A latent variable framework for handling detailed chemistry in reacting flows

Combustion involves multiple scales and intricate physical processes that make its simulation extremely demanding in terms of computational resources. To overcome this challenge, reduced-order models are essential for improving our understanding of the process and speeding up simulations. This work introduces a latent variable (LV) transport framework and demonstrates its compression capabilities through 0D batch reactor simulations. To improve solver efficiency and numerical stability, we introduce projected tolerances, which scale the ODE solver tolerances in the latent space, significantly improving numerical stability and computational efficiency compared to traditional fixed tolerance approaches.The methodology is validated on 0D simulations using methane, propane, and n-heptane combustion with detailed and lumped kinetic mechanisms. Results show that the LV solver achieves substantial computational cost reductions (up to 50%) while preserving solution accuracy. We further analyse the influence of kinetic mechanisms, demonstrating that detailed mechanisms benefit the most from latent space compression due to their higher dimensionality.The proposed LV framework offers a robust and efficient alternative to conventional solvers for detailed kinetics and establishes a foundation for its extension to multidimensional reactive flow simulations.