The use of partially converged simulations in building surrogate models

• We propose here to use partially converged data to construct a surrogate model in solid mechanics.
• We use a correction principle (evofusion) of partially converged data to improve the quality of the surrogate model.
• Two strategies are presented, one to construct a complete precise surrogate model, the other one to localize the minimum area.
• It is shown that for the two strategies, the use of partially converged data enables increases in computing performance.

The main objective of this paper is to propose an optimization strategy which uses partially converged data to minimize the computational effort associated with an optimization procedure. The framework of this work is the optimization of assemblies involving contact and friction.Several tools have been developed in order to use a surrogate model as an alternative to the actual mechanical model. Then, the global optimization can be carried out using this surrogate model, which is much less expensive. This approach has two drawbacks: the CPU time required to generate the surrogate model and the inaccuracy of this model.In order to alleviate these drawbacks, we propose to minimize the CPU time by using partially converged data and then to apply a correction strategy. Two methods are tested in this paper. The first one consists in updating a partially converged metamodel using global enrichment. The second one consists in seeking the global minimum using the weighted expected improvement. One can achieve a time saving of about 10 when seeking the global minimum.