LiteBoost: a lightweight and explainable boosting model for predicting polymer density from SMILES data

Accurately predicting polymer density from SMILES strings remains challenging due to the small size, high noise, and chemically diversity of typical datasets. We introduce LiteBoost, a deliberately minimalist gradient boosting model that employs shallow, three-level symmetric trees and exposes only two tunable hyperparameters (n_estimators and learning_rate). Using a curated dataset of 613 polymers, we benchmark LiteBoost against ExtraTrees, XGBoost, LightGBM, and CatBoost, optimizing each with 100-1000 Optuna trials and evaluating performance across seven complementary metrics: R², RMSE, MAE, median AE, MAPE, maximum error, and explained variance. LiteBoost achieves a MAE of 0.031 g/cm³, RMSE of 0.062 g/cm³, R² of 0.81, and MAPE of 3.03%, all within 2-3% of the best-in-class CatBoost and XGBoost scores and well within the bounds of experimental uncertainty. Crucially, it does so with orders-of-magnitude fewer hyperparameters. These results demonstrates that a streamlined boosting model can rival heavyweight ensembles in accuracy while dramatically reducing tuning effort, computational cost, and interpretability barriers. LiteBoost is thus a practical first-line surrogate model for high-throughput polymer screening and inverse-design workflows where speed, robustness, and transparency are as critical as raw predictive power.