(510b) Stochastic Neural Network-Based Surrogate Model for 2-D Colloidal Self-Assembly.

Colloidal self-assembly is a viable bottom-up approach to the production of materials where the specific properties depend on the structure of the assembly. Effectively controlling colloidal self-assembly into specific structure is a standing challenge, due to the inherent complex, high dimensional, nonlinear, and stochastic dynamics. While detailed force-balance models, such as Brownian dynamics model, exist to help illustrate the system dynamics as a cost-effective alternative to experiments, as the colloidal self-assembly process normally spans a large time scale, solving those models might be time inefficient. Here, in this work we propose a stochastic deep neural network to learn and describe the stochastic dynamics of a 2-D electric field-mediated colloidal self-assembly process, with significantly improved computation efficiency, as compared to detailed force-balance models. Specifically, we deploy a convolutional neural network to represent the system state of the process, and learn the temporal information of the process from an experimentally validated Brownian dynamics simulation. The performance evaluation reveals that the proposed stochastic neural network model is capable of accurately predicting the evolution of a 2D electric field mediated system with significantly reduced computational time. The data-driven nature of the proposed approach also enables the automatability and generalizability to other related systems.