Thermophysical Property Prediction and Optimization of CO₂–Binding Organic Liquids Using Kolmogorov–Arnold Networks and Optimized Quantum Descriptors

Screening emerging CO₂–Binding Organic Liquids (CO₂BOLs) requires accurate, data-efficient tools given scarce experimental data. We introduce a flexible genetic algorithm (FlexGA) to optimize σ-profile segmentation by identifying the most informative regions of the molecular surface charge distribution. An ensemble of models provides robust property predictions and explicit uncertainty quantification (UQ). We implement Kolmogorov–Arnold Networks to extract the symbolic form of the ensemble mean, deriving interpretable analytical mappings between molecular descriptors and macroscopic properties. Leave-One-Compound-Out validation demonstrates high predictive accuracy for density (R² = 0.977), viscosity (R² = 0.996), and vapor pressure (R² = 0.903), performing well on unseen blends of 1,8-diazabicyclo[5.4.0]undec-7-ene (DBU). DBU with hydroxyl compounds is optimized via a multiobjective framework that integrates UQ for risk-aware Pareto analysis. Results identify 1-hexanol blends for minimized solvent loss (0.25 kPa vapor pressure with 3.45 mPa·s viscosity) and ethylene glycol monomethyl ether blends for enhanced circulation fluidity (1.75 mPa·s viscosity).