Surrogate models consist of a Machine Learning model such as a deep neural network that is trained on data previously calculated by a numerical simulation program. This could be a Computational Fluid Dynamics or a FEM application. Assuming the deep neural network is able to generalize well on a broad range of practical cases, then surrogate models are of practical value because inferences on the network are faster than the numerical solution, while they deliver acceptable accuracy. This enables for example a faster what-if analysis, accelerating time-to-market and time-to-solution.
Here is the complete blog.