A Graph-Based Machine Learning Framework for Predicting Fuel Mixture Properties with Limited Data

The accurate prediction of fuel mixture properties is essential for the development of alternative fuels, yet remains challenging under data-scarce conditions due to the combinatorial complexity of multi-component systems. In this study, we present a systematic evaluation of three machine learning (ML) strategies—feature engineering, transfer learning, and delta learning—to improve the prediction of key fuel properties, including derived cetane number (DCN), flashpoint, and viscosity. Using a graph-based deep learning framework based on the Directed Message Passing Neural Network (DMPNN), we encode molecular features for each component and aggregate them using molar-fraction-weighted averaging. Feature engineering incorporates chemically relevant descriptors computed from quantum and cheminformatics tools, while transfer learning leverages synthetic data generated from blending laws to pretrain the model. Delta learning reformulates the task as residual prediction relative to blending law estimates, enabling efficient learning of correction patterns. Among the strategies tested, residual-aware delta learning demonstrates the most robust and consistent performance across all targets, particularly in data-scarce regimes. The proposed framework offers a scalable and resource-efficient approach for high-throughput fuel screening and highlights the benefits of integrating physics-based priors into ML-driven chemical property prediction.