Surrogate models

Definition : Surrogate models

bivariate surrogate model

Surrogate models , or metamodels, are compact scalable analytic models that approximate the multivariate input/output behavior of complex systems, based on a limited set of computational expensive simulations.
Surrogate models mimic the complex behavior of the underlying simulation model, and can be used for design automation, parametric studies, design space exploration, optimization and sensitivity analysis.
Surrogate models are also called response surface models (RSM), emulators, auxiliary models, repro-models, metamodels, etc.

Background

Parameterized scalable computer models are increasingly important as a means of exploring the behavior of complex systems, and for optimization and sensitivity analysis. Application areas include bioinformatics, operations research, network simulation, Electronic Design Automation (EDA), ecological modeling, and many others.

For example, an engineer may explore the behavior of an airplane wing by running a computational model of the air flow multiple times while varying key design parameters such as length, angle, etc. The results of these multiple experiments yield a picture of how the wing behaves in different parts of the design space.

Thorough understanding of the relationship between design parameters and performance over the entire design space is important for optimal design, to reduce the number of design iterations, to lower the costs, and to improve overall quality.

Usually, for parametric computational experiments, the user selects a huge amount of simulation experiments, covering the whole design space of interest, by defining all parameter ranges and sample distributions (min, max, step size). Note that the cross product of the parameter ranges does not guarantee an accurate and efficient coverage of the design space, as certain regions might be under-sampled while others might be over-sampled.

Research goals

We study and develop fully automated sequential metamodeling techniques for efficient design space exploration, which minimize the overall number of computational expensive simulations.

This research is based on, or linked with:
  • Experimental Design / Computer-aided Design
    [Design Of Experiments (DOE), Response Surface Modeling (RSM), Design and Analysis of Computer Experiments (DACE), Kriging methods, Metamodeling, Data-driven information processing]

  • Numerical techniques
    [Data-driven modelling, Multivariate interpolation and approximation, Rational functions, Radial Basis Functions (RBF), Scattered data interpolation, Orthonormal bases for parameter estimation, Model Order Reduction (MOR), Model Based Parameter Estimation (MBPE), Optimization]

  • Artificial Intelligence (AI)
    [Computational Intelligence, Machine Learning (ML), Artificial Neural Networks (ANN), Genetic Algorithms (GA), Evolutionary Computing (EC), Adaptive/Sequential sampling, Reflective Exploration (RE), Knowledge discovery]