Domain Adaptation for Fuel Property Prediction Using Vibrational Spectroscopy

Accurate prediction of fuel properties is essential for accelerating the development and deployment of sustainable energy resources. While vibrational spectroscopy techniques such as attenuated total reflectance Fourier-transform infrared (ATR-FTIR) and Raman spectroscopy offer rapid, non-destructive alternatives to conventional fuel testing, machine learning (ML) models trained on spectral data are often limited by the need for large, technique-specific labeled datasets. Furthermore, models developed for one spectroscopic modality frequently perform poorly on others due to domain shift. In this study, we propose a domain adaptation (DA) framework to enable the transfer of predictive models trained on ATR-FTIR spectra to other vibrational spectroscopies, such as Raman, for fuel property prediction. Our contributions include the design of a DA strategy tailored to spectroscopy data, extensive validation across multiple target domains, and comparative analysis demonstrating significant performance gains over non-adapted models. The results highlight the potential of DA to reduce data requirements and improve model generalization across spectroscopic modalities, thereby enabling more scalable fuel property prediction.