SMILES Token Additivity Model with Interpretability and Generalizability for Fuel Property Predictions

Area of Science:

• Computational chemistry
• Machine learning
• Chemical engineering

Background:

• Quantitative structure-property relationship (QSPR) models face challenges in interpretability and generalizability.
• Deep learning approaches often require complex feature engineering, limiting their application.

Purpose of the Study:

• To develop an interpretable and generalizable deep learning model for predicting fuel properties.
• To establish a novel quantitative structure-property relationship (QSPR) approach using simplified molecular input line entry system (SMILES) token additivity (STA).

Main Methods:

• Utilized stacked multihead self-attention encoders to process SMILES strings.
• Developed the simplified molecular input line entry system (SMILES) token additivity (STA) model.
• Applied the model to predict seven critical fuel properties: standard enthalpy of formation, entropy, isobaric heat capacity, cetane number, boiling point, melting point, and flash point.

Main Results:

• Achieved high predictive accuracy for all seven fuel properties, with R² values exceeding 0.95.
• Demonstrated low mean absolute errors for thermodynamic properties (e.g., 1.86 kcal/mol for Δ_fH°).
• Showcased comparable accuracy to traditional machine learning models while providing token-level insights into structure-property relationships.

Conclusions:

• The STA model offers a powerful and interpretable alternative for fuel property prediction.
• The model's ability to generalize across properties and provide insights into molecular contributions highlights its potential.
• This approach advances the field of quantitative structure-property relationship (QSPR) modeling by integrating deep learning with interpretable token-based analysis.