Multi-Objective Optimization of Conceptual DFT Reactivity Descriptors in Open-Shell Radicals by Reinforcement Learning

Open-shell organic radicals underpin catalysis, energy materials, and spin-based technologies, yet rational design is hindered by the difficulty of tuning electronic reactivity while preserving chemically meaningful local response. Here, a descriptor-driven reinforcement learning framework is used to regulate radical reactivity in terms of conceptual density functional theory (CDFT) based reactivity indices, enabling electrophilicity index (ω) to be driven toward the experimentally motivated target ω ≈ 1.0 eV while maintaining admissible atom-condensed Fukui function behavior. Using a Twin Delayed Deep Deterministic Policy Gradient (TD3) anchored to benchmark tolerances, an 85.7% multiobjective success rate is achieved on held-out radicals, and all of the top 20 candidates ranked by a composite ω-Fukui function score satisfy both global and local criteria. Class- and motif-resolved analyses reveal a clear hierarchy in tractability: electronically flexible scaffolds, including phosphoryl, silyl, boryl, and alkyl-centered radicals, consistently converge to balanced reactivity regimes, whereas rigid π-conjugated or lone-pair-locked motifs, such as aromatic, heteroaryl, and halogen-substituted radicals, resist coordinated tuning. Complementary reward ablation further shows that robust convergence requires coupling global ω control to local Fukui function-based regularization, as single-descriptor objectives lead to unstable or incomplete optimization. Together, these results demonstrate that electronic flexibility governs the tunability of open-shell reactivity descriptors and provide a practical strategy for radical selection and descriptor-space optimization using chemically interpretable electronic metrics.