Accurate prediction of ionic-liquid (IL) surface tension (γ) remains challenging because of limited experimental data, temperature heterogeneity, and widespread data-leakage practices that inflate reported accuracies. Here we present a transparent, leakage-free machine-learning workflow for γ prediction across 263.32–394.70 K. The framework combines cation/anion-grouped data partitioning, recursive feature elimination, Optuna-based XGBoost optimization, and explicit applicability-domain (AD) analysis using leverage and MAD-standardized residuals. Conformal prediction intervals provide calibrated uncertainty bounds, enabling risk-aware screening decisions. The resulting model achieves robust generalization to unseen ILs while retaining physical monotonicity (γ ↓ T) and interpretability via SHAP analysis. Screening within the AD highlights phosphonium and quaternary-ammonium ILs with formate or hexanoate anions as plausible low-γ candidates, spanning 21–43 mN·m−1 in the calibrated range. Rather than competing for marginal accuracy gains, this work establishes a reproducible and auditable baseline for IL surface-tension modeling—one that explicitly separates statistical performance from chemical reliability and defines clear boundaries for valid prediction.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号