Digital Repository at the University of Maryland
AUTONOMOUS EXPERIMENTATIONS FOR MATERIALS OPTIMIZATION
Abstract
dc:description.abstractA new era of scientific discovery is emerging: one in which humans and intelligent machines work together to accelerate the pace of research. Fueled by major advances in artificial intelligence (AI) since 2012, AI-enabled systems are increasingly replacing labor-intensive steps in traditional scientific workflows with automated platforms capable of operating under partial or full autonomy. AI models with perception, control, reasoning, and forecasting capabilities are particularly well suited for these applications, as they can continuously interpret experimental outcomes, make informed decisions, and adapt future actions without requiring constant human intervention.This dissertation presents three distinct demonstrations of autonomous experimentation: autonomous phase-diagram construction, autonomous thin-film growth optimization, and autonomous additive-manufacturing optimization. Together, these studies illustrate the synergistic integration of instrument automation with self-optimizing closed-loop systems, enabling researchers to achieve experimental objectives more efficiently, reproducibly, and cost-effectively. In all three studies, convolutional neural network–based models serve as low-latency perception layers that convert raw experimental data into physically meaningful features. These features are subsequently processed into scalar performance metrics that guide the self-optimizing closed-loop workflow. Across the three platforms, the optimization loop is implemented using Gaussian process–based Bayesian optimization, which iteratively updates its representation of the experimental landscape and selects subsequent experiments accordingly. Beyond the individual case studies, this dissertation demonstrates a general framework for autonomous materials research: automate experimental execution, extract meaningful information from raw measurements in real time, and use that information to guide subsequent decisions. Such systems do not eliminate the role of researchers; rather, they shift researchers away from repetitive operations and toward higher-level tasks, including defining scientific objectives, designing experimental spaces, interpreting physical mechanisms, and validating new discoveries. As experimental instruments become increasingly connected, programmable, and data-rich, the integration of AI-driven closed-loop experimentation will become an important foundation for faster and more systematic materials discovery.
Degree
thesis:*- Department dc:contributor.department
- Material Science and Engineering
- Grantor dc:publisher
- Digital Repository at the University of Maryland
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Liang, Haotong
- Advisor dc:contributor.advisor
-
- Takeuchi, Ichiro IT
Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier
- https://doi.org/10.13016/dyug-nkft