Transferable Free-Energy Priors for MLIP-Driven Sampling of CO2 Chemisorption in Isoreticular Covalent-Organic Frameworks

Covalent-organic frameworks (COFs) are promising solid-sorbent materials for CO2 capture, and amine-functionalized COFs in particular for direct air capture. The capture process is chemisorption, a chemical reaction in which the amine nitrogen attacks the CO2 carbon and a proton moves away, leaving carbamic acid or carbamate. Sampling this rare event with density functional theory (DFT) accuracy is computationally prohibitive. Machine-learned interatomic potential (MLIP)-driven biased molecular dynamics (MD) offers near-DFT accuracy in sampling the underlying free energy landscape at a fraction of the cost. However, every amine site within a COF needs its own biased MD run, and most of that run is spent on filling a deep product well that other sites largely share. This is especially pronounced when studying multiple amine sites within one COF, or several amine-functionalized COFs under realistic (humid) conditions, which is necessary to obtain mechanistic insights into the synergistic or competitive effects of water on CO2 capture. In this work, we introduce transferable bias priors that let the sampling of one target seed the next: each new target starts from a static bias V0 assembled from the wells of its previously sampled relatives. The sampling workflow is applied with explicit water loading to six amine-functionalized COFs: four isoreticular sonoCOFs, NH2-Th-Bta-COF, and COF-999. The prior transfers the wells, which relatives share, and leaves the barriers, which they may not, to each run. On sonoCOF-H3, a prior built from three amine sites reproduces the barrier of a fourth to within 1 kJ/mol of a from-scratch control at equal sampling, well inside chemical accuracy. This work, thus, illustrates how the free-energy wells shared across related amine sites can be written into a bias prior and reused, thereby accelerating sampling while maintaining sufficient accuracy.