Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 150035

150035 - Leveraging Machine Learning for Enhanced Control in Colloidal Self-Assembly Systems. 

Colloidal assemblies with specific structures have the potential of novel functions and applications in fields such as energy harvesting and drug delivery. Using external fields to manipulate the colloidal self-assembly process presents a promising approach to rapidly obtain such structures in a thermodynamically non-equilibrium manner. However, the scalability and automation of external field-mediated control of colloidal self-assembly are still hindered by several challenges, from the assembly state identification to the design of the optimal operating conditions, i.e. optimal control policies. For example, in a conventional colloidal self-assembly control framework, the assembly state is typically described with order parameters, which are aggregate variables designed for the specific problem of interest with either heuristic experience to bare physical meanings or mathematical dimensionality reduction approaches. However, identifying the appropriate order parameters can be a tedious task.

To address these challenges, we propose a data-based control framework for colloidal self-assembly. Our approach leverages image-based representation and classification of the assembly states, to provide a framework that is potentially generalizable to systems where the configuration state can be captured by images. We also employ Deep Reinforcement Learning (DRL) methods for the optimal control policy design, to handle the curse of dimensionality caused by the continuous action space, avoiding potential performance sacrifice from discretizing the action space. Specifically, we implement the Deep Deterministic Policy Gradients (DDPG) algorithm, given its proven success in handling continuous action space for optimal control policy design. Furthermore, with the control policy from the DDPG algorithm, we further infer the set of the most important control actions necessary for the performance, thus shedding lights on the discretization of the control action space, in case the computational demand is an issue that a discrete action space and control design is needed. As the policies are based on artificial neural networks, we anticipate a trained policy to also be adaptable with transfer learning for different objectives, i.e. a different assembly structure, thus reducing the cost involved in designing new control policies as in conventional approaches.

Our Brownian Dynamics simulation results demonstrate that the proposed machine learning-based optimal control approach is able to effectively steer a 2-D electric field-mediated colloidal self-assembly to arbitrarily defined target configurations, including void defective states, polycrystalline states, and perfectly ordered states. By leveraging machine learning, we offer an alternative method to describe and classify the system state, paving the way to automation and generalizability to other systems. The use of reinforcement learning eliminates the need for prior knowledge of the system dynamics in designing the control policy, providing advantages in managing systems with unknown and complex behaviors. The work here represents an effective paradigm for automated feedback control of colloidal self-assembly. We anticipate our approach to be applicable to other external field-mediated colloidal self-assembly systems, such as magnetic and acoustic fields, nanoparticle self-assembly, and self-assembly systems with different particle sizes and shapes.

Presenting Author: Andres Lizano-Villalobos Louisiana State University

Presenting Author Biography: Born and raised in San Jose, Costa Rica in 1994, the author embarked on an academic journey that spanned both Chemical Engineering and Physics. Their undergraduate years at the University of Costa Rica (UCR) were marked by early research endeavors. Specifically, they delved into molecular simulations, exploring excited states and their applications in pharmacology and solid-state systems. This scientific exploration unfolded within the Quantum Chemistry Lab at UCR.

Following graduation, the author wore multiple hats: a lecturer at the University of Costa Rica and a part-time DevOps engineer. Their passion for both teaching and technology converged during this period.

In spring 2021, the author’s path led them to Baton Rouge, Louisiana, where they commenced their PhD program at the Cain Department of Chemical Engineering. Here, they found a home in the Tang Lab—a place where their research now thrives. The focus? The application of data-based modeling and artificial intelligence (AI) to tackle the intricate control of high-dimensional stochastic systems. These systems, often characterized by limited data availability, pose fascinating challenges that the author eagerly embraces.

Authors:

Andres Lizano-Villalobos Louisiana State University
Xun Tang Louisiana State University