In recent years, machine learning (ML) has been increasingly applied to optimize and predict the properties of thin-film nanocomposite (TFN) membranes. However, no existing review has systematically synthesized the ML models employed, the material and performance properties targeted, limiting the ability to identify dominant trends across studies. To address this gap, a scoping review was conducted following Tranfield’s approach and reported in accordance with PRISMA guidelines. Searches were performed in five databases, including PubMed, Web of Science, and Scopus, yielding 299 records published mostly between 2018 and 2026, with 32 meeting the inclusion criteria, mainly focusing on either predicting or optimizing TFN membranes. According to the results, transport-related performance metrics, namely flux, permeability, selectivity, and rejection, dominated the literature, with flux/permeability (12/32, 37.50%) and selectivity/rejection (9/32, 28.13%) being the most frequently predicted outcomes. Artificial Neural Networks (ANNs) were the most employed models (16/32, 50.00%), followed by ensemble learning methods (15/32, 46.88%). Based on within-study comparisons (studies implementing multiple machine-learning models (n = 24)), ANN-based approaches and ensemble learning methods were the most frequently reported best-performing methods, while ML models combined with optimization techniques (e.g., particle swarm optimization) were reported as best performing in 33.33% (8/24) of the studies. Despite promising progress, methodological limitations remain. Most studies relied on internal validation (train–test splits or cross-validation), while only 4/32 used external validation (EV). Without EV, it is difficult to assess performance under distributional shifts. Future studies should prioritize external validation as part of the modeling pipeline. • This study provides the first scoping review synthesising machine-learning applications in thin-film nanocomposite membrane research. • The Tranfield’s methodology was applied to systematically map machine-learning models and target properties across 299 studies published mostly between 2018 and 2026. • Results indicate that artificial neural networks and ensemble learning methods dominate machine-learning applications and are most frequently reported as best performing in comparative studies. • The review identifies limited external validation and inconsistent evaluation practices as key methodological gaps affecting model generalisability.
Ajala et al. (Sun,) studied this question.