Optimizing progress variables for ammonia/hydrogen combustion using encoding–decoding networks

Abstract

We demonstrate a strategy to optimize parameterizations of combustion manifolds using an encoding–decoding artificial neural network architecture. Our focus in this work is on the combustion of ammonia (NH3) and hydrogen (H2) blends. The literature on NH3 combustion, to date, lacks an efficient definition of a reaction progress variable (PV) to parameterize the thermo-chemical state-space. A quality parameterization should be able to represent the thermo-chemical state variables accurately, as well as any functions of those, e.g., the source terms of the non-conserved PVs. Our approach incorporates information about the reaction source term of a PV and about important combustion products into the PV optimization. A gradient descent optimizer is informed by the reconstruction quality of those important quantities of interest (QoIs) that enter the optimization as decoder outputs. The approach can be thought of as an iterative back-and-forth between defining a parameterization (encoding) and reconstructing QoIs from it (decoding). It thus naturally promotes parameterizations where each QoI is uniquely and smoothly represented over the manifold. This work can help advance the adaptivity of combustion models. First, we show that with an adequate definition of a PV, we can steer the model’s accuracy towards improved representation of selected products and pollutants. Second, the definition of a PV automatically adapts to best complement the remaining physics-based parameters, such as the mixture fraction or the enthalpy defect. These two achievements combined were not possible with the existing PV optimization methods which only impose monotonicity and scalar gradient magnitude in defining a PV.

Novelty and Significance Statement

We demonstrate a novel strategy to optimize the definition of a progress variable (PV) using an encoding–decoding artificial neural network. Our approach can be thought of as an iterative back-and-forth between defining a parameterization of a flame (encoding) and reconstructing important scalars from it (decoding). Notably, the PV definition and its corresponding source term are co-optimized. The definition of a PV automatically adapts to best complement the remaining physics-based parameters, such as the mixture fraction or the enthalpy defect. These achievements were not possible with the existing PV optimization methods which only impose monotonicity and scalar gradient magnitude in defining a PV. This work can help advance combustion models, paving the way for adaptive reduced-order models, where the model can be adjusted towards particularly good representation of target scalars, such as pollutants. Our optimization method is applicable to premixed and non-premixed combustion.