Discovering CO₂-Reactive Carbanions via Property-Guided Generative AI

Designing carbanions capable of efficient CO₂ chemisorption is an important approach for advancing reactive capture and conversion technologies. In this work, we integrate a nucleophilicity prediction model, trained using directed message-passing neural networks on Mayr's Reactivity Database (a large, experimentally derived collection of nucleophilicity, electrophilicity, and sensitivity parameters for organic and inorganic molecules), with the Hierarchical Variational Autoencoder (HierVAE) to generate novel carbanions. Fine-tuning the pretrained latent space with predicted log k values yields structurally diverse carbanions with strong CO₂ reactivity and high generation success. Analysis of the top candidates reveals consistent trends in α-substitution, electron-withdrawing groups, and synthetic accessibility. Density functional theory (DFT) validation of the highest-ranked reactive candidates reveals good agreement for electronically stabilized systems bearing strong electron-withdrawing substituents, while deviations increase for weakly stabilized or sterically distinct carbanions, defining a clear domain of applicability for the predictive model. This study demonstrates the promise of property-guided generative AI for discovering novel carbanions for room-temperature CO₂ chemisorption.