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

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting polymer density from SMILES strings is challenging due to dataset limitations.
• Existing models often require extensive hyperparameter tuning and significant computational resources.

Purpose of the Study:

• Introduce LiteBoost, a minimalist gradient boosting model for polymer density prediction.
• Evaluate LiteBoost's performance against established ensemble methods.
• Demonstrate LiteBoost's efficiency in terms of hyperparameters and computational cost.

Main Methods:

• Developed LiteBoost with shallow, three-level symmetric trees and two hyperparameters (n_estimators, learning_rate).
• Curated a dataset of 613 polymers.
• Benchmarked LiteBoost against ExtraTrees, XGBoost, LightGBM, and CatBoost using Optuna for optimization.
• Evaluated performance using R², RMSE, MAE, median AE, MAPE, maximum error, and explained variance.

Main Results:

• LiteBoost achieved competitive results with MAE of 0.031 g/cm³, RMSE of 0.062 g/cm³, R² of 0.81, and MAPE of 3.03%.
• Performance was within 2-3% of top-performing models like CatBoost and XGBoost.
• LiteBoost required significantly fewer hyperparameters and less tuning effort compared to other models.

Conclusions:

• A streamlined boosting model like LiteBoost can achieve high accuracy in polymer density prediction.
• LiteBoost offers a practical, efficient, and interpretable alternative for high-throughput polymer screening and inverse design.
• The model's simplicity reduces barriers to adoption in computational workflows.