A novel deep unsupervised learning-based framework for optimization of truss structures

In this paper, an efficient deep unsupervised learning (DUL)-based framework is proposed to directly perform the design optimization of truss structures under multiple constraints for the first time. Herein, the members' cross-sectional areas are parameterized using a deep neural network (DNN) with the middle spatial coordinates of truss elements as input data. The parameters of the network, including weights and biases, are regarded as decision variables of the structural optimization problem, instead of the member's cross-sectional areas as those of traditional optimization algorithms. A new loss function of the network model is constructed with the aim of minimizing the total structure weight so that all constraints of the optimization problem via unsupervised learning are satisfied. To achieve the optimal parameters, the proposed model is trained to minimize the loss function by a combination of the standard gradient optimizer and backpropagation algorithm. As soon as the learning process ends, the optimum weight of truss structures is indicated without utilizing any other time-consuming metaheuristic algorithms. Several illustrative examples are investigated to demonstrate the efficiency of the proposed framework in requiring much lower computational cost against other conventional methods, yet still providing high-quality optimal solutions.

Automated summary

This paper proposes an efficient deep unsupervised learning-based framework for the design optimization of truss structures. By parameterizing the members' cross-sectional areas using a deep neural network, the framework aims to minimize the total structure weight and satisfy all optimization constraints. Through training the model with a combination of gradient optimization and backpropagation algorithm, the optimal weight of truss structures can be obtained without using other time-consuming metaheuristic algorithms.