Surrogate modeling has become a staple in large-scale scientific computing applications like aerospace and chemical engineering where full evaluations of the model are too expensive to be directly utilized within analysis contexts. However, with the recent advent of pervasive deep learning through differentiable programming, the ability to mix surrogate techniques with neural network code has become a practical issue. Surrogates.jl is a new surrogate modeling library hosted under the DiffEq umbrella which supports differentiable programming to allow for accelerated deep learning. It is compatible with forward and reverse mode automatic differentiation.
My plan is to develop more surrogates methods, such as but not limited to: Compactly supported surrogates, Polynomial expansions, Regularized minimal-energy tensor-product splines, Variable fidelity modeling, Mixture of experts and DENSE.