Interpretable machine learning for predicting the derived cetane number of jet fuels using compact TD-NMR

The derived cetane number (DCN) is a commonly used metric that summarizes fuel ignition characteristics, including ignition propensity and chemical kinetic potential for combustion processes. Traditional methods for determining the DCN of jet fuels are ASTM standards that involve large-scale, laboratory-based experiments. While recent advancements include the estimation of DCN via nuclear magnetic resonance (NMR) and infrared spectroscopy, the search persists for a method capable of real-time and in-situ estimations. This work proposes the use of a compact time-domain NMR (TD-NMR) system for the acquisition of jet fuel T2 relaxation curves. The system is validated using relaxometric experiments and demonstrates the ability to acquire consistent, structurally viable data on a time-scale of just minutes. Furthermore, an interpretable approach for relaxometric data analysis is presented, allowing for the estimation of a sample’s DCN directly from its T2 relaxation curve. Random forests are trained for DCN prediction on both hydrocarbon and jet fuel samples, and the importance of extracted T2 curve features are investigated using both the permutation of out-of-bag predictors and partial dependence plots. A model trained on less than 200 total relaxation curves is tested using two novel jet fuel samples, with predictions achieving an RMSE of just 0.96 DCN. Finally, the applicability and limitations of the proposed scheme are discussed.