SciML/Surrogates.jl

What it solves

It provides a way to approximate computationally expensive simulations (such as solving PDEs or performing complex numerical linear algebra) with faster, cheaper surrogate models. This allows users to optimize functions that would otherwise be too costly to evaluate repeatedly.

How it works

The package implements a three-step process: selecting samples from the original expensive function, constructing a surrogate model that mimics the function's behavior based on those samples, and then performing optimization on that surrogate.

Who it’s for

Researchers and engineers who need to optimize expensive mathematical simulations or functions where each evaluation takes significant time and computational resources.

Highlights

• Supports a wide variety of surrogate models, including Kriging, Neural Networks (via Flux), Gradient Boosted Trees (via XGBoost), and Support Vector Machines (via LIBSVM).
• Includes multiple optimization methods such as SRBF, LCBS, EI, and SOP.
• Integrates with QuasiMonteCarlo.jl for sample selection.