Chris Anderson - University of California, Los Angeles (UCLA), Mathematics In this talk I'll describe a type of approximation for high dimensional functions that consists of combining known functional forms and/or functional forms with free parameters. The method is specifically developed for creating differentiable approximations based upon training data that is computer generated and thus has low noise. This work was motivated by the challenge problem of developing a surrogate that can provide highly accurate values of two electron integrals -- one of the computationally intensive tasks associated with Ab-Initio quantum mechanical simulations.