Date of Graduation

5-2026

Document Type

Thesis

Degree Name

Bachelor of Science in Computer Science

Degree Level

Undergraduate

Department

Computer Science and Computer Engineering

Advisor/Mentor

Luu, Khoa

Committee Member

Churchill, Hugh

Second Committee Member

Gauch, Susan

Abstract

The advancement of next-generation semiconductor and quantum technologies relies on the scalability of the fabrication of two-dimensional (2D) van der Waals heterostructures. However, this process is severely bottlenecked by characterization workflows. Optical microscopy provides high-throughput imaging of 2D material flakes, but lacks the explicit physical priors required for the discernment of sub-nanometer thickness variations, such as distinguishing monolayers from bilayers. The use of computer vision models to automate the localization and characterization process of the flakes was proposed. As a part of this effort, we develop QuantumFlake, an open-source framework to streamline the integration and deployment of computer vision models (e.g. YOLO, DETR, VitDet) for 2D material flake hunting and characterization with intuitive training and inference pipelines. However, an issue of these static vision models is that they have vulnerabilities in dynamic, real-world laboratory environments. As new 2D materials are introduced into a workflow, the traditional models will likely need to be retrained for the new material. However, in doing so, these networks will overwrite the past learned feature representations from the previous material. To stabilize these systems, we introduced CLIFF, a continual learning architecture. In doing so, we introduced a new problem formulation for flake classification in the form of a material-incremental problem. In our experiments, we show that CLIFF preserves material-specific features through the use of prompt pools, delta heads, and memory rehearsal to ensure robust knowledge retention across evolving, multi-material workflows. Yet, even with stable continual learning, typical vision-based architectures remained fundamentally constrained by their inability to inherently model thin-film optical physics. These models consistently fail to distinguish physical truth from illumination noise due to fractional contrast differences between sub-nanometer phase shifts (e.g. monolayer versus bilayer) heavily entangling with background interference. To bridge this divide, we proposed QuPAINT, a physics-aware instruction-tuning approach. By embedding explicit optical interference priors to a multimodal foundation model using a synthetic data engine and a novel Physics-Informed Attention (PIA) module, QuPAINT successfully anchors visual embeddings to physical laws. Finally, as domain-expert Multimodal Large Language Models (MLLMs) like QuPAINT achieve high physical reasoning accuracy, their architectural optimization for cognitive transparency and their fine-tuned approach for answering in a specialized template-style yields verbose reasoning traces for all queries, simple or complex, rather than actionable metrics required for dynamic real-world usage. To evolve from passive visual inference to autonomous laboratory execution, we propose OpenQlaw, an agentic orchestration system. OpenQlaw resolves the interpretability gap by decoupling dense physical reasoning from practical execution. It uses QuPAINT as a localized "Material Domain Expert", while a central orchestrator translates the complex inferences into utility. By autonomously invoking programmatic tools, such as computing deterministic physical surface area, rendering targeted visual annotations, and maintaining persistent metadata context across conversational interactions, OpenQlaw moves from isolated inferencing towards an autonomous, physics-grounded agentic workflow to enable high-throughput quantum material discovery.

Keywords

Agentic AI, Multimodal Large Language Models, Continual Learning, Two-Dimensional Materials, Physics-Informed Machine Learning

Download

Share

COinS