Combining Deep Learning and Coarse Grained Simulation Methods to Study High-Dimensional NanoBiophysical Systems

我们正处于一场人工智能(AI)革命之中。计算技术的进步在让计算机“思考”方面迈出了巨大的步伐，达到顶峰。机器学习(ML)推动了这些进步的大部分，其中使用计算算法和方法(尤其是神经网络)来识别大数据集中的模式，然后根据这种训练，在给定较少量输入数据的情况下预测结果。在过去的几年里，ML从根本上改变了既定的做法，从评估医疗数据到欺诈检测，再到财务预测。它对科学研究的影响同样令人震惊。这项资助申请的重点是使用ML来解决数学方程。虽然我们知道物理事物和过程可以用数学方程来建模，但寻找方程往往比实际求解它们更具挑战性。ML提供了一种高效而强大的方法来求解方程并产生有意义的答案。 我们打算将最大似然技术用于求解偏微分方程组。这是一类可以描述复杂过程和环境的方程。PDE可以描述分子在纳米流体设备中的流动，这些设备的尺寸大约为纳米级(一纳米大约比人类头发的宽度小一万倍)。它们能够分离、表征和修饰单个生物分子，如DNA和蛋白质。这些设备的主要用途是在先进的医疗实践中，如个性化医疗，其中风险、诊断和治疗由一个人的基因构成提供信息。 我们的目标是开发机器学习方法和技术，以显著提高纳米流体设备的研究和设计。ML方法可以解决依赖于大量因素(即高维)的系统，允许在所有可能的情况下研究复杂系统；这是其他方法难以做到的。这使我们能够有效地提炼当前的设备并设计新的设备来执行任务，例如识别指示特定疾病的独特DNA链。虽然我们的重点是为纳米流体设备开发这些技术，但这些知识将转化为从核反应堆设计到理解奶油如何与咖啡混合的大量其他场景。 通过将ML开发为解决偏微分方程的强大工具，我们的研究也可以帮助其他学术实验室和工业研发中基于计算的研究。这项工作将有助于将加拿大的学术研究定位在ML革命的前沿。它还将为加拿大企业提供一个强大的工具，以快速且具有成本效益地开发他们的技术。通过培养本科生和研究生，这一研究项目将培养出所需的高素质人才，以确保加拿大在快速发展和日益科技的全球经济中保持领先地位。

We are in the midst of an artificial intelligence (AI) revolution. Computational advances have culminated in massive strides being taken towards having computers “think”. Machine learning (ML) has driven most of these advancements where computational algorithms and methodologies most notably neural networks are employed to identify patterns in large data sets and then, based on this training, predict outcomes when given smaller amounts of input data. In the past couple of years ML has fundamentally changed established practices from evaluating medical data to fraud detection to financial prediction. Its impact on scientific research is equally astounding. This grant application focuses on the use of ML to solve mathematical equations. While we understand that physical things and processes can be modelled with mathematical equations, finding the equations is often less challenging than actually solving them. ML provides a highly effective and powerful way to solve equations and produce meaningful answers. We intend to advance ML techniques to solve partial differential equations (PDEs). These are a class of equations that can describe complex processes and environments. PDEs can describe the flow of molecules within nanofluidic devices which have dimensions on the order of nanometers (a nm is around ten thousand times smaller than the width of human hair). They are able to isolate, characterize, and modify single biological molecules such as DNA and proteins. A primary use of these devices is in advanced medical practices such as personalized medicine where risk, diagnosis, and therapeutics are informed by one's genetic makeup. Our goal is to develop machine learning methods and techniques to significantly enhance nanofluidic device research and design. ML methods can solve systems that depend on a large number of factors (i.e., high dimensional), allowing the study of complex systems across all possible situations; this is difficult to do with other approaches. This enables us to efficiently refine current and design new devices for tasks such as identifying unique DNA strands that indicate a particular disease. While our focus is on developing these techniques for nanofluidic devices, this knowledge will translate to a vast array of other scenarios from nuclear reactor design to understanding how cream mixes with coffee. By developing ML as a powerful tool for solving PDEs, our research can also benefit computation-based research in other academic labs as well as industrial R&D. This work will help position Canadian academic research at the forefront of the ML revolution. It also will equip Canadian businesses with a powerful tool for rapid and cost efficient development of their technologies. Through training undergraduate and graduate students, this research program will produce the highly qualified personnel needed to ensure Canada remains at the leading edge of a rapidly evolving and increasingly technological global economy.