Open Access

Flammability characteristics of hydrogen mixtures have been extensively investigated at different initial conditions(temperature and pressure). Based on the available experimental datasets, empirical and semi-empirical models are commonly used to calculate flammability limits in dependance to initial conditions and mixture composition to reduce the experimental effort. However, unevenly distributed empirical data and the complex non-linear relationship characteristics of these data present significant challenges to empirical flammability limits prediction methods under various mixture initial conditions. Moreover, the empirical models and semi-empirical models only cover some influencing parameters, respectively. To address these issues, the present study adapts a machine learning (ML) approach for improving the hydrogen-air/oxygen-inert gas mixture flammability limits prediction at different conditions with a holistic approach. A Multi-Layer Perceptron (MLP) model was trained, validated, and tested using key input features such as flammability state, initial mixture temperature, equivalence ratio, inert gas concentration, adiabatic flame temperature, and Lewis numbers. Data augmentation techniques were conducted on experimental datasets to improve the predictive capability of the model. The models’ performance was compared with empirical flammability limit prediction methods. The goal is to deliver fast, reliable, and more accurate predictions across different scenarios with a single prediction model. Most importantly, the machine learning approach offers a cost-effective and robust alternative to existing empirical flammability limit prediction methods, thus also reducing the experimental effort for explosion limits determination.

Predictions on flammability limits had been conducted using semi-empirical and fully empirical models, which are either limited in scope of application and/or affected by significant methodical prediction errors. Some parameters influencing the flammability limits are not considered in the non-comprehensive empirical models. In the present study, optimized multilayer perceptron models are investigated to improve the accuracy of the prediction of flammability limits in dependance from the temperature and any inert gas admixtures. Different input features constituted by the flammability state, mixture composition, initial mixture temperature, adiabatic flame temperature, and Lewis numbers were implemented into the multilayer perceptron model. Data limitation challenges were addressed using data augmentation. The best model indicated averaged deviations comparable to respective experimental flammability limits determination errors. The models also indicated comparable performances in comparison to the empirical models while having a broader range of applicability. This approach seeks to improve flammability limits prediction accuracy and calculation time necessary, minimizing flammability limits determination experimental efforts.