Beyond the Single Pose: Perturbation-Resolved Mapping of Host-CO2 Interaction Landscapes with Interpretable Machine Learning.

Screening carbon-dioxide capture media requires models that examine diverse molecular environments while retaining interpretable host-CO2 descriptions. This work presents a nonreactive finite-complex workflow in which 11,053 parent environments, indexed by the Kulkarni Universal Interaction Descriptor (K-UID), a scale-invariant topological address system, were expanded into 975,578 encounter geometries by appending rigid CO2 at directed and random positions without changing parent coordinates. A reference subset of 149,935 complexes was relaxed using Geometry, Frequency, and Noncovalent eXtended Tight Binding (GFN2-xTB) and processed using an in-house implementation of real-space noncovalent interaction analysis and the published Kulkarni noncovalent interaction fingerprint (KNF) framework. The calculations provided the nine KNF components, the Non-Covalent Interaction Score (SNCI), measuring cumulative attractive interaction load, and the area-weighted COSMO surface-charge variance (VCDI), from which the plotted interaction-character coordinate SCDI = 1 - VCDI was derived. A machine-learning surrogate combined a multioutput multilayer perceptron with a Light Gradient Boosting Machine residual correction to learn the 11 quantities from unoptimized Cartesian structures and placement metadata. On host-grouped evaluation, the model achieved R2 = 0.990 for SNCI and a seven-split mean R2 of 0.8485 ± 0.0127. It reproduced the two-dimensional SNCI-SCDI screening map with 0.8935 region accuracy and 0.8932 macro-averaged F1 score across 149,856 matched perturbations. The map separated high-interaction-load systems by interaction character and revealed chemically organized host families. The workflow prioritizes noncovalent molecular encounters for subsequent condensed-phase simulation or experiment; it does not model covalent CO2 chemistry or directly predict bulk absorption capacity.