modeled using strict, baseline-free configurations where models were either trained exclusively on pristine store-bought items and tested on environmental items or evaluated under strict LOIO–CV where no matched replicate of a held-out environmental item appeared in the training data. Test Case 2 (The Suspect–Source Comparison Approach): This scenario evaluated the workflow’s capacity for targeted forensic attribution when a reference comparison standard is available. This was modeled using a replicate-shared configuration, where physical replicates of the same matched environmental item were distributed across both training and test sets (while ensuring individual instrumental measurements were never leaked). Rather than constituting an artificial optimistic bias, this configuration operationally simulates the real-world deployment of comparing a recovered environmental plastic against a physical suspect reference sample. By comparing these testing approaches, we aimed to evaluate the necessary boundary conditions for forensic source tracking of plastic pollution.

2.3.2.1. Optimization of Missing Data Imputation and Normalization

After splitting the data into training and test sets, to reduce potential bias in the classification, MVs were imputed and the data set normalized. Within each outer training fold, imputation and normalization combinations were compared by a grouped leave-one-item-out inner cross-validation. Once the optimal combination was identified based on the outer training fold, the preprocessing model was refit on the complete outer training fold and applied to the held-out outer test fold as an add-on procedure. This strict separation of train-test ensures that the reported performance metrics reflect a realistic application scenario where the model encounters entirely unseen data, preventing the optimistic bias often introduced by data-dependent preprocessing.

Twenty-four combinations of six MV imputation methods (zero, median, mean, half global minimum, k-nearest neighbor (kNN), and random forest) and four normalization methods (none, total sum normalization (TSN), min-max, and z-score normalization) were evaluated.

Each imputation and normalization combination was hierarchically ranked by weighing three criteria in the following order: (1) cross-validated Matthews correlation coefficient (MCC), , where scores closer to 1 indicate better multiclass classification performance; (2) cluster resolution score, measuring the separation between plastic categories after imputation; and (3) Pearson correlation between original and processed data, ensuring imputation did not distort inherent data trends. The highest-ranked combination was selected for each of the analytical data combinations. Full equations and mathematical definitions for all three metrics are provided in Supporting Information Section S3.

2.3.2.2. Assessing ML Classification Feasibility with Univariate Statistical and Clustering Analyses

To assess classification feasibility of plastic categories based on the individual and combined analytical data sets, hierarchical clustering analyses (HCA) were performed on imputed and normalized chemical features. In addition to the clustering analyses, multiple pairwise tests were performed to explore specific chemical features that could be used to distinguish different plastic categories. HCA was performed using vegan R package using Bray–Curtis dissimilarity index as distance metric and average linkage method. Euclidean distance has been used for nontargeted analysis, , but this method is inappropriate for our data, which is sparse (i.e., data sets with a disproportionate number of 0 s) and high-dimensional (i.e., data sets with hundreds to thousands of variables). Student’s t-test was performed on normally distributed data, and Wilcoxon Rank Sum test was performed on for non-normally distributed data with statistical significance (α) set at 0.05. Data normality was assessed using the Shapiro-Wilk test to determine the appropriate statistical test. P-values from the pairwise comparisons were adjusted using the Bonferroni correction to control for multiple testing.

2.3.2.3. Random Forest Classification

Random forest classification was used to identify diagnostic chemical fingerprints for each plastic category, to link environmental plastics to their store-bought counterparts (test case 1), or to the specific suspect source (test case 2). Random Forest was implemented using the ranger package within the caret framework in R. , The ranger implementation was selected for its computational efficiency with high-dimensional data. To further optimize processing time, parallel processing was employed using the foreach and doParallel backend packages, , thus allowing for concurrent execution across multiple CPU cores.

The accuracy and generalization of machine learning algorithms across various data sets and problems depend on their hyperparameters (i.e., parameters used in model training). For the presented analysis, two hyperparameters were optimized using the expand.grid­() function in R: 1) the number of features considered at each split (mtry) as the square root of p or binary logarithms of p, where p is the number of predictors; and 2) the number of trees (ntree) as a range of values (100, 500, 1000, and 2500). These parameters were optimized with stratified k-fold cross-validation, where k is the minimum sample count of a plastic category in the training data set. The hyperparameter combination that achieved the highest accuracy on average across k-folds was chosen for the final model. The final optimized Random Forest classifier was then refitted over the individual training data set. Finally, RF performance for each of the seven data combinations was evaluated on the testing data set.

2.3.2.4. Feature Selection

In addition to hyperparameter tuning, feature selection was performed to identify subsets of diagnostic chemical features for accurate classification of plastic product types while aiming to reduce noise and optimize computational time. The feature selection methods were compared against the full feature setting to identify optimal feature subsets for source tracking.

Pair-wise tests

Pair-wise tests, as previously described, were performed with statistical significance (α) set at 0.05. Features with corrected p-values <0.05 were selected for a final random forest retraining and then tested on the testing data set.

Recursive feature addition

Recursive feature addition was done by fitting a baseline random forest model on training data using all features. Feature importance was derived using the Gini impurity index, which measures the total decrease in node impurity (Gini index) resulting from splitting on a specific variable. The single feature with the highest importance served as the initial recursive feature addition input. Subsequently, remaining features from the original data set were iteratively tested for addition. To manage the computational load of this iterative process, the candidate feature evaluation was parallelized. In each iteration, the feature that yielded the greatest improvement in the MCC was permanently added to the model. This process was repeated until no remaining feature improved the model. Finally, the optimal feature set with the highest MCC score was selected for final random forest training and evaluated on the testing data set.

Top 10/25/50/100 features based on Random Forest feature importance

In this approach, the Random Forest model was trained on all features from the original data set to obtain feature importance based on the Gini impurity metric, as described previously. Reduced feature sets were created by ranking predictors in a descending order of importance. Subsets of the top 10, 25, 50, and 100 most important predictors were selected. The subsets were utilized to retrain the final models, and classification performance was assessed on the testing data set.

3. Results and Discussion

3.1. Data Processing

3.1.1. Peak Alignment

Nontargeted ATD-GC-MS and HPLC-QToF-MS analyses resulted in data with several hundred to thousands of features per sample that involved variation in a compound’s retention time (RT) and mass-to-charge ratio (m/z). For ATD-GC-MS data, RT and m/z window of RT ± 0.05 min and m/z ± 0.05 were applied. For HPLC-QToF-MS data, RT ± 0.4 min and RT ± 0.28 min for peaks eluted before and after 6 min, respectively, and m/z ± 0.0005 were applied. A larger molecular ion window was applied for ATD-GC-MS data compared to HPLC-QToF-MS data since ATD-GC-MS data had lower mass resolution. In the case of the targeted ICP-MS/MS data the peak alignment process was unnecessary. Further information on determination of retention time and m/z windows is presented in Section S3.1.

3.1.2. Feature Reduction

Across the seven data combinations, the Pearson correlation coefficients between m/z and the number of MVs, retention time and the number of MVs, and mean response and the number of MVs were all close to zero (Table S5). This indicated that the frequency of MVs was not associated with m/z, retention time, or signal response. Therefore, it was assumed that MVs were missing-at-randomthroughout the data set, likely due to those chemical features being genuinely absent in the plastic items at nondetectable concentrations.

Since MVs were determined as missing at random and the percentage of MVs of both GC and HPLC data sets were high, majority of chemical features were removed when either the regular or “modified” 80% approach was applied as compared to in-house approach (Table S5). This was not surprising, considering that these two rules were developed for complete data sets with randomly generated missing values, rather than originally incomplete data sets. Furthermore, the 80% or “modified” 80% rules are only supposed to be used when the percentage of MVs prior to feature filtering is below 50% ,, which was not the case for the data set in this study. Therefore, instead of the 80% or “modified” 80% rules, the in-house filtering method was applied. This method successfully eliminated missing values in the final data set and retained on average 5.5% of chemical features and 17% of percentage of signal response as compared to the regular or “modified” 80% rule (Table S5).

Among the 51 retained features of ATD-GC-MS data after feature reduction, 6 features (11.8%) coeluted with another feature of differing m/z at the 0.005 min retention-time window (used throughout GC peak alignment – Section ). This coelution of feature is substantially lower as compared to the postpeak alignment list (2491 out of 2918 features −85.4%), indicating that residual in-source fragments and isotopes are largely confined to the features that do not survive the data cleaning process. Hence, AnalyzerProXD deconvolution and the downstream data cleaning process reduced the feature set to a low-redundancy set prior to Random Forest classification.

3.2. Chemical Fingerprinting

3.2.1. Optimization of Missing Data Imputation and Normalization

Previous research has reported data imputation and normalization assessment using metrics such as normalized root-mean-square error and PCA-Procrustes analyses to compare the imputed matrices to the complete matrices on a modified version of the complete data (with randomly introduced missing values). ,, However, since the data in this study were originally incomplete (missing at random), alternative metrics were evaluated to provide a more robust assessment. By utilizing MCC score, cluster resolution, and Pearson correlation, combinations of data imputation and normalization that achieved best classification ability with decreased feature correlation and covariance were successfully selected (Table , Table S6).

1. Sensitivity of Classification Performance to the Choice of Missing-Data Imputation and Normalization, Evaluated as a Preprocessing Sensitivity Analysis Rather Than a Single Optimal Selection.

Data set	Forensic approaches	Imputation	Normalization	MCC score	Cluster resolution	Pearson correlation score
GC	Test Case 1	Half global minimum	TSN	0.268	1.01	0.23
	Test Case 2	Zero	Min-Max	0.86	1.26	0.22
HPLC	Test Case 1	Zero	TSN	0.293	0.73	0.22
	Test Case 2	Zero	Z-Score	0.81	1.26	0.37
ICP	Test Case 1	-	None	0.1258	5.07	1
	Test Case 2	-	TSN	0.4	1.79	0.42

The optimal combination of imputation and normalization, characterized by the highest classification accuracy on each cross-validation fold, did not consistently correspond to the highest scores in cluster resolution and Pearson correlation (Table S6). For example, for the HPLC data set, Random Forest and Min-Max scaling had a higher cluster resolution (1.35 vs 0.73) and Pearson correlation score (0.39 vs 0.22) than the zero-TSN combination but had lower MCC score (Table S6). These results highlight a trade-off between the predictive performance and structure-preserving transformations in imputation and normalization. Consequently, the highest MCC score for classification was prioritized, and results on clustering and correlation were considered as complementary diagnostics.

Overall, different combinations of imputation and normalization methods were required for each of the seven data combinations to create an optimal data set for ML classification (Table ). Small value replacements (zero, mean, or median) were the primarily chosen methods for data imputation. The performance of small replacement value imputation differed from previous studies, , which reported poor outcomes for this type of approach as compared to multivariate imputation such as Random Forest or kNN. This is because our data were not simulated, and where weak or no relationships existed between MVs and observed values, the multivariate imputation proved ineffective.

Regarding normalization, our results aligned with recent benchmarking studies for metabolomic data sets, suggesting that no single normalization method is universally optimal. , We observed that the optimal normalization technique was highly context-dependent, varying across different cross-validation folds and data splitting scenarios (Table S6). Consequently, our results do not support the prescription of a single “universal” normalization method but instead, demonstrating the value of an automated, adaptive workflow. By dynamically optimizing imputation and normalization based on a composite of three metrics (MCC score, Cluster Resolution, and Pearson correlation), our workflow offers a flexible and reliable automation to select the most appropriate preprocessing strategy for the specific structure of the data.

3.2.2. Distinction of Plastic Products Based on Cluster Analysis

Plastics from the same category, for example, plastic cups (Figure ), cigarette lighters (Figure S3), and balloons (Figure S4), clustered together, indicating similarity in chemical compositions within each category. Moreover, clustering of store-bought and environmental items within several categories (for example, plastic bottles, Figure ) indicated within-category chemical similarity. Meanwhile, plastic categories were separated from one another, suggesting the feasibility of the category-level classification.

2.

Hierarchical clustering analysis (Bray–Curtis distance, average linkage) using all nontargeted chemical features from ATD-GC-MS samples. The store-bought plastics were labeled gray and environmental plastics were labeled green. Cluster of plastic categories (red rectangle) indicated the feasibility of source tracking between store-bought and environmental.

Nevertheless, the feature patterns showed cases in which environmental and store-bought items from the same category did not cluster together, such as plastic cups and balloons in the ATD-GC-MS data set (Figure ). This could be due to environmental weathering that distorted the feature patterns, making it insufficient for reliably linking environmental plastics back to their store-bought counterparts. Consequently, instead of relying solely on global feature patterns, we focused on identifying and adding specific diagnostic features for the plastic category and source tracking into the ML classification.

The HCA indicated that there were sufficient unique feature patterns among the plastic categories to support the ML classification. Building on this foundation, we proceeded to test the potential source identification based on the two test cases: 1) the universal fingerprint approach and 2) the suspect-source comparison approach.

3.2.3. Classification of Plastic Categories Using Machine Learning

3.2.3.1. Impacts of Feature Selection

We evaluated the impact of feature selection on model accuracy and found that utilizing the full feature set provided optimal or near-optimal results across all of the platforms. The full feature set outperformed all selection methods on ATD-GC-MS and ICP-MS/MS and tied with the top selector on HPLC-QToF-MS. Conversely, recursive feature addition underperformed on all platforms, dropping object-level MCC scores significantly compared to the full feature baseline (ATD-GC-MS: 0.187 vs 0.326; HPLC-QToF-MS: 0.265 vs 0.361; ICP-MS/MS: 0.188 vs 0.231) (Table S7). Consequently, we retained the full feature set for the final classification models. While feature selection did not improve the overall accuracy, it remains valuable for generating compact, interpretable marker panels. The chemical features prioritized using recursive feature addition proved highly diagnostic; for example, the significant feaures identified using HPLC-QToF-MS mostly eluted before the 10 min mark (Figure S5). This elution profile was consistent with polar additives such as colorants, flame retardants, antioxidants, and plasticizers commonly found in plastic products. ,,

While the full feature set performed best for this analysis, feature reduction methods like recursive feature addition and pairwise filtering should remain part of the initial preprocessing pipeline for other types of analytical data.

3.2.3.2. Evaluating Source Attribution Approaches and Diagnostic Chemical Markers

The workflow’s capability to identify the origins of environmental plastic samples was evaluated by comparing the chemical fingerprints of store-bought data against weathered field samples (Table S1). The results showed that, using this workflow, it was not possible to derive generalized chemical fingerprints from pristine store-bought items to identify weathered environmental counterparts without a localized comparison point. Under leave-one-item-out grouped cross-validation, where no replicate of a held-out environmental item appeared in the training set (Universal Fingerprint approach; Test Case 1), none of the matched objects across all three analytical data sets yield correct identification (0 out of 4 for ATD-GC-MS, 1 of 5 for HPLC-QToF-MS, and 0 of 4 for ICP-MS/MS) (Table S8).

The results highlighted the substantial impact of environmental weathering on the chemical compositions of plastics. These findings were consistent with previous studies on the changes in leached additive patterns from plastics over time, ,,− as well as with previously reported significant geographic differences in plastic additive patterns even in plastic item from the same brand. While the low classification accuracy might be improved by using different classification models in the future, it was clear that plastic source tracking is substantially hindered if weathered data is not included in the analysis. This is important, considering that currently most spectral libraries for chemical additives in plastics are based on pristine plastics. The findings were also consistent with previous studies that reported an improved accuracy of spectral matching when spectral libraries of environmental plastic particles and commercially reference libraries were used together. ,

When physical replicates of the matching source product were introduced into the training database (Test Case 2), simulating a forensic investigation where a recovered environmental plastic is compared directly against a sampled local suspect source, classification performance increased substantially. Across the analytical platforms, the proportion of correctly classified products increased to 100%, with an increase in MCC score from −0.365 to 0.51 (ATD-GC-MS), from 0.053 to 0.89 (HPLC-QToF-MS), and from −0.083 to 0.9 (ICP-MS/MS) (Table ).

2. Source Tracking (Store-Bought-to-Environmental) Results Based on the Two Forensic Approaches (Strict LOIO–CV and Suspect Matching).

Platform	Forensic approach	Correctly classified products	MCC score
ATD-GC-MS	Test Case 1: Open-Set (Strict LOIO–CV)	0 of 4 (0%)	–0.365
	Test Case 2: Closed-Set (Suspect Match)	4 of 4 (100%)	0.51
HPLC-QToF-MS	Test Case 1: Open-Set (Strict LOIO–CV)	1 of 5 (0%)	0.053
	Test Case 2: Closed-Set (Suspect Match)	5 of 5 (100%)	0.89
ICP-MS/MS	Test Case 1: Open-Set (Strict LOIO–CV)	0 of 4 (0%)	–0.083
	Test Case 2: Closed-Set (Suspect Match)	4 of 4 (100%)	0.9

This increased performance clearly showed that our method works well for picking out the “guilty” suspect from a candidate pool of local polluters, while it fails to create generally applicable plastic product fingerprints. Consequently, we base the workflow’s primary forensic utility on this targeted, closed-set suspect comparison approach (Test Case 2).

Moreover, our results showed a clear advantage of using data sets from different analytical techniques as complementary chemical fingerprinting tools rather than combining the data sets into a single data set. A truly combined multiplatform analysis with increased sample size would be an interesting future expansion of this work.

To provide chemical context for the selected fingerprints, the recursive feature addition-selected features from each analytical data set were tentatively annotated by matching their monoisotopic masses against the PlastChem database (Table S9). The annotated features spanned a range of functional classes including colorants, flame retardants, plasticizers, antioxidants, monomers, and processing aids. Notably, the industry sectors associated with these chemical fingerprints overlapped with the plastic product categories in our data set, such as food-contact plastics, food packaging, and textiles. For example, one of the important features from the HPLC data set was tentatively identified as triethyl phosphate, an organophosphate ester used as a solvent and flame retardant in plastic food packaging. Among the top-ranked features from the ICP data set, cadmium (Cd), a toxic heavy metal previously reported at elevated levels in coffee packaging, was identified as a diagnostic element. For ATD-GC-MS, N-(2-Hydroxyethyl)­ethylenediamine was among the selected features, consistent with its known use in building and construction plastics and textiles. These tentative annotations provide supporting evidence that the computational workflow extracts chemically meaningful fingerprints that are relevant to the plastic product categories under investigation. However, it should be noted that these annotations remain tentative as definitive identification would require confirmation through reference standards and further fragmentation via tandem MS/MS.

3.3. Implications

The presented ML workflow automates missing data handling, feature selection optimization, and supervised Random Forest prediction to assess the relative chemical similarity between environmental and store-bought plastic products. It can be used to identify high-intensity or high-occurrence chemical features shared between plastics to yield a robust set of representative chemical fingerprints. These fingerprints can be used to track specific local suspect polluters or can be used as a starting point for compound elucidation that can help to identify potential environmental and human health exposure risks from commonly used plastic products. These insights can guide future analytical technique selection, expansion of current plastic databases with environmental fingerprints, and support evidence-based plastic source attribution (i.e., holding polluters accountable for their plastic waste).

Crucially, this workflow was designed to systematically evaluate the feasibility of plastic source tracking using two distinct approaches: an open-set universal fingerprint (Test Case 1) and a localized, closed-set suspect-matching approach (Test Case 2).

Our findings demonstrate that this workflow is best used as a suspect-source comparison rather than open-set source identification. When attempting to match an environmental sample against an unconstrained, unknown universe of possible origins (Test Case 1), transferability failed due to the geographic and manufacturer-specific formulation variability. However, when the workflow was used to compare the chemical fingerprint of an environmental plastic against one or more nominated candidate (“suspect”) sources (Test Case 2) the classification accuracy improved substantially.

We emphasize that this reliance on closed-set matching is not a limitation of the computational design but rather a reflection of real-world environmental enforcement scenarios. In practice, regulatory and forensic investigations rarely confront an unbounded universe of candidate origins; they proceed from a nominated suspect, such as a product line, a local manufacturer, or a specific discharge point, and ask whether the questioned material is chemically consistent with it. The presented study was designed to model this specific source comparison. The goal was to provide a flexible computational method that can be adapted to local plastic pollution problems rather than to establish universal chemical fingerprints for specific plastic products. The sampled U.S. retail products and environmentally collected California beach plastics were utilized to validate the workflow and not to represent the full diversity of global plastics and their chemical additive fingerprints. The core analytical and statistical framework is compartment-agnostic and can be applied to plastics from other environmental matrices (e.g., freshwater sediments, soils, and wastewater effluent) and alternative pollution sources (e.g., industrial discharge, agricultural films, and tire wear particles), provided that sufficient local reference data are collected. Our hope is that researchers and regulatory agencies investigating local pollution hotspots can apply this workflow to their own candidate reference samples, generating locally validated, highly specific chemical fingerprints tailored to the products and weathering conditions relevant to their specific jurisdictions and pollution problems.

However, we would also like to note that the current study design did not determine whether a universal chemical fingerprint for different plastic products can be derived at all or if the lack of generalizability was simply an artifact of the present model, feature set, sample size, and validation framework that did not produce reliable predictions for previously unseen groups. Future studies should address this uncertainty directly by comparing alternative model classes, tuning procedures, feature selection strategies, and fully nested group-based validation designs.

Another important consideration for the broader application of this workflow was the transferability and reproducibility of the identified fingerprints across laboratories and instruments. Because the chemical fingerprints are defined by nontargeted features (retention time and mass-to-charge ratio) rather than confirmed chemical identities, the specific feature sets selected by the workflow may not be directly transferable to laboratories using different instrumentation, chromatographic conditions, or data processing parameters. However, the core steps of the computational workflow are instrument-agnostic and can be applied to any tabular nontargeted data set regardless of the analytical platform used to generate it. Future work to improve cross-laboratory transferability could include the use of standardized reference materials and harmonized analytical protocols, the adoption of retention index systems to enable comparison of chromatographic features across instruments, and the tentative identification of selected features to allow matching by exact chemical identity rather than instrument-specific retention times and mass-to-charge values. Collaborative interlaboratory studies that apply this workflow to shared sample sets would provide a direct assessment of fingerprint reproducibility and are an important next step toward establishing the approach as a broadly applicable tool for plastic source tracking.

A further consideration is the stability of the chemical fingerprints under environmental weathering. For this reason, the training set was constructed to include both pristine store-bought and environmentally weathered samples, capturing a broad range of chemical variability for each plastic product. This was particularly important for fingerprints established from HPLC data sets, which were overwhelmingly based on polar additives and could be susceptible to changes due to environmental weathering; including weathered plastics in the training set ensures that compounds with high leachability or low environmental stability are not selected as diagnostic features. Consistent with this, an exploratory within-item analysis of the multipoint HPLC leachates (days 0, 2, 15, and 34) indicated that the category-discriminating signal of an individual environmental item largely persisted across leaching time points. When one leaching day was held out and classified using that same item from other leaching days, the correct product category was recovered in most cases (Table S10). We interpreted this as an optimistic, within-item estimate, because each item’s other leaching days were present in training (a record-wise rather than item-wise evaluation). Therefore, it indicated that continued weathering did not fully erase an item’s discriminative fingerprint, which is a necessary but not sufficient condition for time-robust source comparison. In future studies, establishing time-robust matching of weathered environmental samples to pristine suspect sources, across rather than within items, will require substantially larger and more chemically diverse matched data sets spanning a range of weathering states.

Beside weathering, several additional factors are important for the construction of robust chemical fingerprints. One of the most important factors was the challenge in diversity and heterogeneity of environmental microplastics and their associated chemicals. This meant that a sufficient number and variety of product categories as well as items per category was essential to identify truly diagnostic features or feature patterns. As noted in Section , this method development used a minimum of 3 to 5 samples per plastic category. While this number was sufficient for the purpose of developing and testing a workflow, increasing this number in future studies would strengthen the representativeness of chemical fingerprints and improve classification robustness, particularly for product categories with high within-category chemical variability. It should also be noted that we deliberately chose samples representative of local environmental plastic occurrence rather than equal numbers of each plastic item type, which increased missing values but improved training set representativeness. Future applications with more evenly distributed data sets could benefit from replicate imputation method, rather than the imputation presented above, which would allow for a more detailed analysis of the model uncertainty. Future research could also explore alternative fusion strategies, such as performing feature selection on each individual data set before merging, decision-level fusion where separate classifiers are trained per platform and their predictions combined through ensemble voting, or block-wise weighting to balance each analytical platform’s contribution.

Overall, the presented method provides a validated proof-of-concept framework for researchers to develop their own plastic fingerprints, facilitating nontargeted environmental monitoring of plastic additives and potential pollution sources. By moving away from the pursuit of universal databases to more targeted, localized suspect-source tracking, this workflow provides a diagnostic tool to identify polluters, track contamination pathways, and hold specific polluters accountable for their environmental impact.

The data used for modeling can be found on Github page (https://github.com/huymanhnguyen0811/Microplastic-Fingerprinting) under folder “Raw Data”.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.est.6c03575.

• Details of sample processing techniques, instrumental analysis, polymer characterization data, and data processing using machine learning; hierarchal clustering analyses of plastic chemical fingerprints; random forest prediction probabilities; principal component analysis of plastic chemical fingerprints; list of plastics tested; detailed results of test case 1 and test case 2 (PDF)

• List of store-bought and environmental plastic categories investigated. Summary of missing values present in all seven data combinations across three feature reduction approaches. Descriptive statistics of retention time of five HPLC benchmark compounds.Model-independence check: regularized linear classifiers evaluated under the identical leave-one-item-out protocol as the Random Forest. Reported at the physical-object level using the Matthews correlation coefficient (MCC) (XLSX)

The authors declare no competing financial interest.