Abstract
The chemical structure of fuels significantly influences the properties of ignition and energy release during combustion, making the exploration of molecular structure-property relationships a key focus for the research and development of new sustainable fuels. Given the vast combinatorial possibilities of potential fuel candidates, prioritization is essential. This study explored the use of generative modeling to propose novel molecular structures for future fuels. Specifically, the long short-term memory (LSTM) autoregressive model was fine-tuned using a hill-climb optimization algorithm to generate structures optimized for high-knock resistance. The generated compounds, unseen during training, were evaluated for their physical properties and research octane number (RON). The generated molecules contained features commonly associated with knock resistance, such as branching and aromaticity, while also uncovering unconventional structures, including oxygenates with ether linkages. This work underscores the promise of generative modeling in fuel design and highlights the strategic advantage of initiating molecular generation from predefined fragments related to known feedstocks and production processes to enhance practicality in synthesis and resource utilization.