Abstract:This study proposes a clustered dynamic adaptive chemistry (CDAC) method, which uses an iterative K-means algorithm to partition the computational cells into different groups according to their similarity in terms of the temperature and significant species compositions. Taking advantage of the clustered cells, the averaged thermo-chemical properties of the cells in the respective clusters are then used to identify the adaptive dynamic reduced chemistry through the method of direct relation graph with error propagation (DRGEP). Moreover, the integration of the chemical source term, which commonly dominates the computational effort in reactive flow simulations, is now performed using the dynamic adaptive reduced chemistry at the cluster level instead of the cell level. With this CDAC method, the on-the-fly DRGEP process as well as the chemistry integration only needs to be conducted at the cluster level, dramatically reducing the unnecessary repeated computation for similar computational cells. In addition, the adaptive dynamic reduced chemistry further accelerates the chemistry integration process due to less ordinary differential equations (ODEs) to be solved for each cluster. This newly proposed CDAC method was tested in multi-dimensional homogeneous charged compression ignition engine (HCCI), direct injection compression ignition (DICI) engine and constant volume chamber combustion (CVCC) fueled with diesel, biodiesel and kerosene through the use of their respective surrogate fuel mechanisms under different operating conditions. The performance of CDAC in terms of accuracy and efficiency were extensively analyzed and discussed using different user-defined parameters, mechanisms with different surrogate components and number of species as well as different meshes with different number of grid cells. Based on this analysis, the error tolerances in DRGEP and the error tolerances of the temperature and significant species' mass fraction are recommended as: epsilon(d) <= 0.001, epsilon(T) 20K and epsilon(Y) <= 0.01 to achieve less than 0.1% integral error. With these recommended user-defined tolerances, it can be observed that the current CDAC method is able to accurately predict the in-cylinder pressure as well as the species profile in HCCI, DICI and the flame lift-off length in CVCC when compared with the full chemistry calculations as well as the experimental results. Moreover, with the recommended user defined error tolerances and a chemical kinetic model of 112 species, this CDAC method is capable of achieving a computational time speed-up factor of more than 3 compared to the conventional DAC and of almost 5 compared to the full chemistry. Finally, the CDAC method coupled CFD was applied to simulate a diesel DICI engine under three different engine speeds with a detailed primary reference fuel (PRF) mechanism of more than 1000 species. The 3-D simulation could be finished in acceptable CPU time while being able to well capture the experimental in-cylinder pressure and heat release rate for all the three engine speeds. (C) 2018 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

Deep learning dynamical latencies for the analysis and reduction of combustion chemistry kinetics

A Robust Autoregressive Long-Term Spatiotemporal Forecasting Framework for Surrogate-Based Turbulent Combustion Modeling Via Deep Learning

Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics

An embedded deep learning model discrepancy for computational combustion simulations

An Improved Approach Towards More Robust Deep Learning Models for Chemical Kinetics

A Deep Learning-Based ODE Solver for Chemical Kinetics.

A deep learning-based model reduction (DeePMR) method for simplifying chemical kinetics

Combustion chemistry acceleration with DeepONets

Dynamic Adaptive Chemistry with Mechanisms Tabulation and in Situ Adaptive Tabulation (ISAT) for Computationally Efficient Modeling of Turbulent Combustion

DLODE: a deep learning-based ODE solver for chemistry kinetics

A New Dynamic Adaptive Chemistry Reduction Method

A Framework for Combustion Chemistry Acceleration with DeepONets

Computational Acceleration of Multi-Dimensional Reactive Flow Modelling Using Diesel/biodiesel/jet-Fuel Surrogate Mechanisms Via a Clustered Dynamic Adaptive Chemistry Method

Simplifying Ignition Delay Prediction of Chemical Kinetic Models by Means of Ignition Correlations

A New Data-Driven Sparse-Learning Approach to Study Chemical Reaction Networks

Deep Learning Reaction Network: a machine learning framework for modeling time resolved data

Kinetics Parameter Optimization via Neural Ordinary Differential Equations

Learning stiff chemical kinetics using extended deep neural operators

On sparse identification of complex dynamical systems: A study on discovering influential reactions in chemical reaction networks

Advancing Reacting Flow Simulations with Data-Driven Models

DeepFlame: A deep learning empowered open-source platform for reacting flow simulations