BRIDGING GAPS IN MULTI-SCALE MATERIALS MODELING WITH MACHINE AND TRANSFER LEARNING
In 2011, the Materials Genome Initiative (MGI) was founded as an effort to unite and drive materials design at an unprecedented pace. By linking computational tools with experimental data, and aligning their data structures to match and interact, scientists across the world have been able to change the way they do science at a fundamental level.
The 3 Mission Statements of the Materials Genome Initiative include: 1) Developing a Materials Innovation Infrastructure 2) Achieving National Goals with Advanced Materials 3) Equipping the Next-Generation Materials Workforce. Since the inception of the MGI the Materials Engineering community has developed numerous cyberinfrastructure repositories for experimental, and varied levels of computational data. This practice aligns with a separate initiative for Findable, Accessible, Interoperable, and Reproducible (F.A.I.R.) principles for data handling and science. By integrating the cyberinfrastructure efforts with continued collaboration from experimental and computational scientists we push the field to evolve improved workflows for research.
This thesis is a collection of applied solutions for materials design with atomistic modeling, and machine learning (ML). In Part 1, we will discuss bridges for the gaps between atomistic simulation and experiment, and what it means for material solutions. A showcase of combining experimental information with ab initio electronic transport calculations will be discussed, as well as the principles of density functional theory (DFT) and molecular dynamics (MD) simulations. In Part 2, our focus will shift to applications of machine learning and the use of composition and chemical featurizers for materials design. Here we leverage cyberinfrastructure efforts with APIs and ML with transfer and active learning for efficient high-dimensional space exploration. In Part 3 local atomic environments and configurations, associative fingerprinting solutions, and workflows for designing deep learning (DL) interatomic potentials for MD are discussed. Finally, a brief section will conclude with efforts made to align with F.A.I.R. principles for Materials Engineering research, and educational development for Mission Statement 3 of the MGI.