A Neural Network-based Optimal Control Framework for Colloidal Self-Assembly

控制胶体自组装形成有序结构是一条很有前途的获得新物理化学性质的途径，在光子学、生物材料、药物、能量收集和先进通讯等领域有着广泛的应用前景。然而，高维数和复杂的动力学是快速生产有序结构的障碍。为了克服这些障碍，需要精确和可推广的状态表示和可靠且计算高效的方法来描述和预测系统动态。该项目提出了一种新的基于神经网络的最优控制框架，借鉴了胶体系统，机器学习和最优控制理论的多学科专业知识。这项工作的成功将：1）有助于快速生产用户自定义组装结构的潜在自动化最佳控制框架，2）有益于相关原子和分子自组装系统的研究，例如用于药物生产和核废料处理的结晶，以及蛋白质自组装; 3）促进跨学科研究，将先进的机器学习技术和控制理论相结合，以解决传统领域方法难以解决的复杂问题。在整个项目中，PI将在分子自组装、先进控制理论和机器学习主题方面教育和指导从中学到硕士的学生。针对电场介导的胶体自组装过程，研究人员提出了一种神经网络-最优控制集成方法，以解决与控制随机和高维小颗粒自组装过程相关的长期挑战。卷积神经网络将用于表示和分类系统状态，避免了广泛的试错探索和验证，以及与顺序参数相关的有限的可转移性。该项目还将部署一个随机神经网络，用装配过程的时间序列进行训练，以捕获和预测系统动态。将卷积和随机神经网络与强化学习相结合，然后计算最佳控制策略，以指导外部电场电压水平的操作，从而快速将组件驱动到所需的结构。所提出的工作提出了一个创新的解决方案，状态表示和分类，和一个有效的系统动力学模拟的随机，高维胶体系统。由于机器学习模型的数据驱动性质，拟议的框架可以转移到由于不同驱动力和/或不同颗粒大小和形状而具有不同内部交互的系统。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Controlling colloidal self-assembly for ordered structures is a promising route to novel physicochemical properties, which stands to benefit applications in photonics, biomaterials, pharmaceutics, energy harvest, and advanced communication. However, high dimensionality and complex dynamics are barriers to rapid production of ordered structures. To overcome these barriers, accurate and generalizable state representation and a reliable and computationally efficient approach to describe and predict the system dynamics are needed. This project proposes a novel neural network-based optimal control framework, drawing on multidisciplinary expertise from colloidal systems, machine learning, and optimal control theory. The success of the work will: 1) contribute to a potentially automatable optimal control framework for rapid production of user-defined assembly structures, 2) benefit studies on related atomic and molecular self-assembly systems, such as crystallization for drug production and nuclear waste handling, as well as protein self-assembly; and 3) promote interdisciplinary research on combining advanced machine learning techniques and control theory to tackle complex problems that are challenging to address with traditional domain approaches. Throughout this project, the PI will educate and mentor students from middle school to the master’s level on molecular self-assembly, advanced control theory and machine learning topics. Focusing on an electric field-mediated colloidal self-assembly process, the researcher proposes a neural network-optimal control integrated approach, to tackle the long-standing challenges associated with controlling a stochastic and high dimensional small particle self-assembly process. A convolutional neural network will be used to represent and classify the system state, avoiding the extensive trial-and-error exploration and validation, and the limited transferability associated with order parameters. The project will also deploy a stochastic neural network, to be trained with time series of the assembly process, to capture and predict the system dynamics. Combining the convolutional and stochastic neural network with reinforcement learning, an optimal control policy will then be computed to guide the manipulation of the voltage level of the external electric field, to rapidly drive the assembly to the desired structure. The proposed work presents an innovative solution to state representation and classification, and an efficient system dynamics simulation of the stochastic, high dimensional colloidal system. Due to the data-driven nature of machine learning models, the proposed framework is transferable to systems with different internal interactions due to different driving forces and/or different particle sizes and shapes.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.