The use of detailed chemistry is becoming an essential requirement in CFD simulations of complex reacting flows, such as the Moderate or Intense Low-oxygen Dilution (MILD) combustion regime.

However, detailed kinetics mechanisms often require to solve hundreds of ordinary differential equations per computational cell, causing the need for massive computational resources. The Sample-Partitioning Adaptive Reduced Chemistry (SPARC) methodology, demonstrated to be effective in the speed-up of the chemical step of such costly simulations. This methodology couples adaptive chemistry and machine learning in order to build a library of locally reduced chemical mechanisms in the preprocessing step.

In this work, we present an enhanced version of SPARC which improves a number of critical aspects of the original methodology. In particular, we focus on the automatic selection of target species for reduction, and on the a-priori error estimation of the reduced mechanisms.