Session F09: Reacting Flows: Modeling and Simulation II

Balancing Transformation for Lightly-damped Dynamics of a Forced Flame using the Eigensystem Realization Algorithm

Increasing applications of reduced-order models (ROMs) in transport problems and the long-standing demand for predictive ROMs encourage further research to reinforce model reduction techniques that leverage theoretical error bounds. The most promising of such methods is balanced truncation (BT) that employs Markov parameters to transform the system to equally controllable and observable coordinates. Stability is preserved under order reduction and the rich dynamical content of the impulse response enables prediction far from the training regime. In this work we utilize BT for model reduction of one-dimensional reacting flow with pressure forcing. Due to the stiffness introduced by the chemical source term we resort to the eigensystem realization algorithm to compute the balancing transformation. We study the sensitivity of the balanced ROMs with respect to sampling properties and show that the system undergoes lightly-damped oscillations following the initial transients, which poses additional computational bottlenecks. We demonstrate the robust performance of BT in comparison to the standard Galerkin and the least-squares Petrov-Galerkin projections in a purely predictive setting.   

Intrusive and Non-intrusive Non-linear Reduced-order Models for Reacting Flows

Reduced-order models (ROMs) offer methods to learn low-dimensional, inexpensive representations of a system's dynamics from a small number of high-fidelity simulations. However, traditional projection-based ROMs, which project the governing equations onto a linear subspace, are typically inadequate to produce accurate models of combusting flows. Artificial neural networks show promise in generating more accurate ROMs for advection-dominated flows by learning a non-linear manifold on which the system dynamics can be described more compactly than on a linear subspace. In particular, autoencoder (AE) ROMs offer a means of learning such non-linear manifolds in an unsupervised fashion. In this work, we develop and apply AE ROMs for advection-dominated reacting flows to create low-cost approximations of reacting flow dynamics. We investigate whether intrusive AE ROMs, which require access and modifications to the governing equations via a numerical solver, provide superior performance over non-intrusive AE ROMs, which only require snapshots of the unsteady flow fields.   

Evaluation of Co-optimized Machine-Learned Manifolds for Modeling Premixed Combustion

Many modeling strategies for turbulent combustion employ a projection of the thermochemical state onto a low-dimensional manifold to reduce computational cost. Frequently, neural networks are applied to compute relevant thermochemical variables from the manifold for both physics-based approaches, such as flamelet generated manifolds (FGM), and for empirical approaches, such as principal component analysis (PCA). Recently, an approach to co-optimize the definition of the manifold and the mapping to (filtered) thermochemical outputs of the model has been proposed and shown to reduce the error associated with the neural network models in either case. In this work, the co-optimized manifolds approach is further evaluated and compared against FGM and PCA for representative premixed turbulent combustion systems. A particular emphasis is ensuring robustness of the neural networks when integrated into practical simulations.    

Deep Reinforcement Learning to Discover Multi-Fuel Injection Strategies for Compression Ignition Engines

Compression ignition (CI) engines typically offer high thermal efficiencies at the expense of harmful gaseous emissions such as NOx. Various strategies for reducing the emissions of CI engines have been developed over the years from exhaust gas after-treatment systems to advanced fuel injection strategies. Some of these strategies involve splitting the fuel injection into multiple pulses (e.g., MCCI) and others leverage multiple fuels with specially designed injection timing to operate in the low temperature combustion (LTC) regime that produces less emissions (e.g., RCCI, DDFS, etc.). In this talk, we show how we can train a deep reinforcement learning (RL) agent to discover multi-fuel, multi-pulse injection strategies for a reduced order model of a CI engine with the objective of maximizing engine work production while maintaining NOx emissions below a certain threshold. Results are shown and discussed for different injection strategies.   

Asynchronous simulations of reacting flows: a path towards exascale computing

Direct Numerical Simulations (DNS) of turbulent combustion at higher Reynolds numbers with detailed reaction mechanisms and at conditions of practical relevance will require efficient utilization of massive computing resources anticipated on the next-generation exascale machines. However, scaling current solvers to these extreme scales requires novel numerical methods and algorithms to avoid performance penalties due to parallelization. A very promising approach is based on so-called asynchrony-tolerant (AT) schemes that can use delayed or asynchronous data at processor boundaries without degrading numerical accuracy, thereby alleviating communication and synchronization bottlenecks. To study stability we developed a new approach that extends (and also highlights important limitations of) the standard von Neumann analysis. We use these AT schemes for asynchronous simulations of several canonical reacting flow problems including auto-ignition and flame propagation. Effects of delayed data on relevant physical processes and the challenging stiff intermediate species is investigated. To further solve problems with shocks and discontinuities, we devised AT-WENO schemes which are validated with DNS data. These first-of-a-kind AT simulations provide a path towards exascale computing.   

Application of Machine learning for turbulent combustion modelling

Application of machine learning to reacting flows is gaining popularity. The use of a Deep Neural Network (DNN) trained on MILD (Moderate, Intense and Low Dilution) combustion data to stratified turbulent flames is explored to estimate sub-grid joint Filtered Density Functions (FDFs) of mixture fraction and reaction progress variable. This objective is achieved by using log of filtered mixture fraction normalised by its stoichiometric value rather than the mixture fraction itself as one the input to the DNN. A good agreement of the estimated FDFs and DNS values is observed for a range of stratified flame conditions despite the use of MILD combustion DNS data for the training phase. Also, filtered reaction rates calculated using FDFs from DNN agree well with DNS data and the root-mean-squared Error is observed to be below 3%.