Automatic parallelization is known to an intractable problem in general. This paper is about a new approach in which domain-specific knowledge is used to facilitate automatic parallelization. The research focuses on three widely used numerical methods: the finite difference method (FDM), the finite element method (FEM), and the boundary-element method (BEM). A prototype tool, called the ParAgent, has been developed to study the feasibility of the approach. The current version of the prototype can parallelize Fortran-77 programs based on the explicit time-marching FDM. The paper provides an overview of the new approach and some results of its application including the parallelization of the NCAR/Penn State Mesoscale Meteorology Model MM5. The manual parallelization of MM5 took about three years whereas the parallelization using ParAgent was done in about two weeks.