Determination of Aggregate Elastic Properties of Powder-Beds in Additive Manufacturing Using Convolutional Neural Networks

摘要

The most popular strategy for the estimation of effective elastic properties of powder-beds in Additively Manufactured structures (AM structures) is through either the Finite Element Method (FEM) or the Discrete Element Method (DEM). Both of these techniques, however, are computationally expensive for practical applications. This paper presents a novel Convolutional Neural Network (CNN) regression approach to estimate the effective elastic properties of powder-beds in AM structures. In this approach, the time-consuming DEM is used for CNN training purposes and not at run time. The DEM is used to model the interactions of powder particles and to evaluate the macro-level continuum-mechanical state variables (volume average of stress and strain). For the Neural Network training purposes, the DEM code creates a dataset, including hundreds of AM structures with their corresponding mechanical properties. The approach utilizes methods from deep learning to train a CNN capable of reducing the computational time needed to predict the effective elastic properties of the aggregate. The saving in computational time could reach 99.9995% compared to DEM, and on average, the difference in predicted effective elastic properties between the DEM code and trained CNN is less than 4%. The resulting sub-second level computational time can be considered as a step towards the development of a near real-time process control system capable of predicting the effective elastic properties of the aggregate at any given stage of the manufacturing process.