| Summary: | The drive to develop materials with superior performance and multifunctional capabilities has underscored the importance of precisely controlling material heterogeneity, establishing it as a pivotal component of materials science and engineering. From a compositional standpoint, the materials-by-design paradigm incorporates multiple constituent phases to combine strengths of each individual components, giving rise to a general concept known as "composite", which applies across scales from the microscale to the macroscale. Introducing heterogeneity into materials can also be achieved by integrating structural variations into their design. By altering topological configurations or embedding structural defects, there exists substantial potential for tailoring materials properties and behaviors. However, the compositional and structural heterogeneity in materials complicates the calculation of their properties and extensively expands their design space. As a result, it poses challenges to conventional computational and experimental methods, limiting their effectiveness in fully exploring and exploiting the potential of heterogeneous materials.
In this dissertation, we aim at addressing these challenges by combining multiscale modeling techniques and machine learning (ML) models to accelerate property calculation and design of heterogeneous materials. We begin with developing artificial intelligence (AI)-based surrogate models for building multiscale "structure-to-physical field" linkage. At the continuum level, our work highlights the prediction of strain/stress fields in complex hierarchical composites through a conditional generative adversarial network, bypassing conventional numerical simulations such as the finite element method (FEM). At the atomic scale, we utilize a graph neural network based method to bridge structural defects and the distribution of atomic properties in mesoscale crystalline solids. Compared to traditional atomistic molecular dynamics (MD) simulations, our proposed method demonstrates enhanced efficiency and broad applicability.
Apart from addressing forward problems (from structure to property), we also showcase the utility of AI-based methods in addressing the notoriously challenging inverse problem (from property to structure). These inverse problems, characterized by limited information from observed data, present significant challenges due to the absence of governing equations or constitutive relations that can be directly solved with multiscale modeling. We introduce a framework that integrates multiple deep learning (DL) models to tackle a typical inverse problem efficiently, circumventing the need for costly iterative methods and provides new solutions to addressing ill-posed inverse problem where multiple solutions may exist for a single observation.
To further design heterogeneous materials, we employ high-throughput screening technique to streamline the generation and testing of materials and structures of our interest. Our investigations encompass a wide range of materials systems, from 3D graphene-based foams showcasing structural heterogeneity to polypeptide self-assemblies exhibiting compositional heterogeneity. We leverage both atomistic and coarse-grained (CG) MD simulations to generate thousands of different data points for each material and train ML or DL models with these created datasets. This high-throughput approach automates the creation of nanoscale structures, facilitates the assessment of their mechanical properties, and accelerates the identification of promising candidates. We believe that our approach not only significantly accelerates the pace of discovery and optimization in material science but also opens new avenues for the creation of innovative materials with tailored properties, marking a pivotal advancement in the field of heterogeneous material design.
|
|---|