Enhancing Hydrogen Laminar Burning Velocity Estimation With Artificial Neural Networks and Simulated Data

Abstract Practically relevant containment scale combustion simulations require fast methods applicable to comparatively sparse meshes, e.g. RANS. Combustion rate in this case can be estimated using simplified progress variable equation closures, for example, Turbulent Flame speed Closure model, and correlations for laminar burning velocity (LBV) instead of chemical kinetics. However, existing correlations may lack accuracy and have restricted validity domains. In the previous work, by developing experimental-data-based Artificial Neural Networks (ANN) model (original), we showed that ANN can be used for fast and accurate prediction of LBV in dry hydrogen-air mixtures from temperature, pressure and equivalence ratio. However, due to limited amount of experimental data available in the literature, database consisted of a relatively low number of experimental datapoints (around 2000). This led to model exhibiting some unphysical behaviour at higher temperatures and specific pressures, requiring more data than is available to enhance its reliability. This work presents an attempt to solve the problem of insufficient database by replacing simulated data instead of experimental. The architecture of new ANN was kept similar to original ANN, ensuring similar prediction time. Instead of a few thousand experimental data points, new model was trained using around 1 000 000 data values generated using chemical kinetics simulation. The larger database improved new model predictions compared to original and led to a more physical LBV behaviour estimation. In addition, new model has an expanded range of applicability. In our estimation, any further improvement in our model accuracy is mostly limited by the number of neurons, restricted by the need for fast ANN calculations.