DATA DRIVEN ANALYSIS OF TRUSS STRUCTURES BY DEEP LEARNING

In this manuscript a new approach to optimum weight design of truss structures with discrete variables is described. The new algorithm is presented based on the machine learning technique. There are three steps are needed to train neural networks and produce new design variable candidates. Moreover, three kinds of neural network structures, namely, regression, multi-class classification, and binary classification, are used to compose the system. Once the neural networks are trained, the optimization cycle needs only the stress or displacement constraints. The constraints are then applied on the proposed system to obtain minimum design results. The algorithm is tested on well-known truss structures and is compared with the results of analysis. The numerical examples show that the proposed method is able to obtain minimum design results, and it is totally different way for size optimization problems.