A manifold learning approach to model reduction in combustion

We use a relatively recent nonlinear manifold learning technique (diffusion maps) to parameterize low dimensional attracting manifolds arising in the description of detailed chemical kinetics mechanisms. With no a priori knowledge about the shape and dimension of the manifold, such an approach provides a way of solving a reduced (and less stiff) set of equations in terms of automatically detected slow variables. Advantages as well as disadvantages of the approach are discussed.