Search for: All records

ABSTRACT In this paper, we present an interface‐filtering structural optimization method that optimizes structural shape and topology through successive interface movements. This interface filtering is achieved via the combination of the variable‐radius‐based partial‐differential‐equation (PDE) filtering and the Heaviside projection on a density representation. In the proposed method, designs are represented by a density field with sharp interface and no internal grey features, and a filter radius field is used as the design variable in the optimization process. With this method, any density distribution with sharp interfaces can be used as initial designs, and sharp density contrast in density distribution is preserved throughout the optimization process. An analytical relation between the maximum movements of interfaces and the maximum filter radius is given, so that the interface movement can be controlled during the optimization process. Sensitivities with respect to filter radius variables are derived. Two numerical treatments, involving the density update scheme and the radius re‐initialization scheme, are developed to achieve smooth successive shape updates and avoid artificial local minima. Numerical examples, including geometric deformation problem, structural compliance minimization, thermal compliance minimization, and negative Poisson ratio problem, are presented to demonstrate the capabilities of the proposed method. 

Abstract This work proposes a combined deep learning based approach to improve thermal component heat sinks involving turbulent fluid flow. A Generative Adversarial Network (GAN) is trained to learn and recreate the new ellipse based heat sinks. Simulation data for new designs is efficiently generated using OpenFOAM 7 (Open Source Computational Fluid Dynamics software) along with high throughput computing. To improve the speed of design evaluation, a Convolutional Neural Network (CNN) is trained to predict the entire temperature field for a given design. The trained CNN is able to predict the entire temperature field for the design with a mean average error of 1.140 degrees kelvin in 0.04 seconds (22,500 times faster than the simulation). A combined model is formed using the trained CNN and GAN networks to create and simulate new designs. The combined model optimizes the latent representation of 64 random designs on a Graphical Processing Unit (GPU) in ten minutes. The optimized designs perform fourteen degrees kelvin better on average than the non-optimized designs. The highest preforming design outperforms any design in the training data by 1.83 degrees kelvin.