Ranking environmental degradation trends of plastic marine debris based on physical properties and molecular structure

The study addresses the global challenge of accumulating plastic marine debris and seeks to determine how many and which descriptors are necessary to model degradation behavior of ocean plastics and assess whether degradation is feasible. The introduction outlines the multifaceted environmental influences (UV radiation, wind, waves, seawater chemistry, and bacteria) that lead to cracking, surface erosion, abrasion, and fragmentation into meso-, micro-, and nanoplastics. It summarizes extensive prior data on physical (density, surface roughness, weight loss), thermal (melting temperature, glass transition temperature), and mechanical properties, along with molecular weight changes measured by spectroscopic and chromatographic methods. It describes three principal mechanisms impacting plastics in the ocean: (1) biotic processes involving bacterial colonization and biofilm formation leading to surface erosion, (2) abiotic hydrolysis of functional groups (esters, carbonates, amides) promoted by seawater alkalinity, and (3) photodegradation via UV/oxygen-induced radical processes that reduce molecular weight and facilitate biofilm formation. The authors hypothesize that combining experimental data and computational predictions can translate polymer molecular structure into predictive models of ocean degradation, and they propose to identify a hierarchy of features that regulate degradation using progressively complex analyses and machine learning.

The paper reviews prior findings on degradation mechanisms and polymer-specific susceptibilities. Biotic degradation depends on surface energy and biofilm formation, with known biodegradation of polyesters (e.g., PCL), polyamides (e.g., Nylon 6), and polyolefins (e.g., PE), with typical rates decreasing in order: polyesters > polyamides > polyolefins. Abiotic hydrolysis is facilitated by seawater pH (~8–8.3) and hydroxide ions, with activation energies indicating propensity PCL (81 kJ/mol) > PC (92 kJ/mol) > PET (125 kJ/mol). Photodegradation proceeds to ~50–100 µm depth, causing chain scission and oxidation to carbonyls that promote biofilms. Polymer structure and additives influence photostability; polymers lacking tertiary hydrogens (e.g., PMMA, PTFE) are more stable than those with tertiary or allylic C–H bonds, carbonyls, or catalyst residues. Reactivity trends of polymers with tertiary C–H bonds decrease roughly as PVC > PS > PP > PE. The review notes extensive environmental and laboratory studies characterizing polymer composition, size distributions, and changes in bulk properties in ocean environments, setting the stage for a data-driven, structure–property–degradation analysis.

The authors constructed a database of over 110 polymers (>5000 descriptors) including commercial and laboratory-made samples across classes such as polyesters (linear, branched, cyclic), polyacetals, polyamides, polyacrylamides, polycarbonates, polyethers, PE, PP, polysiloxanes, PS, PU, PVC. Each record included polymer class, specimen type (films, powders), physical attributes (mass, volume, surface-to-volume ratio), experimental parameters (time, temperature), and weight loss under seawater and abiotic or biotic conditions. Bulk descriptors: density, Mw, Mn, dispersity, Tg, Tm, percent crystallinity, and enthalpy of melting (J/g). Molecular descriptors: atom types and hybridization (sp3, sp2) and their percentages; architectural features (H per monomer, counts of CH3/CH2/CH, rings and % atoms in rings); and hydrophobicity quantified as LogP(SA)−1 derived from octanol–water partition coefficients normalized by Connolly surface area. Hydrophobicity computation used Materials Studio 2019: ALogP98 (QSAR) for LogP, Connolly surface area with 1.40 Å probe after MD-based geometry optimization (Forcite, COMPASS II forcefield, Smart algorithm; convergence thresholds: energy 1.0e-4 kcal/mol, force 0.005 kcal/mol/Å, displacement 5.0e-5 Å). Multiple oligomer models (10, 12, 14 units) were averaged. The hydrophobicity descriptor LogP(SA)−1 captures composition and connectivity, correlating with hybridization, density, and elemental makeup. Feature screening used correlation matrices to pre-select candidate predictors. From an initial set (density, molecular weight, % crystallinity, enthalpy of melting, % sp3 carbons, LogP(SA)−1), correlations reduced key features to five: molecular weight, Tg, % crystallinity, enthalpy of melting, and LogP(SA)−1. Analyses progressively compared single-feature (hydrophobicity), two-feature (e.g., hydrophobicity vs crystallinity/enthalpy) laboratory datasets, and broader comparisons under laboratory and ocean conditions using tiered categorical degradation scales. Degradation rate normalization: Surface erosion was standardized to mg cm−2 day−1 using bulk sample surface area (SAbulk), mass loss, and exposure duration. Due to heterogeneity in reporting (%, mg cm−2 day−1, BOD % day−1) and conditions (temperature, specimen form), results were also converted to 3-tier (slow, medium, fast) and 5-tier (very slow, slow, medium, fast, very fast) categories. Category thresholds: 5-tier: very slow (0–2% BOD/day; 0–0.0003 mg cm−2 day−1), slow (2–4%; 0.0003–0.003), medium (4–6%; 0.003–0.03), fast (6–8%; 0.03–0.3), very fast (>8%; >0.3). 3-tier: slow (0–4%; 0–0.003), medium (4–8%; 0.003–0.3), fast (>8%; >0.3). PBAdip was an internal medium reference. Laboratory rates at elevated temperatures were adjusted to approximate ocean conditions when necessary. Machine learning: Decision tree classifiers (scikit-learn v0.21.2, Gini impurity) were trained with manual maximum depths of 2 and 3 levels to limit overfitting, using features Mn (or Mr), Tg, enthalpy of melting, and LogP(SA)−1. Training accuracies: 72.2% (two-level, two features) and 87.1% (three-level, four features). Ten-fold stratified cross-validation yielded 57.8% and 63.2% accuracies, respectively. Misclassifications tended to occur at class boundaries or for polymers with extreme hydrophobicity (LogP(SA)−1 > ~0.010 Å−2; polyolefins) or negative LogP(SA)−1 (water-soluble polymers like PVA). For visualization, a support vector machine (RBF kernel, gamma 0.2, C 10.0) delineated shaded regions in feature plots. Modeling equations: Based on temperature dependence observed in coastal/deep-sea comparisons for PHBV, PCL, PLA, Nylon, etc., the rate of surface erosion (k) for polymers with LogP(SA)−1 > 0 and enthalpy of melting < ~85–90 J/g was modeled as: k = exp(((Twater−Tg)/LogP(SA)−1 − 28795)/4177.3), predicting slope of erosion vs temperature (mg cm−2 day−1 °C−1). A composite model estimated total erosion: Etotal = k·Twater + b + Ewaves, where Ewaves is the increment due to mechanical forces (e.g., coastal or estuary vs sheltered sites). Data requirements and caveats for these models are noted. Data handling: Missing characterization values were supplemented from databases (e.g., polymerdatabase.com) or extrapolated; densities for calculating mass were estimated via calibration curves for polyesters, nylons, PHBV. Data processing and visualization used Python (Anaconda, Python 3.7.1) with Pandas and Matplotlib; source data and code are provided as supplementary files.

• Hydrophobicity groups: LogP(SA)−1 stratifies polymers into (i) water-soluble (LogP(SA)−1 < 0; e.g., PEG, PVA), (ii) moderately hydrophobic (0 < LogP(SA)−1 < ~0.013 Å−2; susceptible to biodegradation, abiotic hydrolysis, photodegradation; e.g., nylons, many aliphatic polyesters), and (iii) highly hydrophobic (LogP(SA)−1 > ~0.015 Å−2; typically lacking hydrolyzable groups; PE, PP) which predominantly undergo very slow surface erosion and photodegradation.
• Functional groups (esters, amides, carbonates, urethanes) lower hydrophobicity relative to polyolefins and enable faster biotic and abiotic hydrolysis pathways, especially when Tg < ocean temperature. Example LogP(SA)−1: Nylon 6 ≈ 0.0045 Å−2, PCL ≈ 0.0096 Å−2, PE ≈ 0.0236 Å−2.
• Crystallinity and enthalpy of melting slow degradation. Abiotic hydrolysis of polyesters was more sensitive than biodegradation to increases in LogP(SA)−1, % crystallinity, and enthalpy of melting. Biotic processes showed relatively faster rates for more hydrophobic polyesters (e.g., PPPim, PPSub) among those with low Tm.
• Tg and molecular weight effects: Faster degradation correlates with Tg below ocean temperature; trend across polymer types: linear aliphatic polyesters (e.g., PCL) > branched (PHB/PHBV) > cyclic/heteroatom-containing backbones (PBAT, PET, PC) > cyclic all-carbon backbones (PS, poly(vinyl pyrrolidone)). Abiotic hydrolysis was fastest for Mr < ~25 kg/mol, while enzymatic activity could degrade high-MW PHB (200–700 kg/mol) if Tg ≈ 2–5 °C. Polyolefins remain slow despite low Tg due to lack of hydrolyzable groups and stabilizing additives (reported mass loss rates: PE 0.45 wt.%/month; PP 0.39 wt.%/month).
• Decision trees effectively classify degradation categories with few features. Two-feature, two-level tree (Mr and Tg) achieved 72.2% training accuracy; three-level, four-feature tree (Mr, Tg, enthalpy, LogP(SA)−1) achieved 87.1% training accuracy; 10-fold CV accuracies were 57.8% and 63.2%, respectively. The trees avoided misclassifying fast as slow and vice versa. Thresholds included Mr ≤ ~35.3 kg/mol and Tg ≤ ~−33 °C delineating faster classes. Errors clustered at class boundaries or for extreme hydrophobic or water-soluble cases.
• PET categorization varied with conditions but was better placed as slow in the three-level model, aligning with observations of ~20-year-old PET in marine environments.
• Environmental modifiers: Increasing seawater temperature (Twater) elevates erosion rates; effect magnitude depends on polymer type (PCL > PHBV coastal > PHBV deep sea > Nylon 6 > PLA) and scales with (Twater−Tg)·(LogP(SA)−1)−1·SA. Mechanical forces (waves) add measurable increments to erosion (e.g., PHBV: Ewaves ≈ 0.017 mg cm−2 day−1 in coastal water; ≈ 0.005 in estuary vs sheltered mangroves).
• Predictive equations: Eq. 2 quantifies temperature dependence of k for polymers with LogP(SA)−1 > 0 and enthalpy < ~85–90 J/g; Eq. 3 combines temperature-driven erosion with mechanical forcing. Predictions for PC, PU, and PET were consistent in magnitude with data for PCL, PHBV, Nylon 6, and PLA under stated constraints.

The findings link polymer molecular structure and bulk properties to environmental degradation behavior in seawater. Hydrophobicity (LogP(SA)−1), Tg, molecular weight, crystallinity, and enthalpy of melting form a hierarchy of predictors. Polymers with hydrolyzable functional groups and Tg below ocean temperature are more prone to biodegradation and abiotic hydrolysis, while highly hydrophobic, non-functionalized polymers (polyolefins) degrade predominantly via slower photochemical processes, often hindered by stabilizing additives. Decision tree models using 2–4 features classify fast/medium/slow degradation with reasonable accuracy and minimal severe misclassifications, highlighting the practical utility of a small set of descriptors. The applicability spans diverse environmental contexts including varying temperatures (0–30 °C), depths (1–10 m coastal to >600 m deep sea), and pressures. Similar bacterial cell counts to ~225 m depth suggest these trends are relevant to both coastal and open ocean. Variability in Tg due to structure, molecular weight, crosslinking, and plasticizers introduces noise, but category breadth mitigates the impact. Boundary misclassifications largely reflect transitions where categories merge or literature variability between commercial and lab-made materials (e.g., PET, PC). The temperature- and mechanics-informed equations provide a first-order quantitative framework to adjust predicted erosion rates for environmental conditions, though their validity is bounded by polymer crystallinity and hydrophobicity ranges.

This study develops a data-driven framework that ranks and predicts environmental degradation trends of marine plastics by integrating molecular structure, physical properties, and literature-reported degradation data. Key contributions include: (1) establishing hydrophobicity (LogP(SA)−1) as a unifying descriptor that, together with Tg, molecular weight, and crystallinity/enthalpy, predicts degradation categories; (2) demonstrating ML decision trees that classify polymers into fast, medium, and slow degradation categories with few features; and (3) proposing quantitative equations to relate surface erosion rates to seawater temperature and mechanical forcing. The approach offers practical guidance for assessing ocean plastic persistence and informs design strategies for accelerated degradation, such as incorporating hydrolyzable linkages, blending with water-soluble polymers, or using additives that promote photo-oxidation. Future work should expand the dataset—particularly for polyamides, polystyrenes, and polyurethanes—refine models to incorporate mechanical forces and weathering more comprehensively, and validate predictions with standardized, comparable field and lab studies.

• Dataset composition is uneven, with relatively fewer samples for some classes (e.g., PC, PA, PU), limiting model generalizability and contributing to cross-validation accuracies of 57.8–63.2%.
• Heterogeneous experimental conditions (temperature, specimen geometry, reporting metrics) necessitated categorical conversions and adjustments, potentially obscuring fine-grained differences.
• Hydrophobicity extremes pose challenges: very hydrophobic polyolefins (LogP(SA)−1 > ~0.010 Å−2) and water-soluble polymers (LogP(SA)−1 < 0) reduce classifier accuracy and complicate comparisons when plotting against Tg and molecular weight.
• Eq. 2 is intended for amorphous or semi-crystalline polymers with enthalpy of melting < ~85–90 J/g and LogP(SA)−1 > 0; it overestimates rates for more crystalline polyesters (e.g., PBS, PBSeb) and may not capture kinetics for highly crystalline or highly hydrophobic, non-hydrolyzable polymers.
• The simplified Eq. 3 for total erosion aggregates abiotic/biotic temperature effects and mechanical forces without fully capturing complex weathering phenomena; additional data are needed for calibration.
• Some environmental scenarios (e.g., dry-state beach cycling of microplastics) fall outside the database scope.
• Variability in Tg measurements (heating rates, plasticizers) and presence of additives in commercial materials introduce uncertainty; missing data were supplemented or extrapolated, which may add error.