(679a) A Computational Methodology for Learning Low-Complexity Surrogate Models of Processes From Experiments or Simulations

Costly and/or insufficiently robust simulations or experiments can often pose difficulties when their use extends well beyond a single evaluation. This is case with the numerous evaluations of uncertainty quantification, when an algebraic model is needed for optimization, as well as numerous other areas.   To overcome these difficulties, we generate an accurate set of algebraic surrogate models of disaggregated process blocks of the experiment or simulation.  We developed a method that uses derivative-based and derivative-free optimization alongside machine learning and statistical techniques to generate the set of surrogate models using data sampled from experiments or detailed simulations. 

Our method begins by building a low-complexity surrogate model for each block from an initial sample set.  The model is built using a best subset technique that leverages a mixed-integer linear problem formulation to allow for very large initial basis sets.  The models are then tested, exploited, and improved through the use of derivative-free solvers to adaptively sample new simulation or experimental points.  The sets of surrogate models from each disaggregated process block are then combined with heat and mass balances around each disaggregated block to generate a full algebraic model of the process.  The full model can be used for cheap and accurate evaluations of the original simulation or experiment or combined with design specifications and an objective for nonlinear optimization.