A Decade of Machine Learning for Adsorption of Emerging Contaminants: A Systematic Review of Algorithms, Performance, and Methodological Transparency

2.1 Eligibility Criteria

Throughout this review, “machine learning” (ML) denotes the class of methods under study, and “artificial intelligence” (AI) is used only where it is part of a term taken from the source literature or the screening criteria, or where it refers to the language model used as an extraction tool in Section 2.5; the two are not used interchangeably. Studies were assessed against three core pillars for inclusion: (1) ML or AI methods must be applied as a core methodology, not merely cited; (2) adsorption must be the primary removal mechanism; and (3) the target pollutant must be an emerging contaminant. Table 2 summarizes the full inclusion and exclusion criteria.

Review articles are excluded by criterion E5 as sources of primary data, but they were deliberately retained by the database filters described in Section 2.2, because the comparison of prior reviews in Table 1 and the backward citation checking both required them. Removal therefore happened downstream of retrieval rather than at the query, and Section 2.6.1 reports that this removal initially failed for two records.

2.2 Information Sources

Two electronic databases, Scopus and PubMed, were searched for articles published between January 2015 and December 2025. Results were restricted to peer-reviewed journal articles and reviews written in English. Both searches were executed on 24 February 2026, which is the cut-off date for this review: records indexed after that date were not retrieved, and no records were added subsequently. The publication window closes at 31 December 2025, so every article in the final set had been indexed for at least eight weeks before the search ran. A single dating convention is applied throughout: the year assigned to a study is the year it first appeared online, taken from the online publication date recorded in Crossref where one is available and from the date the DOI was registered where it is not. The issue label is not used. This matters for four records that a database’s indexed year would place outside the window. Three appeared online and were indexed during 2025 but carry a 2026 issue; one appeared online in August 2022 and was assigned to a 2024 issue almost two years later. Under the convention they are 2025 and 2022 literature respectively, and every study in the final set therefore falls within 2015–2025 with no exceptions. The four records and their Crossref evidence are listed in the dataset changelog accompanying the Supplementary Data. Two consequences follow and should be borne in mind when reading the temporal trends in Section 3.1. First, coverage of 2025 is likely to be marginally less complete than for earlier years, because indexing of late-2025 articles was still accruing when the search ran; the direction of the upward trend is robust to this, but the 2025 count should be read as a lower bound. Second, because the review is restricted to indexed, peer-reviewed literature, it inherits the publication bias of that literature, in which models that perform well are more likely to be reported than models that do not. The funnel plot in Section 3.5 addresses this directly.

This review was not registered in PROSPERO or another prospective register, and no protocol was published in advance; this is recorded in item 24 of the PRISMA checklist (Table S1) and is stated here because registration, while not mandatory, is the usual safeguard against post hoc changes to eligibility criteria. Two features of the workflow limit that risk. The eligibility criteria and the full Boolean query were fixed on 23 February 2026, one day before the searches were run, and neither was altered afterwards. And because every screening and extraction step was executed by scripts rather than by ad hoc judgement, the decision rules are recoverable in full from the code and the logged per-record decisions, which are available from the corresponding author on request.

2.3 Search Strategy

The search strategy employed a three-block Boolean query combining: (1) machine learning and artificial intelligence terms, (2) adsorption and sorption-related terms, and (3) emerging contaminant categories. The ML/AI block included terms such as “machine learning,” “artificial neural network,” “random forest,” “support vector machine,” “deep learning,” “gradient boosting,” and related algorithm names. The adsorption block covered “adsorption,” “biosorption,” “sorption capacity,” “isotherm,” and specific adsorbent materials. The emerging contaminant block encompassed six sub-categories: per- and polyfluoroalkyl substances (PFAS), pharmaceuticals and antibiotics, endocrine-disrupting compounds (EDCs), microplastics, antibiotic resistance genes (ARGs), and personal care products (PCPs), each with specific compound names and synonyms. The three blocks were combined using the AND operator. The full query strings as executed against each database are reproduced in SI Section S2.1.

The initial search yielded 1,337 records from Scopus and 827 from PubMed, totaling 2,164 records. After removing 419 duplicates (identified by DOI, PMID, and title matching), 1,745 unique records remained for screening.

2.4 Selection Process

The PRISMA flow diagram summarizing the study selection process is presented in Fig. 1.

2.5 Data Collection Process and Data Items

A structured data extraction framework was developed to capture over 30 fields from each included study, organized into five categories: (1) bibliographic metadata (DOI, title, authors, journal, year, country); (2) ML methodology (algorithms used, best-performing model, validation method, dataset size); (3) adsorption system (adsorbent material, target contaminant, EC category); (4) model performance (R2, RMSE, MAE); and (5) adsorption outcomes (maximum adsorption capacity qmax, removal efficiency). The complete field list and definitions are given in SI Section S2.4.

Three properties of the process should be stated before the phases are described, because each bounds what the extracted dataset can support. Data were collected by a single automated pass over each report, reconciled between two extraction methods but not duplicated by an independent second extractor. Study authors were not contacted to supply missing or ambiguous values. And no values were digitised from figures: extraction operated on text recovered from each PDF, so a quantity reported only inside a plot is systematically absent from the dataset. The verification in Section 2.6.2 measures the consequences of all three.

Full texts were extracted from the 383 candidate PDFs using the pypdf library, achieving a 92.1% good-quality text extraction rate. Data extraction was then performed in two complementary phases.

2.6 Evaluation of Automation Tool Performance

Screening and extraction were both automated, so the reliability of this review rests on how well those two tools performed on this specific body of literature. Neither the agreement statistics of Section 2.4 nor the coverage gains of Section 2.5 answer that question: agreement measures whether two rule sets concur, and coverage measures how many cells were filled, while what matters is whether the decisions were right and the values correct. We therefore evaluated both tools against re-reading of the source records. Both evaluations were carried out by one of the authors, without an independent second assessor, which is a limitation we state plainly here and again in Section 3.10.

2.7 Study Quality and Risk of Bias

No validated instrument exists for appraising risk of bias in machine-learning adsorption studies. Instruments developed for clinical prediction models, such as PROBAST, presuppose a diagnostic or prognostic outcome and a defined patient population, and they do not transfer to a regression of adsorption capacity on physicochemical descriptors. We therefore did not attempt a formal risk-of-bias appraisal of the included studies, and none should be read into Section 3.7: the eight-indicator score reported there measures what could be recovered from each study by automated extraction, which Section 2.6.2 shows is a lower bound on what the studies actually report, and it is a measure of machine-readability rather than of quality. The absence of a study-level appraisal is a limitation of this review and is recorded as one in Section 3.10.

Two related assessments were made and should not be confused with the one that was not. Risk of bias arising from missing results across the synthesis, that is publication bias, was assessed by funnel plot and Egger’s regression and is reported in Section 3.5. And the methodological property that most threatens the pooled estimates, namely that reported R2 values are almost all obtained by partitioning a single experimental dataset, was established by reading the 30 audited papers in full rather than by scoring all 202, and is reported in the same section.

Certainty in the body of evidence was not graded with GRADE or an equivalent framework. Those frameworks are built around an intervention-and-outcome question with a directional effect estimate, which this review does not pose: the pooled quantity here is a reported goodness-of-fit statistic, not a treatment effect, and its heterogeneity is reported directly instead (Section 3.5).

2.8 Effect Measures and Synthesis Methods

The effect measure carried into quantitative synthesis is the R2 reported by each study for its best-performing model. Studies were eligible for the meta-analysis if they reported both an R2 and the identity of the model that achieved it, which 110 of the 202 studies do. The remaining 92 contribute to the descriptive syntheses only. Because R2 is bounded and its sampling distribution is skewed near unity, pooling was performed on Fisher’s z-transform of R2 rather than on R2 itself, with the pooled estimate back-transformed for reporting. Studies were weighted by inverse variance, with the standard error of each z value approximated as 1/n−3 from the reported dataset size n. Dataset size is itself sparsely reported, and where a study eligible for pooling did not report one, the median dataset size of the pooled set was imputed in its place. This is an analytical choice that affects weights rather than point estimates, and it is stated here because it cannot be recovered from the reported results. Heterogeneity was quantified by Cochran’s Q and I2. Publication bias was assessed by funnel plot and Egger’s regression test.

Algorithm names were normalized to the canonical forms of Section 2.5 before pooling, so that a study reporting “GBR” and one reporting “GBDT” contribute to the same stratum. Pooled estimates are reported for the six most frequently reported best-performing algorithms.

Comparisons of R2 distributions across algorithms used the Kruskal–Wallis rank sum test with ε2 as the effect size, followed by Dunn’s post hoc test with Holm correction across all pairs and Cliff’s δ as a pairwise effect size. Two-group comparisons used the Mann–Whitney U test with Cliff’s δ, proportions were compared with Fisher’s exact test, and monotone associations were tested with Spearman’s ρ. Binomial proportions arising from the audits are reported with Wilson confidence intervals. All analyses were carried out in Python using pandas, NumPy, SciPy and scikit-posthocs. The analysis scripts, which pin the exact versions used, are available from the corresponding author on request.

3.1 Bibliometric Overview

The 202 included studies, listed individually in Table S9, were published between 2015 and 2025, with a clear upward trajectory (Fig. 2a). Only eight appeared before 2019, the earliest of them in 2015 [26], whereas the period from 2021 onward accounted for 179 of the 202 studies (88.6%). The most productive year was 2025 (n=64), followed by 2024 (n=41) and 2023 (n=32). This growth mirrors the broader expansion of ML applications in environmental engineering and suggests that the field is still in an accelerating phase.

Geographically, research output was dominated by Asian institutions (Fig. 2b; country assigned from the Scopus affiliation field of the first author). China contributed the largest share of studies (n=52, 25.7%), followed by India (n=26, 12.9%), Iran (n=24, 11.9%), and South Korea (n=16, 7.9%). Together, these four countries accounted for 58.4% of all included papers. Brazil (n=11) and Japan (n=10) were the next largest contributors. The concentration of research in a limited number of countries leaves the regions with the least treatment infrastructure largely unrepresented among the studies that define current model performance.

The 202 studies were distributed across more than 90 peer-reviewed journals, indicating a broad interdisciplinary reach (Fig. 2c). The most frequent outlets were the Journal of Environmental Chemical Engineering (n=12), Chemosphere (n=11), and the Journal of Hazardous Materials (n=11), followed by Environmental Research (n=10) and the Journal of Water Process Engineering (n=9). The top five journals collectively published 26% of all studies, while the remaining 74% were spread across journals spanning environmental science, chemical engineering, and water treatment, reflecting the multidisciplinary nature of this research domain.

To characterize the broader research landscape beyond the 202 included studies, a bibliometric network analysis was performed on the full set of 1,337 Scopus-indexed records (Fig. 3). Country nodes were resolved from the last comma-separated element of each Scopus affiliation string. The keyword co-occurrence network (Fig. 3a) reveals distinct thematic clusters: a central core around “machine learning” and “adsorption” is tightly linked to “artificial neural network,” “random forest,” and “optimization,” while satellite clusters emerge around specific contaminant classes (“tetracycline,” “antibiotics,” “pharmaceuticals,” “microplastics,” “PFAS”) and adsorbent materials (“biochar,” “activated carbon”). The temporal overlay indicates that keywords such as “XGBoost,” “PFAS,” “microplastics,” “deep learning,” and “SHAP” (shown in lighter tones) represent the most recent research frontiers, consistent with the trends observed in the included studies. In contrast, earlier keywords such as “response surface methodology” and “genetic algorithm” appear in darker tones, reflecting their established but plateauing role.

The co-authorship country network (Fig. 3b) reveals that China and the United States serve as the two dominant hubs, with the strongest bilateral collaboration link between them. India, Iran, and South Korea form secondary hubs with extensive connections to both China and Western countries. European nations (United Kingdom, Germany, France, Spain) form a densely interconnected cluster, while countries in Southeast Asia, Africa, and Latin America appear as peripheral nodes with fewer collaboration links. This pattern underscores the geographic concentration noted earlier and highlights opportunities for expanding collaborative networks to underrepresented regions.

3.2 Machine Learning Algorithms

A total of 59 unique ML algorithms were identified across the 202 studies, with 475 algorithm mentions in total (Fig. 4a; the complete inventory, with the number of studies using each algorithm, is given in Table S3). Artificial neural networks (ANN) were by far the most widely adopted, appearing in 131 studies (64.9%). Random forest (RF) was the second most popular (n=52, 25.7%), followed by response surface methodology (RSM; n=47, 23.3%), extreme gradient boosting (XGBoost; n=23, 11.4%), and support vector machines (SVM; n=21, 10.4%). Other frequently used algorithms were the adaptive neuro-fuzzy inference system (ANFIS; n=22), decision trees (n=18), gradient boosting machines (GBM; n=18), and linear regression (n=18).

Regarding algorithm multiplicity, 45.5% of studies employed a single algorithm, 53.5% used two or more (mean = 2.35, median = 2), and the remaining two studies (1.0%) have no algorithm recorded at all, an extraction gap of the kind quantified in Section 2.6.2. Multi-algorithm comparison studies typically benchmarked three to five algorithms against each other to identify the best performer for their specific adsorption system (e.g. Refs. [11, 27, 28]). Regression was the dominant task type (84.2% of studies), with 41.1% also incorporating optimization components. Classification-based approaches were relatively rare (n=14, 6.9%), suggesting that continuous prediction of adsorption capacity or removal efficiency remains the primary objective.

When evaluating reported best-performing models, ANN maintained its leading position, being selected as the top model in 88 studies (43.6%), followed by RF (n=20, 9.9%), ANFIS (n=14, 6.9%), and XGBoost (n=12, 5.9%) (Fig. 4b). Other algorithms that achieved best-model status included GBM (n=6), CatBoost (n=5), and stacking ensembles (n=3). ANN’s share as the best model (43.6%) is lower than its usage rate (64.9%), indicating that in head-to-head comparisons ensemble methods such as RF and XGBoost frequently outperform neural network approaches (e.g. Refs. [29, 30, 31]).

Temporal analysis of the top five algorithms revealed distinct adoption trajectories (Fig. 4c). ANN usage has remained consistently high since 2019, whereas ensemble methods have gained momentum since 2021. RF usage grew from one study in 2020 [32] to 18 in 2025, and XGBoost, virtually absent before 2022 [33, 34], rose to 15 studies in 2025 alone. This shift follows a broader trend in the ML community toward gradient boosting frameworks, which often achieve competitive performance with less hyperparameter tuning than deep neural networks. CatBoost and LightGBM have also appeared as alternatives in recent studies, while crystal graph convolutional neural networks (CGCNN) have so far been applied in a single study [35]. RSM usage has remained relatively stable, owing to its long-established role in experimental design optimization within chemical engineering.

Regarding software platforms, MATLAB was the most frequently reported tool (31.2% of all studies), followed by Python (19.8%) and Design-Expert (14.4%). A validation method could be recovered from the text of 45.5% of studies, cross-validation being the most common (n=31); as Section 3.5 shows, that figure reflects what the extraction could capture rather than what the studies describe.

3.3 Emerging Contaminants

The reviewed studies addressed 12 distinct emerging contaminant categories encompassing 146 unique compounds (Fig. 5a; the 50 most frequently studied compounds are listed in Table S7). Antibiotics were the most frequently studied category, appearing in 105 studies (52.0%), followed by pharmaceuticals (n=72, 35.6%), non-steroidal anti-inflammatory drugs (NSAIDs; n=38, 18.8%), and other pharmaceuticals (n=31, 15.3%). Combining the four pharmaceutical categories, antibiotics, pharmaceuticals, NSAIDs and analgesics, gives 165 studies (81.7%), which underscores the strong research focus on drug residues in water systems.

Beyond pharmaceuticals, endocrine-disrupting compounds (EDCs) were targeted in 24 studies (11.9%), per- and polyfluoroalkyl substances (PFAS) in 19 (9.4%), and microplastics in 12 (5.9%). Pesticides (n=13, 6.4%) and personal care products (PCPs; n=14, 6.9%) received comparatively less attention. Heavy metals appeared as co-contaminants alongside emerging pollutants in 22 studies (10.9%) and dye removal in 15 (7.4%), reflecting the multi-contaminant nature of real wastewater streams.

Regarding co-occurrence, 47.6% of studies focused on a single EC category, while 52.4% addressed contaminants spanning two or more categories (mean = 1.8 categories per paper). This indicates a growing recognition that real-world water matrices contain mixtures of emerging pollutants, necessitating ML models capable of handling multi-contaminant scenarios.

At the individual compound level, tetracycline was the most frequently modeled contaminant (n=36, 17.8%), followed by ciprofloxacin (n=32, 15.8%), ibuprofen and diclofenac (each n=16, 7.9%), and carbamazepine (n=15, 7.4%) (Fig. 5b). Among non-pharmaceutical compounds, bisphenol A (n=13, 6.4%) was the most studied EDC, while PFOA (n=7) and PFOS (n=5) were the most targeted PFAS compounds. The dominance of a small number of model compounds, with the top 10 contaminants appearing in over 80% of studies, suggests that future research should broaden the scope to include less-studied but environmentally relevant pollutants.

Temporal analysis of EC category trends (Fig. 5c) revealed that antibiotic adsorption modeling has grown consistently from 4 studies in 2019 to 30 in 2025, with pharmaceutical studies following a similar trajectory. The most striking temporal trend was the emergence of PFAS-related ML studies, absent before 2021 but rising to 10 studies by 2025, which reflects heightened global regulatory attention to “forever chemicals.” Microplastic studies appeared only from 2023 onward (n=12 in total), and EDC studies showed steady but modest growth. These trends collectively suggest that while the field remains anchored in pharmaceutical contaminants, it is diversifying toward PFAS and microplastics in response to evolving environmental priorities.

3.4 Adsorbent Materials

Adsorbent information was available for 201 of the 202 studies (99.5%), spanning 18 distinct material categories (Fig. 6a). Biochar was the most frequently studied adsorbent, appearing in 43 studies (21.3%), reflecting its low cost, tunable surface chemistry, and growing availability from agricultural and industrial waste streams. Activated carbon ranked second (n=30, 14.9%), followed by carbon nanotubes (n=14, 6.9%), microplastics as adsorbent substrates (n=9, 4.5%), graphene oxide (n=8, 4.0%), and nanocomposites (n=8, 4.0%). Advanced materials such as metal–organic frameworks (MOFs; n=8, 4.0%), iron oxide and magnetic particles (n=6, 3.0%), and layered double hydroxides (LDHs; n=4, 2.0%) constituted a smaller but growing share. Traditional adsorbents including clay and bentonite (n=5), chitosan (n=5), and zeolite (n=2) were comparatively underrepresented, likely because their well-characterized adsorption behavior leaves less scope for ML-driven discovery. A further 33 studies employed adsorbents classified as “other,” encompassing diverse or less-common materials.

Carbon-based materials collectively dominated the dataset. The four carbon categories account for 94 studies (46.5%) when each study is assigned to a single category, as in the counts above, and for 104 studies (51.5%) if studies combining a carbon material with another class are also counted; thirteen studies carry more than one adsorbent category and are the difference between the two figures. This dominance aligns with their high surface area, chemical modifiability, and established effectiveness for organic micropollutant removal. The remaining studies featured diverse materials spanning biosorbents, agricultural waste, polymers and resins, silica, and soil or sediment, indicating that ML modeling is being applied across the full spectrum of adsorbent technologies.

Temporal analysis revealed notable shifts in material preferences (Fig. 6b). Biochar usage grew sharply from 2022 onward, with 12 studies in 2024 and 13 in 2025, together 25 of the 43 biochar papers (58.1%). Activated carbon showed a similar recent surge, with 13 studies in 2025. The most striking trends were the emergence of MOF-based studies, which appeared only from 2023 [36] and reached 5 studies in 2025, and microplastic adsorption studies, which followed a parallel trajectory from 2023 through 2025. These trends mirror the broader materials science literature, where MOFs and microplastic–contaminant interactions are active research frontiers.

The cross-tabulation of adsorbent categories with EC categories (Fig. 6c) revealed clear material–contaminant pairing patterns. Biochar was predominantly applied to antibiotic adsorption (12 studies) and pharmaceutical removal (6 studies), consistent with its proven affinity for aromatic organic compounds. Activated carbon showed a similar antibiotic focus (7 studies) but was more evenly distributed across contaminant classes, including PFAS, EDCs, and heavy metals. Carbon nanotubes were applied exclusively to antibiotics and pharmaceuticals. PFAS removal studies showed no strong preference for a single adsorbent type, with papers distributed across biochar, activated carbon, MOFs, LDH, graphene oxide, polymer and resin materials, soil and sediment, and other sorbents, which reflects ongoing efforts to identify optimal sorbents for these recalcitrant compounds. EDC studies were concentrated in the biochar and graphene oxide categories, while pesticide adsorption was primarily modeled using activated carbon and biochar.

To provide a quantitative perspective on adsorption performance across material types, the reported maximum adsorption capacity (qmax) values were compared across adsorbent and contaminant categories (Fig. 7). Among the 86 studies reporting qmax (Table S6), values spanned over four orders of magnitude, from 0.69 mg/g for heavy metal uptake on low-cost biosorbents to 14,499 mg/g for high-capacity engineered materials, with an overall median of 115 mg/g. When stratified by adsorbent category (Fig. 7a), MOFs and nanocomposites exhibited the highest median capacities, consistent with their high surface areas and designed pore architectures, while traditional materials such as clay/bentonite and iron oxide/magnetic particles showed lower but more consistent values. Biochar and activated carbon displayed wide ranges reflecting the substantial variability in feedstock, preparation conditions, and target contaminants. When grouped by EC category (Fig. 7b), antibiotic and pharmaceutical studies reported the highest median qmax values, whereas PFAS studies, despite employing high-performance adsorbents, showed lower capacities, likely reflecting the lower initial concentrations and higher recalcitrance of fluorinated compounds. These cross-categorical comparisons highlight that ML models in this field must accommodate adsorption capacities varying by several orders of magnitude, underscoring the importance of appropriate data normalization and log-transformation in model training.

3.5 Model Performance

Model performance was assessed primarily through the coefficient of determination (R2), which was reported in 110 of the 202 studies (54.5%). Other metrics were reported less frequently: RMSE in 52 studies (25.5%), removal efficiency in 116 studies (56.9%), maximum adsorption capacity (qmax) in 86 studies (42.2%), and MAE in only 16 studies (7.8%) [37, 38, 39]. The inconsistent reporting of performance metrics across studies represents a significant barrier to systematic comparison and meta-analysis.

Among the 110 studies reporting an R2 value, model performance was generally high (Fig. 8a). The median R2 was 0.984 (mean = 0.969): 86 of the 110 studies (78.2%) reached R2≥0.95, 46 (41.8%) reached R2≥0.99, and only 10 (9.1%) reported a value below 0.90. These values indicate strong reported predictive capability across the literature, but such uniformly high figures may partly reflect publication bias toward favorable outcomes and the tendency to report only the best-performing model configuration.

Comparison of R2 values across the five most frequently reported best-performing models (Fig. 8b) revealed that ANFIS achieved the highest median R2 (0.997, n=8), followed by ANN (median = 0.991, n=48) and GBM (median = 0.990, n=6). XGBoost showed slightly lower but competitive performance (median = 0.970, n=9), while RF exhibited the most variable results (median = 0.947, n=11) with the widest interquartile range. A Kruskal–Wallis rank sum test showed that these differences were statistically significant (χ2=19.39, df =4, p=6.6×10−4, ε2=0.20; Table S4), although the unequal sample sizes and potential confounding by study-specific factors (dataset size, contaminant type, adsorbent material) warrant cautious interpretation. A global test of this kind establishes only that the five distributions are not interchangeable. It does not identify which algorithms differ. We therefore added Dunn’s post hoc test with Holm correction across all ten pairs, reported in full with Cliff’s δ effect sizes in Table S10. Two comparisons survive correction: ANN against RF (p=0.005, δ=0.64) and ANFIS against RF (p=0.012, δ=−0.77). The remaining eight pairs, including ANN against ANFIS and ANN against XGBoost, are not distinguishable at the 5% level. The significant global result is thus driven almost entirely by RF sitting below the others, and the apparent ordering among ANFIS, ANN, GBM and XGBoost in Fig. 8b is not statistically supported. The sample sizes bound this conclusion as well: GBM (n=6) and ANFIS (n=8) contribute few studies, so the power to detect anything but a large difference involving them is low, and a non-significant pair here indicates insufficient data rather than equivalence. Less common algorithms also performed strongly in individual studies: CatBoost achieved top results in multi-algorithm benchmarks (e.g. Refs. [11, 27, 40]), while Cubist [41], LS-SVM [42, 43], and gene expression programming (GEP) [44, 45] were reported as best-performing in specific adsorption systems.

Dataset size was reported in 78 studies (38.2%), revealing a concerning distribution (Fig. 8c). Nearly 40% of those studies (n=31) worked with fewer than 50 data points, and the median dataset size was only 131 samples. Studies using 200 or more samples represented 38.5% of those reporting dataset size, while only 15.4% exceeded 1,000 samples (e.g. Refs. [46, 11, 30]). The prevalence of small datasets raises questions about model generalizability and the risk of overfitting, particularly for complex models such as deep neural networks that require substantially more training data [47, 48]. The reported removal efficiencies were consistently high (median = 95.3%, n=116), and qmax values spanned a wide range from 0.69 to 14,499 mg/g (median = 115 mg/g, n=86), reflecting the diversity of adsorbent–contaminant systems studied.

Small training sets are the most frequently voiced concern about this literature, and the 202 studies allow it to be examined rather than assumed. The relationship between R2 and dataset size was tested on the 51 studies that report both quantities, 46% of the 110 with a usable R2. The analysis therefore rests on under half the performance data, and every statement below is conditional on that subset (Fig. 9). The result is not the one usually expected. The rank correlation between dataset size and reported R2 is weak and not significant (Spearman ρ=−0.147, p=0.304), and a direct comparison of the two strata contradicts the inflation hypothesis outright: studies with fewer than 50 samples report a median R2 of 0.964, slightly below the 0.977 of studies with 50 or more (Mann–Whitney p=0.70, Cliff’s δ=−0.07, negligible). Nor are implausibly perfect fits concentrated among the small studies: R2 exceeds 0.99 in 3 of 16 small-sample studies (19%) and 6 of 35 larger ones (17%), a difference indistinguishable from chance (Fisher exact p=1.0). Dispersion runs the same way, with the smallest stratum no tighter than the largest (SD 0.044 against 0.044). On the evidence available, then, reported performance in this literature does not degrade with sample size, and we can find no statistical signature of small-sample overfitting in the values that authors report.

The correct inference from this is not reassurance. It is that reported R2 is not a quantity in which overfitting would be visible, because in most of these studies nothing was held back that could reveal it. This is where the extraction audit of Section 2.6.2 changes what can responsibly be claimed. The extracted validation-method field is populated for only 92 of 202 studies (46%), which invites the conclusion that half the literature reports no validation at all; the audit shows that conclusion would be wrong. Reading 30 randomly selected papers in full, 27 (90%, 95% CI 74–97%) do state a validation scheme, so most of the empty cells are extraction failures rather than silent authors. What those 27 schemes have in common is the substantive point. Every one of them is an internal partition of the same experimental dataset: a random train/test or train/validation/test split, k-fold cross-validation, or both. Exactly one study describes any check against separately collected data, and that check comprises four additional experimental conditions run in duplicate. None validates against an independent dataset from another laboratory or study. The label is unreliable as well: of the four studies the extracted dataset records as performing “external validation,” the one that fell into the audit sample turns out to mean a held-out test set from a 7:3 split of its own data. A model evaluated on data it was fitted to will report a high R2 whatever its sample size, and that is what the flat relationship in Fig. 9 most plausibly reflects: not an absence of overfitting, but an absence of the reporting that would detect it. The same reasoning limits what the algorithm comparison above can mean, since the same undisclosed validation practices underlie every R2 entering the Kruskal–Wallis and Dunn tests, and they need not be uniform across algorithm families. The stratified analysis makes the point concretely. Splitting the 51 studies by sample size leaves 9 and 23 studies respectively in the five-algorithm subset, several algorithms represented by a single study, and no interpretable test of whether algorithm ranking depends on dataset size; the honest answer to whether small datasets distort the comparison between algorithms is that these studies cannot say. What can be said is that 77% of reported R2 values exceed 0.95, and that essentially all of them were obtained by splitting a single experimental dataset rather than by confronting a model with data gathered independently. That, and not sample size in itself, is the substantive threat to the reliability of the pooled estimates reported here, and it is the practice this review would most like to see change. This underscores the urgent need for standardized external validation, reporting of confidence intervals, and the use of held-out test sets that are independent of the training data.

3.5.1 Meta-analysis of algorithm performance

To move beyond descriptive comparison and provide a statistically rigorous cross-study synthesis, we pooled the reported R2 values by the method set out in Section 2.8, across the 110 studies reporting both an R2 and the identity of the model that achieved it (Fig. 10a). Among the six most frequently reported best-performing algorithms, ANFIS achieved the highest pooled R2 (0.998, 95% CI [0.998,0.998], n=8), followed by ANN (0.995[0.995,0.995], n=50), XGBoost (0.979[0.978,0.980], n=9), GBM (0.978[0.977,0.978], n=7), RF (0.955[0.951,0.958], n=11), and CatBoost (0.929[0.926,0.932], n=5). However, heterogeneity was extremely high across all algorithms (I2>98%, Cochran’s Qp<0.001 for all), indicating that the substantial between-study variability (driven by differences in adsorbent–contaminant systems, dataset characteristics, and feature engineering) precludes a simple “universal best algorithm” conclusion. Notably, the apparently lower pooled R2 for CatBoost likely reflects its preferential application to more challenging multi-contaminant or multi-adsorbent datasets [11, 27, 40], rather than inferior algorithmic capability.

A funnel plot analysis was performed to assess potential publication bias (Fig. 10b). Visual inspection revealed notable asymmetry, with a concentration of studies reporting high R2 values (>0.95) regardless of sample size, and a conspicuous absence of studies in the lower-left region (low precision, low R2). While Egger’s regression test did not reach statistical significance (p=0.97), the visual pattern strongly suggests that studies reporting low predictive performance are either not submitted or not published, consistent with publication bias. The funnel plot further reveals that small-sample studies (high standard error) cluster near the upper boundary of R2=1.0, reinforcing the overfitting concern identified earlier. These meta-analytic findings collectively argue that the field’s reported predictive performance is likely inflated, and that future studies should pre-register analysis plans, report negative results, and provide external validation on independent datasets.

Subgroup analysis by adsorbent and contaminant categories (Fig. 11) revealed that model performance varies substantially across adsorption systems. By contaminant category, pesticide and antibiotic studies achieved the highest median R2 (≥0.98), likely reflecting the more extensive datasets available for these well-studied compound classes. PFAS studies, despite employing advanced algorithms (XGBoost, CatBoost), showed slightly lower and more variable performance (median R2≈0.96), consistent with the greater chemical diversity and lower concentrations characteristic of PFAS mixtures. By adsorbent category, carbon nanotube-based studies reported the highest median R2, followed by activated carbon and biochar, while MOF-based studies, being relatively new, showed wider performance variability.

3.6 Feature Importance Synthesis Across Studies

While most ML models in the adsorption literature function as “black boxes,” a growing subset of studies (n=15, 7.4%) have employed model-agnostic interpretability tools, primarily SHAP (SHapley Additive exPlanations), to identify the key drivers of predicted adsorption performance. To date, these feature importance analyses have been reported in isolation, with each study interpreting its SHAP results within its own adsorbent–contaminant context. Here, we synthesize feature importance rankings across those 15 studies to identify universally important predictors and evaluate their alignment with classical adsorption theory (Fig. 12).

Across the 15 SHAP-enabled studies, initial contaminant concentration and BET surface area emerged as the two most universally important features, each appearing as a top-ranked predictor in all 15 studies with mean relative importance scores of 0.81 and 0.81, respectively (Fig. 13a). This finding is directly consistent with classical adsorption theory: the Langmuir and Freundlich isotherms both predict strong dependence on equilibrium concentration (Ce), while maximum adsorption capacity (qmax) scales with available surface area. Solution pH ranked third (mean importance = 0.56, 15/15 studies), reflecting its well-established role in governing adsorbent surface charge (relative to pH_pzc) and contaminant speciation. Adsorbent dosage (0.55, 12/15 studies), pore volume (0.47, 12/15 studies), and temperature (0.31, 14/15 studies) constituted the remaining universally important features.

Beyond universal features, the synthesis revealed system-specific predictors that carry mechanistic information (Fig. 13). For biochar-based adsorption, elemental ratios (H/C, O/C, O+N/C) and pyrolysis temperature emerged as important predictors: features not captured by classical isotherm models but directly related to surface aromaticity and functional group density. For PFAS adsorption, contaminant molecular weight and perfluoroalkyl chain length were highly ranked [40, 27, 33], reflecting the dominance of hydrophobic interactions and size-dependent partitioning. For soil/sediment systems, organic matter content and clay content outranked surface area [49], consistent with the known role of soil organic matter in contaminant sorption. When categorized by feature type, contaminant properties (molecular weight, log Kow) showed the highest mean importance among system-specific features, suggesting that future ML models should systematically incorporate molecular descriptors alongside traditional physicochemical parameters.

The strong alignment between ML-derived feature importance and classical adsorption theory validates the physical plausibility of these data-driven models. However, the synthesis also highlights that ML can extract insights beyond classical frameworks (particularly the importance of elemental ratios and molecular descriptors), suggesting a complementary role for explainable ML in hypothesis generation for adsorption mechanism research.

3.7 Machine-Extractable Reporting

To characterize how readily methodological detail can be recovered from this literature at scale, each study was scored on eight binary indicators (Table S5): whether a value could be extracted for (1) R2, (2) RMSE, (3) MAE, (4) dataset size, (5) validation methodology, (6) number of input features, (7) software or tool used, and (8) multi-algorithm comparison (Fig. 14). The score measures what the extraction recovered, which the audit in Section 2.6.2 shows is not the same as what the studies contain: populated cells were 93.3% accurate, but 35 of 90 empty numeric and methodological cells in the audited sample should have carried a value. Each rate below is therefore a lower bound on reporting. The mean score across the 202 studies was 2.97 of 8.

Recovery rates differed sharply between indicators (Fig. 14a). Software disclosure was the most readily extracted (61.9%), followed by R2, recovered for 110 of 202 studies (54.5%), and multi-algorithm comparison (108 studies, 53.5%). Methodological detail was far less accessible: validation methodology was recovered for 45.5% of studies, dataset size for 38.6%, RMSE for 25.7%, number of input features for 9.4%, and MAE for 7.9%. The gradient itself is the finding. Fields that a study reports in a fixed, labelled place, such as software and headline R2, are recoverable; those reported in prose, in figure captions, or not at all are largely lost. Validation methodology is the clearest case: the audit found that 90% of studies do describe a scheme, against the 45.5% recovered here. Whatever these studies contain, it cannot at present be read out of them at the scale a synthesis requires.

The score improved modestly over time (Fig. 14b; Table S8; Spearman ρ=0.202, p=0.004, n=202). Restricted to the 2019–2025 window shown in Fig. 14b, where all but eight of the studies lie, the same test gives ρ=0.157 and p=0.029. The trend is positive and reaches conventional significance in both windows, but the effect is small: a rank correlation of this size accounts for well under a tenth of the variance in the score, and the practical gain over six years is less than one indicator of eight. The mean score rose from 2.69 in 2019–2021 to 3.30 in 2024–2025, driven by the adoption of multi-algorithm comparison and more consistent R2 reporting. Even in 2025 it remains below 4 of 8. Standardized placement, rather than more disclosure alone, is what would raise it, and that is what the Minimum Reporting Checklist for Machine Learning in Adsorption Studies (MRCAS) proposed here is designed to supply: a 22-item checklist organized into five categories covering data description, model specification, validation protocol, performance metrics, and interpretability or reproducibility (Table S15). Analogous to the PRISMA guidelines for systematic reviews and the TRIPOD statement for clinical prediction models, MRCAS gives each item a fixed reporting slot, so that a future synthesis can recover it without reading every paper in full.

3.8 Explainable AI and the Distance to Deployment

The preceding sections establish which algorithms are used and how well they are reported to perform. Neither question is the one a utility engineer would ask, which is whether any of these models could be relied upon outside the laboratory that produced it. Two conditions bear on that: the model must be interpretable enough to be trusted when it is wrong, and it must have been shown to work on water resembling the water it would treat. The reviewed studies speak to both, and in each case the answer is that the field is earlier in its development than the reported R2 values suggest.

Interpretability is advancing but remains a minority practice. Fifteen studies (7.4%) apply SHAP, and their cross-study synthesis in Section 3.6 is encouraging: the features that emerge as most influential, initial contaminant concentration and BET surface area, are the ones classical adsorption theory would nominate, which is a meaningful, if modest, validation that these models are learning adsorption rather than the idiosyncrasies of a particular experiment. Its limits should be equally clear. A consensus drawn from 15 studies is a consensus among the small subset of authors who chose to look, and those authors were disproportionately working with tree ensembles, for which SHAP is computationally convenient. Whether the same features dominate in the 48 studies whose best model was a neural network is not established by these studies, and the agreement with theory does not by itself license using any of these models predictively.

Deployment is the larger gap. The water matrix was recorded for 132 studies (65%), and among these at least 46 (23% of all studies) report working with a real or non-synthetic matrix (wastewater, groundwater, drinking water, surface water, urine or tap water), while 86 (42%) report synthetic solutions alone. These counts should be read as lower bounds on real-matrix work, given the extraction audit’s finding that empty cells substantially reflect extraction failure rather than silence. Even read generously, the picture is one in which most models are trained and tested on single-contaminant solutions in deionised water, whereas the competitive adsorption, natural organic matter and variable ionic strength of a real influent are precisely what would degrade their predictions. That concern compounds with the validation practice documented in Section 3.5: a model fitted and evaluated on random splits of one synthetic dataset has not been tested against any of the conditions that deployment would impose on it.

Three further reporting deficits stand between this literature and reuse of its models by others. Software is identified in 125 studies (61.9%), with MATLAB the most common environment and Python-based stacks a growing minority, but the extracted dataset contains no field for code or data availability, so the number of studies releasing either cannot be stated here. Establishing it would require a separate full-text pass. Only 11 studies (5.4%) report both a dataset size and a number of input features, so the sample-to-feature ratio, the single most direct indicator of whether a model had enough data to support its complexity, is computable for one study in twenty. Among those 11 the median ratio is 47:1, with two studies below 10:1. And no reviewed study reports a prospective test, in which a model trained on one adsorbent–contaminant system predicts an experiment not yet performed. Until such tests become routine, the performance figures synthesised here should be read as descriptions of model fit, not as evidence of predictive capability in service.

3.9 Research Gaps and Future Directions

The synthesis of the 202 reviewed studies reveals several critical gaps that should guide future research in ML-assisted adsorption of emerging contaminants (Fig. 15).

3.10 Limitations of This Review

Five limitations bound what can be concluded from this synthesis, and one of them we regard as material.