Surrogates for Computationally Expensive and Failure-Prone Simulators

Simulations used for navigating design spaces or finding optimal design points and Pareto fronts require accuracy and resolution to guide the designers towards effective decisions. High-fidelity or high-resolution simulators are computationally expensive. Due to inconsistent parameter settings and physically invalid outcomes, such simulators can also fail to return a solution during design automation or optimization loops. In CPS systems, failures can be expected when there is a hard execution-time deadline on the simulator for producing a solution. Time deadlines imposed by other considerations drive simulation failure behaviors in such cases.Current strategies for modeling simulator failure incorporate failures as an additional discrete parameter or entirely disregard the failed point. This paper describes a method for classifying candidate design points as predictable or failure-prone using a Bayes Classifier. Either a global surrogate or a mixture of the expert local surrogates will be identified in design spaces where the high-fidelity simulators yield a prediction. The developed technique is illustrated with two geometry assembly examples. The simulator fails if the composition does not lead to a valid 3D geometry or produces collisions in the assembly. This paper shows that global or MOE surrogates can be trained for both these cases with validation accuracy exceeding 90 %. The results also show that the best surrogate model can be a global model or a mixture of experts models and can vary by the approximated output parameter.