Principal Investigator Paul Barton
Modern experimental techniques in conjunction with quantum computational chemistry are facilitating the construction of detailed chemical kinetic models that can predict accurately the formation and destruction of byproducts and pollutants in processes such as combustion, pyrolysis, and super-critical water oxidation. These detailed ODE/DAE models can involve 100s-1,000s species and 1,000s-10,000s reactions. In recent research we have shown how automatic differentiation technology can exploit the inherent sparsity in these large-scale chemical kinetic models to reduce the computational cost required for simulation and sensitivity analysis dramatically.
It is often desirable to embed these chemical kinetic models in reacting flow simulations where it is necessary to model the chemistry at a large number of spatial grid points associated with the semi-discretization of partial differential equations. The need to repeat the large-scale chemical kinetic model at every grid point can easily overwhelm state-of-the-art algorithms running on advanced computing architectures. It is therefore necessary to consider approaches for reducing the size of the chemical kinetic model while still retaining accuracy in the simulation results. We are developing a fully automated local kinetic model reduction procedure that deletes reactions and/or species from the model using a mixed-integer optimization formulation. This procedure can guarantee finding the smallest possible kinetic model that satisfies user specified error tolerances. In addition, because the model reduction procedure is based on local information, we are developing an interval analysis based algorithm that can generate rigorous regions of validity for these reduced models (i.e., the reduced model satisfies error tolerances in a region of composition/temperature space). This is based on a new approach for solving semi-infinite optimization problems, also being developed in this project.