Evaluation of deep learning networks for predicting truss topology optimization results

Abstract

The applicability of artificial neural networks (ANNs) on the prediction of the structural optimization results of a truss structure is investigated. Two different ANN architectures are employed and the effect of using various optimizers and activation functions on their prediction performance is evaluated. Unlike the traditional machine learning network strategies where usually a physical response of the truss optimization (such as compliance, stress, etc.) is predicted, in this study, a new way of prediction is utilized for the truss-like structures; particularly predicting the optimized thickness values of the strut members by the ANNs. Thus, the input data used in these networks are the thickness values of the strut members at a certain initial iteration while the optimized thickness values are predicted as the outputs. A cantilever beam example is presented for the truss optimization to show the efficacy of the presented ANNs. The results indicate that using the thickness values at a certain initial iteration as inputs and final iteration thicknesses as outputs in ANNs are promising for the structural optimization prediction of the presented truss problem with the appropriate selection of the architecture, optimizer, activation function, and the input-output data formation.