Machine learning for gasification systems: from process modeling to sustainability-oriented decision making

Gasification provides a flexible route for converting coal, biomass, plastics, and waste-derived feedstocks into syngas, hydrogen, fuels, chemicals, power, and carbon materials. Its modeling and operation remain difficult because feedstock composition, reaction regimes, multiphase transport, reactor configuration, sensor availability, and downstream integration vary simultaneously. Machine learning (ML) has consequently expanded from static performance prediction to CFD surrogate modeling, soft sensing, fault diagnosis, predictive control, digital twins, and sustainability assessment. However, the literature remains dominated by application-specific accuracy reports, and R 2 , RMSE, and MAE values are often compared without accounting for sample size, data provenance, validation strategy, domain shift, uncertainty, or deployment conditions. This review combines a broad Web of Science corpus of 951 records with a 417-record direct-topic sensitivity subset and a separately screened technical evidence layer. It reorganizes the evidence through a system-level framework linking gasification science, descriptor engineering, task–model matching, trustworthy validation, cyber–physical integration, and sustainability-oriented decision making. Representative studies demonstrate large reductions in surrogate-computation time and credible industrial optimization, but few establish closed-loop deployment. In an adjacent solar-thermochemical membrane-reactor study, a learning-assisted feedforward–feedback controller reduced the control integral error by 82.4%, increased hydrogen recovery and yield by 43% and 33%, respectively, and reduced the input preheat ratio by 30%; these results are informative for controller architecture but are not direct gasification evidence. The review proposes gasification-specific descriptor classes, a task–data–model selection matrix, evidence levels for validation and deployment, and a decision chain coupling uncertainty-aware prediction with scenario evaluation, robust Pareto optimization, risk filtering, governed execution, and feedback learning. The central conclusion is that progress depends less on marginal in-distribution accuracy than on transferable, physically consistent, uncertainty-aware, and governed ML systems that convert predictions into verifiable operating and sustainability decisions.