Machine learning predictions of thermochemical properties for aliphatic carbon and oxygen species

In thermochemistry, predicting fundamental properties such as entropy and specific heat capacity remains underexplored, with most studies primarily focusing on the enthalpy of formation. This limits our understanding of the thermochemical landscape, particularly in combustion research, where precise thermochemical data is essential for optimizing fuel and propellant efficiency. Traditional methods, such as group additivity and quantum calculations, are often costly when dealing with large and complex molecular species, presenting a challenge in predicting their thermochemistry. Recent advancements in machine learning (ML) present a promising solution for efficiently predicting combustion-related thermochemical properties. Despite this potential, challenges persist in optimizing molecular representations and selecting appropriate models. This study aims to bridge this gap by introducing a carbon, hydrogen, and oxygen-containing species dataset. We systematically evaluate the performance of fourteen featurization methods and nine ML models, incorporating error estimations and hyperparameter tuning. Our results demonstrate that the Composite or Custom Descriptor Set (CDS) combined with the Random Forest (RF) model yields a chemical accuracy (95 % confidence interval) of 2.21 kcal/mol for enthalpy of formation at 298.15 K, 2.20 cal/(molK) for entropy at 298.15 K, and an average of 2.63 cal/(molK) for specific heat capacity across temperatures from 300 K to 1500 K. Such results highlight the effectiveness of using a single ML method to predict multiple thermochemical properties, underscoring the contribution of our study to the field of thermochemistry.