On-the-fly uncertainty-aware deep ensembles for stiff chemical kinetics

Chemical kinetics calculations present a major computational bottleneck in combustion simulations during the direct integration (DI) of thermochemical differential equations. In this work, deep ensembles (DE) are presented as a data-driven approach for predicting chemical source term in reacting flow simulations, aiming to replace the stiff CVODE solver traditionally used for computing the temporal evolution of thermochemical species in detailed chemistry simulations. The training data is generated using Latin Hypercube Sampling (LHS) for initial conditions in zero-dimensional (0D) constant pressure autoignition simulations, with inputs and outputs preprocessed for training. Then, ensembles are trained on datasets that capture the full range of thermochemical states in autoignition processes. Two test cases of increasing complexity are investigated: H 2 /air and NH 3 /H 2 /air mixtures, each exhibiting varying degrees of chemical stiffness and non-linear behaviour. The study shows that iterative predictions align well with DI results, while providing uncertainty estimates that grow with the prediction horizon. The DE approach achieves up to ≈ 9.4-fold computational speed-up compared to the CVODE solver while maintaining accuracy through the potential for uncertainty-based solver selection, where uncertainty estimates serve as a potential criterion for switching back to the CVODE solver when exceeding predetermined thresholds. This offers a foundation for future implementation of an adaptive solver selection mechanism.