An adaptive-learning framework for chemistry tabulation in turbulent reacting flows

Many applications in turbulent combustion rely on some form of reduced-order model. Flamelet-type models with pre-computed, tabulated chemistry are widely used and often perform well. However, the creation of an appropriate and accurate table requires considerable knowledge and time. Data-driven models have shown a lot of promise in reducing this workload. They are fast, simple to use, and can provide high accuracy. But the model is often a black box. Moreover, data-driven extrapolation to new, unknown data is not obvious; a feature required in most applications.In this work, we implemented the RONAALP algorithm (Reduced Order Nonlinear Approximation with Active Learning Procedure) which combines offline training with an adaptive learning approach to handle extrapolation events on-the-fly. The offline training relies on a combination of nonlinear auto-encoders, community clustering, and radial basis function networks to develop an efficient and compact base model from limited training data. The active learning procedure expands the base model to consider extrapolation outside the original training data. The algorithm was developed and validated for non-equilibrium hypersonic flows.We have adapted the algorithm to the turbulent reacting wall-bounded flows typical of gas turbine combustion. The model is trained on inexpensive 1D laminar flame data and assessed using a turbulent 3D DNS database.The base model performs as well as conventional chemistry tabulations methods. Due to the built-in extrapolation procedure, adaptive learning with additional training data from strained counterflow flamelets and head-on quenching flamelets improves the predictions.