The work we present deals with the problem to provide learning assistance systems in the context of simulation and modelling. We develop a classification scheme for learning assistance systems and the ir use cases. Beyond this, we discuss how learning from simulation data differs from traditional knowledge discovery from data bases. The discussion contains a classification and review of existing approaches followed by a n enclosing case study of an assistance system for load balancing purpose in FEM simulation. The pres ented application case uses a two -stage architecture to minimize additional computational costs. The approach does not require labeled data in the sense of a quality rating for a load distribution nor a teacher for the initial setup and can improve itself unsupervised. For the FEM simulation on heterogeneous distributed systems we introduce a novel feature set and perform an evaluation for several problem sets.