Artificial intelligence for novel fuel design

Co-optimization of fuels and engines can help decarbonize the transportation sector. Traditional methods for fuel-engine co-optimization rely on experimental or theoretical approaches for determining relationships between fuel molecular properties and engine performance. However, these methods can be expensive and time consuming to develop and are prone to errors when there are complex non-linear relationships between molecular structure and combustion properties. Data-driven approaches that rely on artificial intelligence (AI) and machine learning (ML) algorithms are gaining interest for their versatility, adaptability, and ease of use. This paper provides an overview of AI and ML methods and their applications in fuel design. ML algorithms are now routinely developed to predict quantitative structure property relationships (QSPR) for combustion applications. However, molecular representations relying on descriptors are limited to iterative screening of molecules or mixtures for fuel design; they do not enable inverse fuel design wherein a fuel's structure or composition is predicted based on desired target properties. Inverse fuel design is enabled by using AI/ML methods to train QSPR models using molecular graphs and neural networks for a fully differentiable mapping of molecular structure to fuel properties. Gradient-based search methods can then be employed on the latent space representation to generate novel fuel components or mixtures. While AI for fuel design is an evolving domain, the need for uncertainty quantification in training data and model parameters cannot be overlooked. Methods for addressing aleatoric and epistemic are discussed herein together with outlooks for their applications in fuel design.