Real-Time Structural Topology Optimization Using Deep Learning Methods

Traditional topology optimization methods rely on time-consuming iterative algorithms that lead to an exponential increase in the computational time as the degree of background grid freedom increases,the so-called "dimensional curse" problem.In order to solve this problem,based on the convolutional neural network and conditional generation confrontation network in deep learning,we propose two new real-time topology optimization methods.This method gives more accurate optimization results based on displacement and load boundary conditions in near real time.Topology-optimized real-time deep learning methods.(1)A real-time topology optimization method based on symbolic representation and deep convolutional neural networks is proposed.The method replaces the topological design variables with two-dimensional images,and then uses the volume and neural network to train,and then obtains a convolution model of the real-time generated topology.Abandoning the traditional iterative algorithm for topology optimization can significantly reduce the computational source.The 25000 optimization structure is generated as a training data set by the SIMP algorithm using the open source topology optimization code.We use different symbols to represent Dirichlet boundary conditions,Neumann boundary conditions and material properties.Numerical examples show that our convolutional network requires about 10 ms to generate static topology optimization,while the traditional SIMP method requires 80 s,which shows that our method can significantly reduce computation time.(2)We propose a real-time topology optimization method based on conditional generation adversarial networks.Firstly,the simulation data training set corresponding to the topology design variable and the topology structure is prepared,and then the conditional generation adversarial confrontation network and the Pix2 pix network are used for training,and then the generation confrontation model of the real-time generated topology structure is obtained.This method adopts the denoised CWGAN-GP model,which uses the generation to resist the powerful generation ability and convergence ability to quickly generate the topology structure,and then uses the Pix2 pix model to improve the image clarity of the topology structure,which could obtain high definition,deblurring and smooth edges.At present,both methods can predict the real-time structure of the topology without relying on tedious and time-consuming iterative calculations.Our research shows that deep learning can achieve near real-time topology optimization,which has important theoretical and practical value for the rapid design of automotive lightweight and the combination of deep learning and mechanics.