Artificial Neural Networks for Predicted Bending Properties of Additively Manufactured Pyramidal Lattice Truss Core Sandwich Structures

Abstract

An Artificial Neural Network (ANN) model is developed to predict the mechanical behavior of pyramidal lattice truss core sandwich structures under bending load. The development process aims to optimize material use, enhance structural efficiency, and reduce analysis time for the developed ANN model. Key phases include specimen fabrication via additive manufacturing, experimental testing in four-point bending, and validation of the finite element model (FEM). Experimental tests on five specimens validated FEM simulations with a 4.5 % error rate. The ANN, trained on FEM data, accurately predicts reaction forces and stress components (sigma,, sigma 2, tau,2). Comparison of training algorithms (LM, Levenberg-Marquardt, BR, Bayesian Regularization, SCG, Scaled Conjugate Gradient) highlights LM's superior performance in convergence and MSE reduction (max. MSE value: 2.287), while BRexcels in generalization and robustness. Scaled Conjugate Gradient's performance was lower than the others. The ANN demonstrates high accuracy within the training range but shows limitations in extrapolation. Overall, this ANN model offers engineers a rapid and precise tool for predicting the mechanical behavior of these sandwich structures, reducing reliance on time-consuming FEM simulations and facilitating efficient design optimization across various engineering applications.