A randomization and unrolling framework to improve deep neural networks modelling chemistry integration across multiple combustion regimes

Solving finite-rate chemical kinetics through direct integration is computationally expensive and often becomes a performance bottleneck in high-fidelity combustion simulations. Deep neural networks (DNNs) offer a promising alternative by directly modelling integrated chemical source terms. However, conventional training approaches often suffer from error accumulation, especially when modelling detailed chemistry, leading to unreliable performance in a posteriori simulations. In this work, we propose an unrolling training strategy based on randomized states that exposes the network to accumulated errors during training and enables it to observe the evolution of species across varying chemical time scales. The effectiveness of this approach is systematically evaluated in a two-dimensional laminar counterflow flame, which spans ignition, propagation and extinction regimes using the detailed GRI 3.0 mechanism. The influence of different unrolling step numbers on model performance as well as computationally efficiency on both CPU and GPU is also analysed. To demonstrate the method’s generalizability, the trained networks are further assessed in a three-dimensional turbulent planar jet flame with strong turbulence–chemistry interactions.