Session: 01-05-01: Applications of Artificial Intelligence/Machine Learning for Aerospace Structures 1

Paper Number: 152509

152509 - Multiscale Modeling of Lattice Metamaterials Using Machine Learning and Heat Kernel Images 

Lattice metamaterials are lightweight materials with enhanced mechanical, thermal, and acoustic properties. Their microstructures can be customized to meet specific performance requirements across a variety of applications. However, the extensive design space makes predicting material properties based on microstructures computationally expensive. While machine learning (ML) models have been developed to mitigate these costs, different lattice designs often require separate models, which significantly increases computational resources and effort. Additionally, ML models that use direct image data (e.g., convolutional neural networks) provide an alternative, but the pixelated images may fail to accurately capture the complex lattice geometries, leading to compromised prediction accuracy.

To address these computational challenges, this work proposes a novel parameterization method using heat kernel images. This approach converts lattice metamaterials with diverse and complex geometries into a unified format for ML modeling. Four distinct two-dimensional (2D) lattice metamaterials—chiral, re-entrant, hexagonal, and sinusoidal—will be selected for this study. These lattices exhibit unique geometries that cannot be parameterized using a single set of design variables. Python scripts will be developed to automatically generate finite element meshes with periodic boundary conditions (PBCs) for each lattice design. We will use the mechanics of structure genome (MSG) multiscale modeling method to compute the effective material properties of the selected lattices. Additionally, the lattices will be fabricated using a PolyJet 3D printing process, and compression tests will be performed to experimentally measure their material properties, validating the MSG model results. The MSG method provides an efficient and accurate way to generate simulation data for developing an ML model.

The microstructures in the training dataset will be converted into heat kernel images using diffusion geometry theory. These heat kernel images will serve as input data for an ML model to predict the effective material properties. The ML model will be based on the variational autoencoder (VAE) method, mapping key structural features from the input data into a reduced-dimensional latent space. Features in the latent space will then be input into a feedforward neural network to predict the material properties. For comparison, the same ML model will also be trained using direct image data of the 2D lattices to predict the material properties, providing a baseline to demonstrate the improved accuracy achieved with heat kernel images. By incorporating heat kernel images, the proposed ML framework will enhance the generality and applicability of ML models for material modeling, especially when dealing with complex microstructures.

Presenting Author: Xin Liu University of Texas at Arlington

Presenting Author Biography: Dr. Xin Liu is an Assistant Professor in the Mechanical and Aerospace (MAE) Department at the University of Texas at Arlington. He is also a member of the Institute for Predictive Performance Methodologies (IPPM) at the UTA Research Institute. Dr. Liu obtained his PhD in 2020 and Master of Engineering in 2016 from Purdue University in Aeronautical and Astronautical Engineering. His expertise is in data-driven multiscale modeling of advanced materials and structures.