Predictive Modeling of Yield Sooting Index Using Machine Learning with Uncertainty Estimation.

Area of Science:

• Computational chemistry
• Materials science
• Chemical engineering

Background:

• Quantitative structure-property relationship (QSPR) connects molecular structure to fuel properties.
• Predicting fuel behavior, including yield sooting index (YSI), is crucial for developing efficient fuels.
• Machine learning (ML) offers advanced tools for developing predictive models.

Purpose of the Study:

• To develop and validate two predictive models for the yield sooting index (YSI) of various fuels.
• To utilize multilayer perceptron (MLP) networks and QSPR methodology for accurate fuel property predictions.
• To identify key molecular descriptors influencing YSI through advanced feature selection techniques.

Main Methods:

• Development of two ML models: one using Gini importance and another using genetic algorithms for feature selection.
• Application of QSPR methodology to link molecular descriptors with fuel properties.
• Rigorous data preprocessing, feature selection, hyperparameter tuning, and uncertainty estimation.

Main Results:

• The genetic algorithm model demonstrated superior performance over the Gini importance model by reducing feature autocorrelation.
• Key molecular descriptors significantly impacting YSI were identified.
• A strong correlation was found between specific 2D matrix-based descriptors and YSI, offering new predictive insights.

Conclusions:

• The developed ML-QSPR models are robust and reliable for predicting fuel properties.
• This study highlights the synergistic potential of ML and QSPR in fuel research.
• The findings contribute to the advancement of computational methods for sustainable and efficient fuel alternatives.