The execution time of many multi-scale simulations is dominated by the frequent evaluation of computationally expensive functions. The use of a scientific, on-line database for function approximation can significantly reduce the computational demands for such applications. For example, in previous work, sequential database implementations have have been developed that have proven effective when used in complex combustion simulations. We review these sequential algorithms and present an extension of these algorithms to parallel computing environments. Our parallel algorithm is based on a global partitioning of the search space and its BSP tree representation. For a representative combustion application, we present experimental results that detail the trade-off between minimizing interprocessor communication and the number of function evaluations. We introduce a new hybrid strategy that attempts to achieve a good load balance while minimizing the number of redundant function evaluations. Our results indicate good efficiencies and function retrieval rates when compared to running a a separate version of the sequential database on every processor.