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

1. Physics-Informed Descriptor Engineering

1.1. Descriptor Engineering Strategies

An important step toward understanding the intricate potential energy surfaces (PES) of porous materials is the incorporation of spatially aware energy descriptors. Researchers have obtained R²>0.97 for nitrogen and R²>0.87 for carbon dioxide isotherms by adding surface energy histograms and radial distribution functions (RDFs) to conventional geometric features and Henry's constants. By giving the ML model a physically valid representation of surface form, this method specifically addresses the "intermediate pressure bottleneck"—the region where both binding energy and spatial heterogeneity govern uptake.

Importantly, using XGBoost trained on more than 10,000 structures shows that the spatial decay of interaction sites (recorded via RDF) is the main differentiator for capacity in frameworks with comparable energy levels, whereas affinity distribution (f(E)) establishes the energetic baseline. The model's poor performance in the CO₂ chemisorption regime, however, emphasizes the single-charge probes' present limitations. Future physics-informed machine learning approaches must use dipole and quadrupole probes to account for the orientation-dependent packing and electrostatic multipoles of CO₂, bridging the gap between simplified physisorption models and practical selective separation, in order to achieve industrial-grade reliability.[8]

By moving beyond standalone pore or doping engineering, researchers have identified a critical micropore-dopant coupling mode that determines CO₂ transport in carbon-based adsorbents. Using a Random Forest architecture trained on multi-scale simulation data (R²=0.934), this study introduces Free volume (Vf) as a primary descriptor (accounting for 25% relative importance) to quantify the steric effects induced by surface functionalization. This approach reveals a significant thermodynamic shift: while basic dopants (e.g., NH₂) utilize Lewis’s acid-base interactions to optimize adsorption at 7 Å, physisorption-dominant dopants (e.g., oxygen groups) occupy available nano space, necessitating an enlarged optimal pore size of 8–10 Å to maximize capacity. These results were validated experimentally, as adsorbents designed with this particular coupling mode attained a leading-level capacity of 4 mmol/g, a 130% improvement over unoptimized frameworks. These findings highlight the need for physics-based descriptor engineering to settle long-standing disputes over the promoting vs. inhibitory effects of heteroatom doping in porous carbons. [9]

Compared to conventional structural models, the incorporation of advanced descriptor engineering—more precisely, the insertion of more than 700 "calculated" molecular features—has been demonstrated to increase CO₂ adsorption prediction accuracy by 15%–20%. The difficulties in capturing sparse gas-solid interactions are highlighted by the XGBoost framework's performance sensitivity at low pressures (0.01 bar), even though it achieves a high coefficient of determination (R²>0.94) at industrial pressures (2.5 bar). Importantly, the application of SHAP interpretability reveals a basic thermodynamic transition: as pressure rises, charge distribution and electronegativity features (representing Coulomb forces) gradually replace the predictive weight of atomic mass and number (representing van der Waals forces) at low pressures. However, there is a risk of overestimating real-world performance due to the use of a hypothetical dataset (hMOF) and simulated GCMC labels; in the absence of experimental validation, the model's capacity to account for structural defects or competitive adsorption in complex flue gases continues to be a crucial gap for industrial scalability.[10]

1.2. Hybrid Textural-Optimization and Outlier-Aware Predictive Modeling

One significant step toward overcoming the "black-box" constraints of conventional adsorption modeling is the transition from solo machine learning to hybridized optimization frameworks. Researchers successfully neutralized the hyperparameter tuning mistakes that cause underestimation in solo models at high uptake regimes by combining the Growth Optimization (GO) method with Least Square Support Vector Machines (LSSVM), achieving an astounding R² of 0.9798. SBET is the main structural driver of capture, outperforming pore volume and Langmuir surface area in prediction sensitivity, according to a database-driven analysis of 475 experimental points.

Crucially, Williams plot analysis provides a level of statistical robustness necessary for quick screening by confirming that 94.95% of the experimental data fits into the model's applicable domain. The current model's reliance on experimental textural measurements suggests a continuous barrier in purely computational discovery processes. Future physics-informed machine learning approaches must close the gap between these high-accuracy experimental models and high-throughput theoretical proxies to improve industrial scalability. This will ensure the utilization of the "metal-affinity" and "pore-filling" dynamics discovered here in the de novo design of next-generation carbon capture materials.[11]

1.3. Ensemble Interaction Mapping and Node-Affinity Hierarchies

A major advancement in closing the accuracy gap between individual machine learning models and intricate experimental adsorption data is the implementation of unique stacking ensemble structures. Researchers were able to attain a benchmark-surpassing R² of 0.9833 for CO₂ uptake across 1,212 experimental data points by combining the predictive powers of tree-based, kernel-based, and neural network learners. By using ablation studies and permutation importance to separate the impact of thermodynamics from framework architecture, this study goes beyond straightforward performance reporting to offer a multi-criteria interpretability framework.

Crucially, partial dependence plots that measure the stepwise shift in CO₂ binding efficacy support the analysis's identification of a "metal-affinity hierarchy," where copper and magnesium centres clearly outperform conventional zinc-based nodes in terms of productivity. The 70% prevalence of zinc-based items in existing experimental archives, however, draws attention to a recurring problem with data imbalances that puts well-studied materials in danger of model overfitting. To ensure that the discovery of high-performance MOFs extends into the underrepresented chemical subspaces of rare or complicated metal centres, future physics-informed ML strategies must include stratified sampling and synthetic oversampling (e.g., SMOTE) to improve commercial scalability.[12]

The fundamental language for quantitatively representing MOF structures is established via physics-informed descriptor engineering, but large-scale predictive procedures are where these capabilities truly shine. Supervised learning methods can quickly predict adsorption parameters over large chemical spaces once chemically meaningful descriptors are defined. The first significant acceleration point in ML-driven CO₂ remediation research is this transition from feature construction to high-throughput prediction.

2. High-Throughput Screening and Universal Property Prediction

2.1. High-Throughput Screening and Discovery

The move to hybrid physics-ML models makes it possible to screen composite materials more effectively than with conventional molecular simulations. Researchers demonstrated that transfer learning is a reliable method for predicting "unseen" materials like 6FDA-DAM based on known polymer features by training a stacked ensemble regression model on over 54,000 hybrid membranes and achieving an outstanding forecast accuracy of R² = 0.96. Importantly, big data mining of this dataset revealed that the MOF filler becomes the bottleneck for elite performance instead of the polymer matrix; in top-performing membranes, PLD and LCD importance spikes to over 30%, indicating that once a high-permeability baseline is established, pore-size engineering becomes more important than polymer selection.

Constructive evaluation, however, indicates a reliance on the ideal interface assumption of the Maxwell model. Interfacial gaps and polymer chain rigidity frequently undermine the theoretical "Robeson limit" performance in practical industrial scalability. To guarantee that high-throughput screening yields materials that maintain their separation efficiency under the mechanical stresses of flue gas streams, forthcoming physics-informed machine learning initiatives should integrate interfacial morphology descriptors to reflect the complex bonding between organic linkers and polyimide matrices.[13]

The shift from simulation-heavy datasets to experimental-based machine learning models provides a more rigorous benchmark for actual CO₂ capture. Researchers have shown that ensemble learning can traverse the diverse terrain of 236 distinct experimental MOFs with a 15% RMSE improvement over conventional gradient-boosting frameworks by employing the CATBoost architecture. The integration of SHAP interpretability reveals a deeper layer of chemical control: localized atomic environments and electronic state distributions (e.g., PEOE_VSA7) are the true differentiators for optimizing adsorption, even though the model confirms that pressure and surface area remain the dominant thermodynamic drivers.

Nevertheless, this study also functions as a critical assessment of the limits of generalization. The discrepancy between the model's validation score (R2 = 0.84) and training performance (R2 = 0.99) highlights a recurring problem with overfitting when training sophisticated models on sparse experimental data. In order to guarantee that high-throughput screening of hypothetical frameworks can consistently transfer into industrially scalable, high-stability sorbents, future efforts in physics-informed machine learning must shift toward dataset diversification, embracing a larger spectrum of chemical functions.[14]

2.2. Universal Property Prediction and Isotherm Generalization

An important development in zeolite-based carbon capture is the shift away from material-specific empirical models and toward unified experimental datasets. Researchers have developed a "universal" predicting capability that greatly surpasses conventional Langmuir isotherm fittings in both accuracy and generalization by training a Gradient Boosted Trees (GBT) model using more than 5,700 experimental data sets. By identifying the Si/Al ratio and cation composition as crucial chemical factors, this method enables the model to generalize over a variety of frameworks (such as FAU, ZSM-5, and 13X) and high operating environments up to 45 bar.

Crucially, the model's resilience is confirmed by external validation on unseen datasets, which successfully resolves the overfitting issues typical of datasets derived from literature. A constructive criticism, however, shows that the dataset is still biased toward low-uptake regimes (<2 mmol/g) and that 32% of the training data did not report surface area. Future physics-informed machine learning approaches should prioritize the inclusion of high-uptake experimental benchmarks and uniform reporting of structural parameters to further optimize model sensitivity in peak performance regimes in order to improve industrial scalability.[15]

2.3. Property-Driven ML Applications

In this study, two particular properties—CO₂ working capacity and CO₂/N₂ selectivity—are predicted using machine learning. In order to quickly screen Metal-Organic Frameworks and achieve high predictive accuracy (R²) for carbon capture metrics, Artificial Neural Networks (ANNs) offer a computationally efficient substitute for Graph Neural Networks. This model quantifies a crucial design insight by combining industrial, field, and simulation data: CO₂ working capacity is primarily determined by pore size and surface area, exhibiting a weak negative association with greater chemical complexity. The CO₂ capacity predictions are reliable (MAE = 0.8 mmol/g), but the CO₂/N₂ selectivity predictions show significant dispersion (MAE = 25), indicating that intrinsic properties might not adequately capture the competitive adsorption physics needed for high-selectivity forecasting. In addition, the model's ability to generalize to new, uncharacterized structural fragments is limited in the absence of specific thermodynamic benchmarks (fixed T, P) or external experimental validation. Future physics-informed directions must address the high uncertainty in selectivity modeling to guarantee that quick AI-based screening translates into reliable industrial carbon reduction technologies.[16]

Predictive accuracy by itself does not ensure scientific comprehension, even when high-throughput screening greatly reduces computing costs. The physicochemical principles regulating adsorption behaviour may be obscured by black-box models. In order to derive mechanistic insights from trained models and guarantee that predictions are rooted in adsorption physics, interpretability frameworks and thermodynamic mapping techniques are crucial.

3. Model Interpretability and Thermodynamic Mapping

3.1. Model Interpretability and Physical Insight

Machine-learning potentials (MLPs) based on quantum chemistry data offer a high-fidelity substitute for the rigid-lattice assumptions commonly used in molecular simulations of gas transport. Researchers were able to achieve remarkable forecast accuracy (R² = 0.9916) for system energies by using the DeepPot-SE model to reflect the dynamic flexibility of MgMOF-74. This method shows that structural flexibility plays a major role in CO₂ diffusivity prediction; rigid models greatly overestimate adsorption free-energy barriers, leading to diffusion coefficients that are over ten times slower than those found in flexible frameworks.

The "hopping" mechanism between open metal sites is highlighted by this integration of thermodynamics and machine learning, although the model's capacity for generalization still has to be critically assessed. The MLP's success is inextricably linked to the accuracy of the underlying semi-empirical lot since it was trained solely on simulated DFT-MD trajectories from a single 30 ps frame. Additionally, the absence of direct experimental validation for Mg-MOF-74 diffusivity highlights a more general requirement for standardized experimental benchmarks to confirm the industrial scalability of such physics-informed machine learning models. Although studies that depend on small simulated datasets are advantageous in identifying physical patterns such as flexibility-enhanced diffusion, they should be used with caution in real-world applications where complicated gas mixtures and structural flaws may change the kinetics of transport.[17]

3.2. Ensemble-Averaged Thermodynamics and Potential Energy Surface (PES) Mapping

The creation of transferable machine learning force fields (MLFFs) represents a paradigm shift in the high-throughput screening of porous materials for direct air capture. Researchers have achieved thermodynamic ab initio quality at computing rates equivalent to classical approaches by fine-tuning a foundation model (MACE-MP) using a specific GoldDAC dataset. This method shows that in chemically complex hybrid settings, typical UFF+DDEC models have systemic errors, especially in lanthanide-based frameworks where H2O adsorption energies are overstated by up to 17.8%. The DAC-SIM package has identified 161 promising candidates by transitioning from single-point interaction energies to ensemble-averaged properties. This study demonstrates that sustaining a high CO₂/H₂O selectivity (KH ratio > 100) necessitates the "selective pocket" mechanism facilitated by parallel benzene rings (PAR) and uncoordinated nitrogen. The existing reliance on rigid-framework assumptions continues to be a barrier to adequately capturing the regeneration costs and chemisorption dynamics essential for large-scale industrial deployment, even if these models successfully root screening in real-world physics.[18]

The "cooperative insertion mechanism" is less about static binding and more about a dynamic redistribution that favours faster kinetics, according to this blend of thermodynamics and machine learning. The study's dependence on a solely computational MLP framework underscores the difficulties in simulating chemical transitions to carbamates or carbonates in humid streams, even though it uses actual NMR data for confirmation. Future physics-informed machine learning approaches must close the gap between these transport models and the intricate stability and regeneration cycles needed in high-humidity flue gas conditions to guarantee industrial scalability

4. Molecular Transport and Multicomponent Separation Mechanisms

4.1. Molecular-Level Transport and Adsorption Mechanisms

The long-standing computational bottleneck of modeling flexible frameworks with open metal sites (OMS) at ab initio accuracy is addressed by the application of fragment-based Neural Network Potentials (NNPs). In comparison to SCXRD experimental standards, this study obtained a remarkable force RMSE of 0.039 eV/Å while preserving great transferability with only 0.54% variance in lattice parameters by using an E (3)-equivariant architecture (NequIP) trained on a small collection of ~2,000 DFT conformers. Importantly, a hybrid MD/GCMC workflow that integrates thermodynamics and machine learning shows that structural dynamics are crucial for realistic adsorption modeling: framework flexibility promotes structural relaxation, which delocalizes CO₂ molecules and optimizes the energy landscape at low pressures (0.1–1.0 bar).

At 298 K, conventional rigid-lattice GCMC models offer a respectable approximation, but, at higher temperatures, they greatly underestimate uptake due to their inability to account for the required structural modifications between the host and guest. Even though this active-learning-driven screening was successful, a critical assessment of the fragment-based approach indicates that, although transferable, these models need to be thoroughly validated against various chemical environments to make sure they don't "forget" complex interactions outside of their original training fragment. These high-fidelity potentials should be used in future physics-informed machine learning approaches to investigate the cooperative insertion processes and regeneration cycles necessary for large-scale commercial carbon capture.[19]

4.2. Multicomponent Gas Separation and Structural Design Strategies

An important step towards the practical implementation of MOF-based carbon capture is the shift from binary/ternary gas models to real 6-component natural gas mixtures (N₂, CO₂, CH₄, C₂H₆, C₃H₈, H₂S). This work achieved a predictive accuracy of R²=0.922 for material renderability by effectively identifying 10 elite MOF candidates from the 12,020-structure CoRE-MOF database by combining GCMC simulations with a Random Forest architecture. A "volcanic" structure-property relationship is revealed by a database-driven analysis of the data, where the best separation performance is limited to a density window of 0.5–1.7 g/cm³; frameworks outside of this range experience either kinetic exclusion or a loss of selectivity because of oversized pores.

The study outlines a clear tripartite design strategy for next-generation adsorbents: (i) replacing metal nodes, such as substituting Cd with Mn to quadruple working capacity; (ii) incorporating nitrogen-rich organic linkers, such as pyridine or azoles, to take advantage of electrostatic interactions. While the rigid-framework assumption successfully identified top-tier materials like XIGWUF and ETECOX, future physics-informed ML directions must go beyond static lattice models to incorporate the thermodynamic and kinetic impacts of framework flexibility in real-world, high-pressure natural gas streams.[20]

Local chemical conditions inside the framework ultimately determine adsorption selectivity and catalytic activation, even though transport modeling clarifies how CO2 moves through porous designs. In addition to merely structural optimization, engineering Lewis acid–base interactions, electronic structure modification, and synergistic site arrangements connects chemical reactivity and adsorption thermodynamics.

5. Active-Site and Electronic Structure Engineering

5.1. Lewis Acid-Base Site Engineering and Catalytic Kinetics

The merging of explainable machine learning with low-cost descriptors makes rapid screening of catalytic frameworks possible without the prohibitive expense of density functional theory (DFT). Researchers were able to predict CO₂ cycloaddition activity with an astounding 97% accuracy by using 372 high-yield experimental data points to train a Random Forest architecture. By measuring the "optimal Lewis acidity" of metal nodes using SHAP and PDP analysis, this study goes beyond "black-box" predictions. The findings indicate that a metal partial charge between 1.2 and 2.0 maximizes epoxide substrate activation while preventing active site deactivation through saturated coordination.

The effective experimental validation of MOF-76(Y), which attained a top-tier TOF of 64.72 h⁻¹, demonstrates that machine learning (ML) may bridge the gap between CO₂ use in the real world and hypothetical structure formation. However, a rigorous analysis indicates that to assure comparability across non-linear reaction profiles, the dependence on datasets collected from the literature requires extensive data cleaning (e.g., filtering for high yields). Future physics-informed machine learning directions should concentrate on standardizing experimental benchmarks to further enhance the commercial scalability of screened catalysts for a variety of catalytic reactions.[21]

5.2. Electronic Property Modulation and Electrocatalytic Selectivity

A superior class of 2D conjugated MOFs (2D c-MOFs) that outperform conventional Cu(211) benchmarks in CO₂ electroreduction performance has been found by combining DFT-driven discovery with Gradient Boosting Regression (GBR) designs. This study establishes a fundamental stability hierarchy (TMN4>TMN2O2>TMO4) and finds catalysts with very low limiting potentials, like NiN4−HDQ (UL=−0.04V for CO), by methodically altering both the metal active sites and organic ligands (HDQ series).

The primary electronic "handles" for tuning activity are electron affinity (EA) and electronegativity (χ), which together account for more than 34% of feature importance, according to a critical ML-based sensitivity analysis. The analysis indicates that ligand-induced orbital overlap is the main cause of intermediate adsorption strength, even though these 2D c-MOFs maintain pristine surfaces in aqueous conditions and show great thermal stability at 400 K. Future physics-informed machine learning approaches must connect these high-fidelity C1 models with multi-carbon (C2+) coupling kinetics in order to guarantee industrial scalability and fully realize the potential of these highly conductive, adjustable frameworks for artificial carbon cycle applications.[22]

5.3. Synergistic Site Engineering and Composite Pore Modulation

The computational "time trap" of simulating complicated composite adsorbents is addressed by integrating Convolutional Neural Networks (CNNs) with inception modules, offering a reliable framework for multi-criteria screening. Through training on 700 GCMC-validated structures, our model successfully navigated the inherent trade-offs between CO₂ working capacity and selectivity, achieving a high predictive fidelity (R² ≈ 0.90). Importantly, big data analysis of the 1,631-composite pool shows that unusual "synergistic" frameworks like IL@MARJAQ use the IL to create completely new potential energy minima for CO₂, whereas most ionic liquids improve selectivity by physically limiting N₂ transport.

Additionally, this study quantifies the non-linear impact of IL loading, proving that "more is not always better"; for instance, by adjusting the amount of injected molecules, the selectivity of IL@GUBKUL may be modified from 614 to over 7,000. Precision loading solutions replace trial-and-error post-synthetic alteration in this field. However, the rigid-framework assumption of the current model and its propensity to underestimate uptake in frameworks with open metal sites (OMS) indicate that lattice dynamics must be included in future physics-informed ML directions to fully capture the regeneration efficiencies and process economics needed for industrial flue gas separation.[23]

Insights gained from electrical modulation and active-site engineering naturally inform the next evolutionary stage, inverse design. Generative machine learning techniques use learned structure-property connections to suggest completely new MOF designs optimized for many aims instead of screening existing materials. This transition from predictive modeling to autonomous material generation essentially redefines the discovery paradigm.

6. Multi-Objective Inverse Design and Generative MOF Discovery

6.1. Multi-Objective Inverse Design and Chemical Subspace Exploration

The difficulty of navigating the almost infinite chemical space of porous crystals is addressed by the switch from brute-force screening to Deep Reinforcement Learning (DRL) for inverse design. Researchers have successfully produced physisorbents for Direct Air Capture with Qst values more than 40 kJ/mol and CO₂/H₂O selectivities greater than 1 by incorporating transformer-based predictors into a reward-driven environment. A database-driven analysis of these findings reveals a crucial trade-off in material "genes": high selectivity is controlled by particular Cu and Zn clusters that occupy a distinct subspace in the chemical design landscape, while high affinity for CO₂ is frequently driven by open metal sites (such as Mn-based N₁₃₁ nodes) that concurrently attract water.

The DRL approach's current reliance on classical force fields restricts its capacity to predict chemisorptive interactions involving charge transfer, despite its strong extrapolation capability—producing structures that compete with top-performing experimental MOFs like KAUST-7. Future physics-informed machine learning approaches must use DFT-derived charges and active learning loops to refine the predictors to improve industrial scalability. This will guarantee that inverse-designed candidates can continue to be effective in the humid, high-dilution environments typical of real-world atmospheric capture.[24]

The discovery of materials suited for particular carbon capture regimes has advanced significantly with the transition from manual high-throughput screening to reinforcement learning-enhanced generative design. Using the MOFGPT framework, researchers have shown that a reward-guided RL strategy can achieve 100% validity even when aiming for extreme CO₂ adsorption performance (mean + 2σ), but standard supervised fine-tuning is unable to produce a single chemically valid structure (0% validity). By optimizing MOFid string sequences, which encode both SMILES-based chemistry and RCSR-based topologies, this method enables the exploration of an almost unlimited chemical space without requiring the manual assembly of building blocks.

The RL model incorporates underlying structure-property correlations, as demonstrated by a database-driven examination of the generated candidates. For instance, open Cu²⁺ paddle wheel units and nitrogen-rich linkers that improve electrostatic interactions are consistently associated with strong CO₂ uptake. Additionally, the framework shows a capacity to investigate underrepresented areas of the chemical landscape that conventional screening could miss, with novelty rates consistently exceeding 63%. However, because it ignores the dynamic flexibility that is frequently essential for CO2 transport kinetics, the existing reliance on rigid-lattice assumptions continues to be a significant drawback. To close the gap between artificial intelligence-designed strings and artificially robust industrial sorbents, future physics-informed machine learning approaches should incorporate 3D structural building and DFT-based stability filters.[25]

Material-level optimization by itself does not ensure commercial viability, notwithstanding impressive gains in generative discovery. In the end, adsorption performance needs to be assessed under actual process settings, such as cycles of temperature and pressure swings. Therefore, integrating machine learning with process simulation closes the gap between system-scale performance and molecular design, allowing for the conversion of computer prediction into deployable carbon capture solutions.

7. Multiscale Generative Design and Process-Oriented Optimization

7.1. Process-Integrated Generative Design and Material Optimization

A major step forward from basic property screening to process-level material design is the creation of a multiscale generative workflow. Researchers have successfully navigated a terrain of trillions of potential structures using MOF-NET, an architecture based on NLP-style word embeddings, to identify candidates that "substantially outperform" benchmarks like CALF-20 and 13X zeolite. Strict size exclusion (3–5 Å pores) or the creation of high-density binding pockets (6–30 Å pores) where CO₂ molecules are stabilized by numerous oxygen-rich nodes are the two ways to achieve optimal performance, according to a database-driven analysis of the best-performing materials.

Crucially, this study provides experimentalists with a clear design guideline by identifying Cu-based nodes and fluorinated short-linkers as the "genetic markers" of elite adsorbents. The study identifies a continuous innovation gap even while the best-in-class small-pore material (HJM + N387 + E44) shows a notable productivity advantage over CALF-20 because of its improved N₂ rejection. The next frontier for physics-informed machine learning (ML) should be the integration of synthetic accessibility (SA) scores and lattice dynamics, ensuring that these "theoretical possibilities" can withstand the transition from computer to laboratory, as many computationally designed materials still encounter challenges in synthesizability and water stability.[4]