Abstract

Developing high-performance fuels from low-carbon or carbon-neutral streams can significantly advance decarbonization efforts in the transportation sector. Given that practical fuels are complex mixtures of hundreds of components, and considering the highly non-linear nature of fuel properties, the task of exploring a broad chemical space for potential blendstocks becomes a cyclical, costly, and time-consuming process. In the era of matter engineering, inverse design has become a vital component in the sophisticated framework required for the rapid design of fuels. The swift advancements in artificial intelligence, particularly in machine learning, have propelled the application of inverse design in various fields, where specific properties are pre-selected to identify suitable new candidates.

This dissertation introduces an inverse, data-driven framework for liquid fuel formulation that employs a constrained optimization approach. Central to this framework are predictive deep learning (DL) models and robust search algorithms for efficient chemical space navigation.  A novel contribution of this work is the integration of an algorithmic advancement into the training loop, directly connecting molecular structures with mixture representations and facilitating mixture-level fuel screening. This level of screening evaluates several crucial combustion-related properties, including octane rating, sooting propensity, and volatility, as well as the emissions of unburnt hydrocarbons. To effectively represent unburnt hydrocarbons derived from gas trapped in the combustion crevice, a generic 1D premixed stagnant flame modeling approach was adopted.

The framework's effectiveness is demonstrated through the design of high-octane, low-sooting fuels that adhere to gasoline specification constraints, using a variety of gasoline blendstocks. This work also assesses various molecular representations and state-of-the-art deep learning practices. Finally, the dissertation presents a search workflow through several case studies that directly optimize mixtures within the domains of fossil, oxygenated, and synthetic streams. We expect our simple and practical framework will serve as a solid baseline and help ease future research designing liquid energy carriers.