Data-Driven Exploration of Spinodoid Materials for Mechanical and Thermal Properties

TR Number

Date

2026-09-11

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Spinodoids are a new class of architected materials with smooth, spatially correlated topolo gies, inspired by spinodal decomposition, a phase separation phenomenon. Generated stochas tically through a mathematical formulation based on Gaussian Random Fields (GRFs), they enable an extensive design space covering both periodic and non-periodic architectures. Their smooth, low-curvature geometric features offer enhanced material response and highly tun able anisotropic properties compared to traditional surface- or strut-based architected ma terials, making them promising candidates for engineering applications. This dissertation explores the design spaces of spinodoids for mechanical and thermal properties using nu merical and computationally efficient data-driven frameworks. In Chapter 2, a numerical and data-driven design study is carried out to explore the optimum design parameters of two-dimensional (2D) metal-based porous spinodoids for improved mechanical performance. In the proposed approach, spinodoids are designed by parameterizing the orientation vectors defined in the GRF formulation. Finite element analysis (FEA) is performed to compute their elastic stress-strain responses, which are compared to those of traditional truss-based lattice structures. The strain energy density function is calculated under tensile and shear deformations, and an optimization problem is formulated to minimize strain energy frac tions. This problem is solved with the pattern search algorithm using the physics-based framework and a deep-learning-based surrogate model. Chapter 3 then presents a numer ical homogenization and design exploration framework for the thermal conductivities of ceramic composite spinodoids. The orientations of the spinodoids are designed to achieve enhanced homogenized thermal conductivity along different directions. For this purpose, their anisotropic thermal conductivities are determined via steady-state thermal conduc tion simulations combined with a volume-averaging method. A physics-based optimization framework is then implemented to perform topology optimization. Exploring the design of these architected materials with conventional physics-based methods becomes computation ally prohibitive when multiphysics properties must be evaluated. To address this, Chapter 4 presents a computationally efficient data-driven framework for exploring and optimizing the mechanical and thermal properties of dual-phase metallic spinodoids. A convolutional neural network (CNN) is developed to predict homogenized elastic and thermal properties by relating images of 2D spinodoids to properties computed via FEA. Next, the CNN model is coupled with a global optimization solver to design dual-phase spinodoids with improved mul tiphysics objectives. Finally, Chapter 5 presents a deep learning-assisted surrogate modeling framework for the inverse design of spinodoids with targeted anisotropic thermal conduc tivity. The GRF-based formulation is extended to three-dimensional (3D) representations of porous spinodoids, which are then modeled as representative volume elements (RVEs) and analyzed via FEA to extract homogenized thermal conductivities in three orthogonal directions. The resulting simulation dataset is used to train a feed-forward neural network (FNN) model that serves as a surrogate for predicting homogenized thermal properties. This surrogate replaces computationally expensive numerical simulations and is embedded within a genetic algorithm (GA) to perform efficient inverse design of spinodoids with prescribed thermal properties.

Description

Keywords

Optimization, finite element analysis, deep learning, architected materials, materials design, inverse design, homogenization

Citation