From Screening to Generative Design: Advances in ML-Assisted MOFs for Carbon Capture

Carbon Capture and Storage (CCS) and Direct Air Capture (DAC) technologies must improve quickly due to the escalating climate problem, which is caused by annual CO₂ emissions reaching 37 billion metric tons. Because of their remarkable surface areas and customizable pore topologies, Metal–Organic Frameworks (MOFs) have become very intriguing sorbents; yet, exploring their large chemical design space is still computationally prohibitive. The growing significance of machine learning (ML) in advancing CO₂ capture research within MOFs is methodically examined in this review. We evaluate state-of-the-art models in four important areas using a structured evaluation protocol: process-level application, mechanistic interpretability, descriptor physical relevance, and predictive performance. Recent developments in Machine Learning Interatomic Potentials (MLPs) challenge traditional rigid-lattice assumptions by showing that framework flexibility greatly affects diffusivity and adsorption thermodynamics. While physics-informed descriptor engineering achieves R² values ranging from 0.81 to 0.97 depending on gas species and pressure regime, generative techniques, such as Deep Reinforcement Learning and transformer-based topologies, enable the inverse construction of high-affinity frameworks. Crucially, the field is moving away from isolated property prediction and toward multiscale, process-integrated optimization, where machine learning models combine material characteristics with industrial performance indicators like recovery and CO₂ purity in pressure swing adsorption systems. All of these advancements point to the need for physics-informed, comprehensible designs that can connect molecular-scale discovery with experimentally reliable, water-stable materials appropriate for commercial use.