Deep learning and computational nanoscience

这项研究结合了人工智能，机器学习，材料科学和纳米技术。将在两个主要领域开展工作。研究的第一条线集中在使用深度神经网络来学习计算要求高的问题的近似解决方案，例如高度精确的原子尺度量子力学模拟。降低计算成本和这种计算的缩放限制将大大提高研究人员快速搜索，设计和评估新形式的纳米级增强材料的能力。所提出的方法是非常普遍的，并可适用于其他形式的数值模拟，如等离子体，计算流体动力学，预测。这项研究将开发该方法，并测试各种材料，如水-液界面，高压固体和液体，以及二维材料（包括纳米管）。第二条研究路线将考虑是否有可能通过示例的方式教授人工智能代理化学和物理直觉。我们将使用最近被证明足够强大的算法来征服围棋，而无需任何人类知识或干预。深度强化Q学习是AI的一种方法，通过一系列模拟体验，教AI学习在环境中移动或控制环境。我们将把这样的代理耦合到化学过程的物理现实模拟，使其能够控制温度，压力和存在的化学物质等参数。智能体将以与人类相同的方式观察系统-通过传感器和实验的组合。如果成功的话，一个训练有素的代理人将能够开发新的工艺来生产大规模的纳米增强材料，如纳米管，量子点，或二维材料。这将在许多领域产生巨大的影响，并将构成化学和材料发现的新方法。

This research combines artificial intelligence, machine learning, materials science, and nanotechnology. Work will be carried out in two main areas.******The first line of investigation is focused on using deep neural networks to learn approximate solutions to computationally demanding problems such as highly accurate atomic-scale quantum mechanical simulations. Reducing the computational costs and scaling limitations of such calculations will greatly increase researchers' ability to rapidly search for, design, and evaluate new forms of nanoscale enhanced materials. The proposed method is extremely general and may be adaptable to other forms of numerical simulation such as plasmonics, computational fluid dynamics, and forecasting. This research will develop the method and test it for a wide variety of materials such as water-liquid interfaces, high-pressure solids and liquids, and 2d materials (including nanotubes).******The second line of research will consider whether it is possible to teach an AI agent chemical and physical intuition by way of example. We will use algorithms which have recently been shown to be powerful enough to conquer the game of Go without any human knowledge or intervention. Deep reinforcement Q-learning is an approach to AI which, through a series of simulated experiences, teaches an AI to learn to move within, or exert control over, an environment. We will couple such an agent to physically realistic simulations of chemical processes, giving it control over parameters such as temperature, pressure, and chemical species present. The agent will observe the system the same way that a human does - through a combination of sensors and experiments. If successful, a trained agent will able to develop new processes for producing large-scale amounts of nano-enhanced materials such as nanotubes, quantum dots, or 2d materials. This would be hugely impactful in many fields and would constitute a new approach to chemical and material discovery.