Pretrained 3D Molecular Representations Enable Data-Efficient Discovery of High-Energy-Density Fuels

Abstract

High-energy-density hydrocarbon (HEDH) fuels are essential for aerospace propulsion, yet the design of such fuels is limited by the scarcity of reliable property data. In this work, we fine-tuned Uni-Mol, a pretrained three-dimensional (3D) molecular representation learning framework, on a dataset of 316,069 hydrocarbons from GDB-13. Each molecule in the dataset was labeled with six physicochemical properties calculated using the group contribution method. With only 1% of the training set, Uni-Mol achieved a flash-point mean absolute error (MAE) of 1.67 K. After further training on hydrocarbons with a broader carbon-number range, the model achieved coefficients of determination (R2) of 0.9672–0.9998 across six GDB-17 properties. High-throughput screening of this subset identified seven polycyclic candidates with exceptional energy density and thermal stability. These results demonstrate the potential of pretrained 3D molecular representations for data-efficient molecular discovery and provide a scalable framework for the accelerated identification of next-generation energetic materials. This work provides a foundation for future experimental validation.