Autonomous Reaction Discovery of CO₂ Capture in Aqueous Ammonia through Active-Learning Neural Networks

The mechanistic origin of CO₂ capture in aqueous ammonia has long remained debated, with competing proposals invoking concerted versus stepwise pathways, ambiguous catalytic roles of ammonia, and uncertain product distributions. Here, we introduce an active learning, data-driven framework (ADRML) that integrates reactive molecular dynamics (RMD) with dimensionality-reduced sampling. RMD simulations employing the trained machine-learned interatomic potentials (MLIPs) on cluster models with periodic boundary conditions reveal that the [CO₂]/[NH₃] concentration ratio (R_[C]/[A]) is a critical determinant of product distributions, kinetics, and underlying mechanisms. At high R_[C]/[A], carbonic species are favored via water-mediated hydration, whereas low R_[C]/[A] markedly promotes carbamate formation through a concerted ammonia-ammonia pair mechanism that becomes accessible only under ammonia-rich (low R_[C]/[A]) conditions. This disparity underscores a distinct concentration-dependent mechanistic shift in carbamate formation─from a stepwise to a concerted pathway. In addition, the hydronium ion (H₃O⁺) generated in the carbonate-formation channel ultimately suppresses further reactivity by promoting the reverse process of carbonate hydrolysis and by depleting NH₃ through protonation. Overall, at low R_[C]/[A] ≈ 0.2, carbamate production surpasses carbonate formation in both rates and yields, occurring before significant H₃O⁺ accumulation from the carbonate channel, thereby maximizing CO₂ uptake. Both carbonate and carbamate formation reactions compete across the entire range of R_[C]/[A]. However, the dramatic enhancement of carbamate formation at low R_[C]/[A] is likely the primary source of long-standing ambiguities in these systems.