Abstract
Design of microarchitected materials (Materials-by-Design) seeks to identify material structures that achieve desired performance, but in practice this process often remains slow, sparse, and experience-guided. Candidate microstructures are typically proposed from prior knowledge, evaluated through experiments or high-fidelity simulations, interpreted by domain experts, and revised through an iterative trial-and-error cycle. While this workflow is essential for material discovery and validation, it is difficult to scale. Microstructure spaces are high-dimensional, each physical evaluation can be costly, and human-guided exploration can examine only a limited portion of the possible design space. Consequently, promising candidates may remain unexplored, and selected microstructures may reflect the constraints of the search process rather than the broader structure of the design space. This dissertation develops an AI-assisted closed-loop framework for microstructure characterization toward target performance. The framework is organized around three barriers that limit conventional MbD workflows. The first is the representation barrier: microstructures must be expressed in a form that is compact enough for search while preserving performance-relevant geometric detail. The second is the structure–property–performance (SPP) evaluation barrier: candidate responses must be rapidly approximated to support broad exploration. The third is the systematic search barrier: the design space must be explored strategically rather than through unguided trial-and-error. Addressing these barriers together enables a computational loop in which microstructures are represented, predicted, searched, and validated in a more systematic and data-efficient manner. Specifically, generative AI is used to construct a continuous microstructure design space from image-based data, enabling candidate structures to be generated, interpolated, and systematically explored. Physics-aware neural surrogates are developed to approximate structure–processing–property–performance (SPP) linkages. Within this SPP linkage modeling component, specific efforts include: (1) improving the fidelity and reliability of neural surrogate models; (2) extending these models to account for chemical initial conditions, enabling a single model to explore different material configurations; and (3) accounting for epistemic uncertainty. Hybrid optimization then couples the learned design space with rapid response prediction to support target-guided exploration of candidate microstructures while balancing exploitation and exploration. Together, these contributions establish a computational loop connecting microstructure representation, SPP linkage approximation, and systematic optimization. The objective is not to replace high-fidelity simulation, experiment, or domain expertise, but to make their use more targeted through human–AI collaboration. By addressing representation, SPP evaluation, and search/decision barriers within one integrated framework, this dissertation provides a scalable foundation for AI-assisted microstructure characterization toward target performance.
