Beyond Permeability Prediction: Machine Learning for Gas-Separation Membranes from Data Curation to Experimentally Validated Discovery

Machine learning (ML) is transforming gas-separation membrane research, yet high retrospective accuracy does not necessarily translate into experimentally viable or process-relevant materials. This review reframes the field around the evidence chain required to convert a membrane record into a defensible discovery decision. Across polymers, metal–organic and covalent organic frameworks, mixed-matrix membranes, carbon molecular sieves, and two-dimensional systems, we critically examine data curation, physics-aligned representation, multitask and multi-fidelity learning, uncertainty quantification, inverse design, and prospective validation. We introduce a five-level evidence framework spanning retrospective interpolation, domain-disjoint generalization, physics or simulation confirmation, prospective membrane validation, and deployment-relevant testing. We further propose a contextualized observation unit that binds chemical identity to membrane state, fabrication history, feed composition, operating conditions, time, protocol, and provenance, thereby exposing leakage and structural uncertainty hidden by conventional datasets. Our analysis shows that ML has demonstrably accelerated data completion, simulation, and candidate prioritization, whereas evidence for routinely discovering manufacturable, durable, mixed-gas-performing membranes remains limited and concentrated in polymer studies. To close this gap, we develop a constrained, uncertainty-aware closed-loop architecture that treats feasibility and fabrication failures as learnable outcomes, couples acquisition to multiscale process utility, and advances candidates through staged experimental fidelities. A minimum reporting standard and measurable research priorities are provided. Together, these contributions propose “decision-grade ML” as an emerging framework for evaluating machine learning in membrane science: models should know their applicability domain, preserve physical context, propose feasible experiments, quantify uncertainty under domain shift, and demonstrate value under the conditions of intended membrane use rather than merely reproduce familiar labels or idealized proxies.