Comparative Machine Learning Approaches for QSPR-Based Prediction of Thermochemical Properties in Rocket Fuel Candidate Screening

Abstract Experimental screening of compounds based on thermochemical properties is a time-consuming and expensive process. In this work, a multi-output Artificial Neural Network is developed using molecular descriptors to predict fuel-related thermochemical properties, such as flash point, vapor pressure and surface tension. The input features are selected using domain knowledge and information-gain metrics such as mutual information. Using cross-validation techniques and exhaustive feature engineering, the developed model produced competitive performance, with that of ensemble learning models like XGBRegressor and supervised machine learning models like SVR, for several target properties while enabling simultaneous prediction of multiple thermochemical properties. This architecture demonstrate the potential of using multi-output shared representation to accelerate the computational screening of candidate compounds for rocket fuel applications.