Fast and accurate prediction of H2 combustion via data-augmented physics-informed operator learning

Abstract

This study models hydrogen combustion within a zero-dimensional constant volume reactor through a physics-informed operator approach. 11 thermochemical variables are considered, including 9 chemical species along with two thermodynamic state variables. Due to the short lifetime of the intermediate species, these variables differ up to 11 orders of magnitudes, resulting in stiff differential equations. Therefore, a scalable architecture designed to predict numerous thermochemical variables varying in order of magnitudes and nonlinearities is proposed. Moreover, small amounts of pre-labeled training data are incorporated to increase training efficiency. Additionally, a dedicated set of training strategies is introduced to improve both training efficiency and convergence. As a result, the completely trained model accurately predicts time-evolving profiles of thermochemical variables in combustion phenomena at given initial conditions. Furthermore, a quantitative analysis of the correlation between the amount of pre-labeled training data and the accuracy of the model is investigated. Results show that the data-augmented physics-informed model can achieve accuracy comparable to that of a data-driven model trained with up to a fivefold increase in training datasets. Consequently, the physics-informed model showed approximately 1% relative error for temperature and pressure, while showing errors lower than 10-3 for the chemical species concentrations when employing only 25 training datasets.