In this paper we investigate how to obtain high-level adaptivity on complex scientific applications such as Finite Element (FE) simulators by building an adaptive version of their computational kernel, which consists of a sparse linear system solver. We present the software architecture of FEMS, a parallel multifrontal solver for FE applications whose main feature is an install-time training phase where adaptation to the computing platform takes place. FEMS relies on a simple model-driven mesh partitioning strategy, which makes it possible to perform efficient static load balancing on both homogeneous and heterogeneous machines.