Computational screening methodology identifies effective solvents for CO2 capture

The study addresses the challenge of improving CO₂ capture efficiency in post-combustion applications. Conventional primary and secondary amines (e.g., MEA, DEA) react rapidly with CO₂ to form carbamates but require high regeneration energy and reduce plant net output substantially. Tertiary amines (e.g., MDEA) capture CO₂ as bicarbonate with lower heat of reaction and higher capacity but suffer from slow absorption kinetics, making them unsuitable for low-pressure flue gas without an activator. Industry practice adds piperazine to boost rates, but systematic discovery of faster tertiary amines remains time- and labor-intensive. The research question is whether a combined molecular simulation and machine-learning methodology can accurately predict CO₂ absorption kinetics and efficiently screen large chemical libraries to identify tertiary amines with significantly improved absorption rates (especially when paired with piperazine) while maintaining favorable thermodynamics for regeneration.

Prior work has used molecular simulations and machine learning extensively for physical adsorption materials (e.g., MOFs), enabling large-scale virtual screening. For chemical absorption in amines, prior QSPR efforts have focused on predicting properties related to CO₂ absorption but were constrained by small datasets, limiting model applicability. Experimental datasets such as Chowdhury et al. provided rates for a limited set of tertiary amines. Industry has adopted piperazine as an activator to enhance kinetics of tertiary amines like MDEA. However, robust predictive models that capture solvent effects on reaction barriers and scale to diverse tertiary amines remained lacking, motivating the integrated approach developed here.

The authors developed a multistage workflow combining molecular dynamics (MD) simulations, QSPR machine learning models, virtual screening, and targeted experimental validation.

• Molecular simulations: An MD-based kinetic model (per Romanazzo et al.) predicts CO₂ absorption rates by modeling the key reaction CO₂ + OH⁻ → HCO₃⁻ in aqueous tertiary amine solutions where the amine acts as a base. The activation Gibbs free energy ΔG⦵ is derived via a Polanyi–Eyring/Erenfest-type relation using solvation free energy differences of OH⁻ + CO₂ (reactants) and HCO₃⁻ (product), with parameters fitted to experimental rates in water and various amine solvents. Concentrations of CO₂ and OH⁻ are computed by solving pH equations. The model is robust to variations in amine concentration and temperature.
• Simulation datasets: The model was validated against an experimental dataset of 24 tertiary amines (313 K, 30 wt% amine) and applied to a diverse set of 100 tertiary amines (1 mol% amine, 323 K) sourced from in-house and public datasets, including linear/cyclic amines, diamines, and sulfur-containing derivatives. MD outputs include predicted absorption rates (R_MD, including R_max) and absorption free energies (ΔG_MD).
• QSPR modeling: Using the MD-predicted endpoints (R_max, ΔG_max) and experimental data, the team built QSPR models. Structures were standardized (RDKit/KNIME). Descriptors included ISIDA fragments (sequences, atom pairs, triplets) and OPERA physicochemical properties (e.g., predicted pKa, logP, melting/boiling points, vapor pressure, solubility, substructure counts). Machine-learning algorithms (Random Forest, SVR, XGBoost) were tuned via grid search and assessed using nested cross-validation with multiple reshuffles; performance metrics included Q²_cv, RMSE_cv, and MAE_cv. pKa predicted by OPERA served as a strong single predictor; consensus models across descriptor sets and algorithms were used for robust ranking.
• Virtual screening: The ZINC23 database was filtered for tertiary amines meeting criteria (MW < 250 g/mol, −1 < logP < 1, commercially available). Additional structural filters removed non-tertiary amines, double bonds/aromatics where excluded per criteria, primary/secondary amines, certain functional groups. Hundreds of candidates (>800) were retained and ranked by QSPR-predicted absorption rate (R_QSPR) and ΔG°_RSPR.
• Experimental validation: A thermo-regulated, constant-flow Lewis-type reactor with controlled stirring and temperature was used to measure CO₂ absorption kinetics for selected amines. The initial CO₂ amount was fixed, and the temporal evolution of gas-phase CO₂ partial pressure was monitored; the slope at 50% of equilibrium uptake, r(CO₂), quantified the absorption rate. Eighteen amines (7 from the Chowdhury set, 3 from the 100-MD set, and 8 novel amines from screening) were tested. Additional experiments assessed the effect of adding piperazine (PZ) activator; solutions contained 13 mol% amine (or 11 mol% amine + 2.5 mol% PZ). Data analyses used established EOS (GERG-2008 via REFPROP) and mass-balance approaches for vapor–liquid quantification.

• The MD model accurately reproduced experimental CO₂ absorption rates for 24 tertiary amines and showed near-perfect rank correlation: Spearman r_S = 0.99 between MD and experimental rates at 313 K; ΔG_MD strongly correlated with R_MD (Spearman r ≈ 0.98), with slower absorption corresponding to higher ΔG_MD.
• Application to 100 diverse tertiary amines at 323 K identified many with faster predicted kinetics than industrial MDEA; rapidly absorbing structures frequently contained piperidine or pyrrolidine rings, consistent with literature.
• QSPR models built on ISIDA and OPERA descriptors (including predicted pKa) achieved robust cross-validated performance adequate for ranking large libraries; while RMSE values were moderate, the applicability domain was larger than prior studies due to a threefold increase in training data.
• Virtual screening of >800 ZINC23 tertiary amines found numerous candidates predicted to outpace MDEA in R_QSPR. Substituted piperidines and related cyclic tertiary amines populated the top ranks.
• Experimental validation of 18 amines showed strong correlations between predicted metrics (R_QSPR, ΔG°_RSPR, predicted pKa) and measured r(CO₂) for eight novel amines (Spearman ≈ 0.93). Four of eight piperidines absorbed CO₂ faster than MDEA. Two screened amines were especially effective among the tested set.
• A newly suggested tertiary amine (e.g., EPOL as cited) displayed faster CO₂ absorption than MDEA without activator; adding piperazine further increased rates, with the amine+PZ mixture showing the fastest absorption among tested conditions.

By integrating high-precision MD simulations that capture solvent effects on the reaction barrier with QSPR machine learning models trained on simulated and experimental endpoints, the study demonstrates an efficient route to prioritize tertiary amines with improved CO₂ absorption kinetics. The strong rank correlations indicate that thermodynamic descriptors (ΔG of absorption) and basicity (pKa) can serve as reliable, transferable predictors of kinetic performance across structurally diverse tertiary amines. The virtual screening successfully enriched for piperidine/pyrrolidine-containing candidates, and experimental tests confirmed that several outperformed MDEA, validating the screening methodology. The demonstrated enhancement with piperazine further underscores industrial relevance, as activator-augmented tertiary amines can meet kinetic requirements without incurring the high regeneration penalties associated with primary/secondary amines. Overall, the approach addresses the core challenge of accelerating solvent discovery for CCS by drastically reducing experimental burden while maintaining predictive fidelity and broad applicability.

The work introduces and validates a computational-experimental workflow for discovering effective tertiary amine solvents for CO₂ capture. A solvent-aware MD kinetic model accurately predicts absorption rates and energies, enabling the construction of broadly applicable QSPR models. Large-scale virtual screening identified multiple candidates with faster absorption than MDEA, particularly among piperidine/pyrrolidine derivatives, and experimental validation confirmed improved kinetics for several novel amines, including strong performance when combined with piperazine. This methodology can guide rapid, cost-effective exploration of chemical space for CCS solvents. Future research should extend to broader property optimization (e.g., volatility, corrosion, oxidative stability), multi-objective screening balancing kinetics and regeneration energy, testing under varied process conditions and gas compositions, and expanding the training datasets to further improve model accuracy and domain coverage.

• Some textual inconsistencies suggest that certain compound identifiers/acronyms in the narrative may be garbled; exact chemical identities of top performers beyond general classes are not exhaustively detailed in the excerpt.
• QSPR models, while robust for ranking, do not achieve very low RMSE; predictions are best interpreted for prioritization rather than exact rate estimation.
• Training data, though expanded, remain limited compared to the size of chemical space; model performance outside the applicability domain may degrade.
• The MD model focuses on the OH⁻-mediated pathway and treats tertiary amines as bases; potential alternative mechanisms or specific solute–solute interactions may not be fully captured for all structures.
• Experimental validation covered 18 amines; broader testing and full process evaluations (e.g., long-term stability, degradation, corrosion, foaming) are needed for industrial translation.