Abstract

Optimization of complex structures using standard evolutionary algorithms, like genetic algorithm (GA), is known to be computationally expensive, because a very large number of finite element analysis (FEA) must be conducted for each possible structural design during the optimization. In the literature, different pretrained surrogate models have been used to replace the FEA for improving the computational performance of the structural optimization. However, the optimal solution found by the optimizer often depends on the accuracy of the pretrained surrogate model itself. Moreover, large datasets may be needed for getting the required accuracy of the surrogate model. An adaptive machine learning technique called active learning is used in this paper to accelerate the evolutionary optimization of complex structures. An active learner is a machine-learning-based model that can interactively query the outputs of certain data points, whenever the model would be uncertain about those outputs. In the presented approach, the active learner helps the GA by predicting if the possible design is going to be feasible or infeasible, meaning if it satisfies the required constraints or not. If the active learner is uncertain about the output, an actual FEA is conducted, and it improves its own accuracy for future evaluations. The approach does not need a trained surrogate model before the optimization. The active learner adaptively learns about the structure during the optimization to improve the computational performance. The approach is used to optimize the classic 10-bar truss problem, the Hesse function, curvilinearly stiffened panels under buckling and stress constraints, and some benchmark constrained test functions. The results show that the approach has the potential to reduce the total required constraint evaluations by more than 50%.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call