Degree

Doctor of Philosophy (PhD)

Department

Cain Department of Chemical Engineering

Document Type

Dissertation

Abstract

Controlling colloidal self-assembly is essential for manufacturing materials with well-defined structural properties. However, developing effective control protocols is challenging, requiring domain expertise and significant computational resources---from identifying suitable state representations to designing control policies. In this dissertation, we introduce a fully data-driven framework for automating the optimal control of two-dimensional colloidal self-assembly, reducing reliance on domain knowledge. The framework integrates machine learning-based state representation with deep reinforcement learning (DQN), and is demonstrated on an experimentally validated Brownian Dynamics (BD) model of a quadrupole electric field-mediated assembly system. To construct effective state descriptors, we first apply an unsupervised learning pipeline that combines convolutional autoencoders, UMAP, and HDBSCAN to extract and cluster latent features from particle configuration images. This clustering is further refined using Siamese neural networks (SiNNs), improving cluster-based classification accuracy from 72.26\% to 92.23\%. We then train supervised classifiers---convolutional neural networks (CNNs) and graph neural networks (GNNs)---to map new configurations to structural states, using either images or particle graphs as inputs. Validation against pre-labeled data shows that these machine learning-based representations capture structural subtleties more effectively than traditional order parameters such as $\psi_6$. Using these learned state representations, we train Deep Q-Network (DQN) controllers to drive the system toward target states. When tested in BD simulations, CNN-based DQN policies achieved a 99\% success rate relative to the CNN-defined target cluster, but only a 48\% actual success rate upon visual inspection, due to classification errors propagating through the control process. In contrast, GNN-based DQN policies achieved a 97\% actual success rate, demonstrating the value of incorporating topological information into state representations. Comparative analysis also confirms that $\psi_6$ is not an ideal descriptor for this system, whereas GNN-derived representations improve both state classification and control performance. Overall, this framework provides a modular and generalizable approach for closed-loop control of self-assembling materials. By eliminating the need for hand-crafted order parameters and analytical process models, it offers a pathway toward fully automated control pipelines applicable to broader micro- and nano-particle assembly systems.

Date

5-20-2025

Committee Chair

Xun Tang

Download

Available for download on Friday, May 19, 2028

Share

COinS