Fuelprop: Fuel property prediction from ATR-FTIR spectroscopic data

Abstract

Synthetic fuels are crucial for decarbonizing the transportation sector. A significant challenge lies in the rapid and efficient characterization of these fuels. Chemometric methods using ATR-FTIR data offer a potential alternative to conventional techniques. This study expands the applicability and performance of chemometric models by providing an extensive ATR-FTIR spectral dataset and exploring various data enhancement strategies. Data enhancement was achieved by semi-supervised data generation, consistency enforcement through unsupervised data augmentation, and data imputation using synthetic spectra blending and pseudo-labeling. Models were trained on surrogate fuels and rigorously tested on real fuels, representing out-of-distribution testing conditions. We believe that this work will enhance the adoption of chemometric models for fuel characterization.