Data-driven, explainable machine learning model for predicting volatile organic compounds' standard vaporization...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Physical Chemistry

Background:

• Accurate prediction of standard vaporization enthalpy (Δ_vapH_m°) for volatile organic compounds (VOCs) is crucial for environmental, industrial, and regulatory applications.
• Traditional experimental methods are time-consuming and costly.
• Existing machine learning (ML) models have limitations in prediction accuracy and applicability.

Purpose of the Study:

• To develop a data-driven, explainable supervised ML model for predicting the standard vaporization enthalpy (Δ_vapH_m°) of VOCs.
• To provide a high-throughput and cost-effective alternative to experimental property estimation.
• To enhance the accuracy and applicability of ML models in chemical property prediction.

Main Methods:

• Utilized a supervised ML regression approach, specifically the Random Forest algorithm.
• Trained the model on an experimental database of 2410 unique molecules, including 223 VOCs categorized by chemical groups.
• Validated the model through prediction on known VOC databases and molecular group hold-out tests.

Main Results:

• The Random Forest model accurately predicted VOCs' Δ_vapH_m° with a mean absolute error of 3.02 kJ mol^-1 and achieved a 95% test score.
• Chemical feature importance analysis identified VOC polarizability, connectivity indexes, and electrotopological state as key predictors.
• The model demonstrated replicability and explainability.

Conclusions:

• The developed explainable ML model offers a reliable and efficient method for predicting the standard vaporization enthalpy of VOCs.
• The model's explainability provides insights into the key molecular descriptors influencing vaporization enthalpy.
• This approach can be expanded for predicting other thermodynamic properties of VOCs.