Nidhi Pandya
Post-combustion CO₂ capture (PCC) using amine-based solvents is widely regarded as one of the most technologically mature near-term pathways for decarbonising existing fossil fuel power generation and heavy industry. Despite decades of industrial development, the energy penalty associated with solvent regeneration remains the primary economic barrier to largescale deployment. Machine learning (ML) has emerged as a powerful data-driven paradigm for process modelling, optimisation, and control of PCC systems. The research gap addressed by this review is the absence of a consolidated, algorithm-level synthesis of ML performance across the diverse solvent systems and process configurations studied between 2014 and 2024-a gap that hinders evidence-based selection of ML methods for future PCC research and industrial application. This review systematically analyses 72 peer-reviewed publications from 2014 to 2024 in which ML methods were applied to amine-based PCC, encompassing monoethanolamine (MEA), piperazine (PZ), AMP/PZ blends, ionic liquids, and potassium carbonate slurry systems. A meta-analysis of predictive performance across 47 modelling studies reveals that, among studies reporting comparable test-set metrics, XGBoost and hybrid mechanistic-ML approaches were frequently associated with high predictive performance (mean R² of 0.969 and 0.974 respectively), although cross-study comparability is limited by differences in dataset size, validation methodology, and process conditions. ML-guided optimisation has delivered reported energy consumption reductions of 7- 18% relative to baseline conditions across five solvent systems; however, the majority of these reductions are model-predicted rather than experimentally validated. Critical challenges including data scarcity, model interpretability, uncertainty quantification, and limited experimental validation are identified. The review concludes with a proposed ML workflow framework for PCC process development.