The recent advances in 3D printing has made it possible to create new materials from naturally available materials. These new materials, also known as architected materials, are built by varying architecture at a sublevel of the actual structure thereby altering the material properties. Our research consists of investigating the finite deformation characteristics of these material for finite deformations using computational homogenization method,. Also we are working on developing efficient machine learning (ML) based multiscale frameworks for doing computational homogenization.
ML assisted multiscale analysis of architected materials
Currently deep neural network (DNN) is being utilized for replacing finite element based microscale analysis within the multiscale analysis. DNN, also known as universal approximators, have proven their capability to learn complex functions in various fields, especially for “big data” problems. Consequently, studies have been done on implementing DNN in solid mechanics problems, including multiscale analysis with positive findings. A standard DNN, fully connected network (FCN), consist of an input layer, hidden layers and an output layer that combines weights and biases from previous layer and passes it into an activation function to generate output. The weights and biases of the DNN are set by training the network, which involves solving an optimization problem.
In multiscale analysis, DNN is used to replace the microscale analysis for homogenized stress prediction; thereby eliminating the need for nested finite element analysis (FEA). However, when limited data is available, governing physics of the problem has to be incorporated in the DNN framework for obtaining good predictions. One way of accomplishing this is by using microscale variational energy as the loss function for the DNN. We are currently working on developing a DNN model for doing homogenization of RVEs made up of hyperelastic material, for finite deformations.