Probing the Prediction of High-Temperature Ignition Delay Times of Jet Fuels via Machine Learning Approaches

• The first machine learning method for ignition prediction of complicated jet fuels. • The random forest algorithm is suitable for the prediction of ignition delay times. • Neural network model can be significantly improved by more descriptors. • Machine learning methods are better than traditional linear regression methods. • SHAP analysis highlights the important inputs on ignition models. Understanding the relationship between jet fuel’s composition and ignition property is not only critical for airworthiness certification process in aviation industry, but also useful to optimize manufacturing technique to produce high quality jet fuels. Traditional experimental method to measure the ignition delay times (IDTs) is time-consuming, expensive and requires expert knowledge. Herein, this work probes the development of machine learning (ML) approaches to predict the high-temperature IDTs for conventional, alternative and sustainable aviation fuels. By using the concentrations of reactant mixtures, combustion conditions, molecular formula and fuel compositions as input, several ML models are developed with good accuracy. The random forest (RF) method is the most suitable algorithm for the prediction of IDTs of jet fuels. The error indicators of the ML models including root mean square error, mean absolute error and coefficient of determination are around 0.50, 0.25, and 0.94, respectively. The gradient boosting (GB) model can achieve prediction accuracy that competes with the RF model. The performance of neural network (NN) models can be improved by the usage of detailed fuel compositions from two-dimensional gas chromatography (GC × GC) analysis. The developed ML models are significantly better than multiple linear regression models. SHAP and PFI analysis reveals that the molecular formula and quantity of aromatics of the jet fuels are important descriptors affecting the model performance besides the compulsory combustion conditions. This work is useful for the prediction of IDTs of various jet fuels, and also provides valuable information for further development of more accurate models for combustion property predictions.