Local combustion regime identification using machine learning

A new combustion regime identification methodology using the neural networks as supervised classifiers is proposed and validated. As a first proof of concept, a binary classifier is trained with labelled thermochemical states obtained as solutions of prototypical one-dimensional models representing premixed and nonpremixed regimes. The trained classifier is then used to associate the regime to any given thermochemical state originating from a multi-dimensional reacting flow simulation that shares similar operating conditions with the training problems. The classification requires local information only, i.e. no gradients are required, and operates on reduced-dimension thermochemical states, in order to cope with experimental data as well. The validity of the approach is assessed by employing a two-dimensional laminar edge flame data as a canonical configuration exhibiting multi-regime combustion behaviour. The method is readily extendable to additional classes to identify criticality phenomena, such as local extinction and re-ignition. It is anticipated that the proposed classifier tool will be useful in the development of turbulent multi-regime combustion closure models in large scale simulations.