Gradient-enhanced stochastic optimization of high-fidelity simulations

TL;DR

This paper introduces a gradient-enhanced stochastic surrogate-model based optimization method for high-fidelity CFD simulations, improving convergence and reducing computational costs in flow optimization tasks.

Contribution

The paper proposes a novel gradient-enhanced version of the DYCORS algorithm, integrating gradient information into surrogate models and optimizing radial basis function parameters for better accuracy.

Findings

Gradient-enhanced method improves convergence rate.

Stochastic optimization outperforms gradient-based methods at low Reynolds numbers.

Open-source implementation available.

Abstract

Optimization and control of complex unsteady flows remains an important challenge due to the large cost of performing a function evaluation, i.e. a full computational fluid dynamics (CFD) simulation. Reducing the number of required function evaluations would help to decrease the computational cost of the overall optimization procedure. In this article, we consider the stochastic derivative-free surrogate-model based Dynamic COordinate search using Response Surfaces (DYCORS) algorithm and propose several enhancements: First, the gradient information is added to the surrogate model to improve its accuracy and enhance the convergence rate of the algorithm. Second, the internal parameters of the radial basis function employed to generate the surrogate model are optimized by minimizing the leave-one-out error in the case of the original algorithm and by using the gradient information in the case of the gradient-enhanced version. We apply the resulting optimization algorithm to the minimization of the total pressure loss through a linear cascade of blades, and we compare the results obtained with the stochastic algorithms at different Reynolds numbers with a gradient-based optimization algorithm. The results show that stochastic optimization outperforms gradient-based optimization even at very low $Re$ numbers, and that the proposed gradient-enhanced version improves the convergence rate of the original algorithm. An open-source implementation of the gradient-enhanced version of the algorithm is available.