DMREF - Material Intelligence for Accelerated Design of Biologically-Interfaced Single-Layered Devices

快速扩展的机器学习（ML）和人工智能（AI）算法为材料科学与工程（MSE）提供了工具，这在几年前是前所未有的。即使是当今最先进的传统工具（微流体、高级建模和超级计算机）也无法跟上数据密集型ML/AI工具提供的机会。实现材料基因组计划（MGI）战略计划中规定的目标的主要障碍是难以获得可用于工程系统预测设计和建模的干净，精心策划，全面，有意义，最重要的是标准化数据，特别是在生物学和材料科学之间高度复杂和动态的接口中。在领域科学中部署这些强大算法的能力仍然有限，这是由于参数空间的维度数量和数据的巨大变化。为了克服当前的障碍，该项目旨在开发一个模块化软件框架，称为材料智能（Mat-I），以加速MSE的发现和创新。通过利用Mat-I技术，科学界很有可能加速研究，（微软、亚马逊、谷歌、英伟达、真实的网络、普罗克特和甘布尔、艾伦人工智能研究所和英特尔）将拥有关键的工具，在生物学、固态系统、以及设计设备的信息学，如用于癌症诊断的生物纳米传感器，用于能量收集的生物分子燃料电池，以及面向类脑计算机的神经形态网络。该项目将培养下一代创新科学家、本科生、博士和博士后研究人员，巩固美国在世界舞台上的传统竞争优势。汇聚科学团队的技术目标是构建一个模块化的Mat-I软件框架，以加速发现和创新。该研究将在广泛的实验条件下生成并提供结构输入空间（肽和单原子层固体，分别是生物学和物理科学中最小的可行实体）到物理特性输出目标空间之间的综合地图。我们的目标是学习三个参数之间的相关性，以便在给定序列/结构表示和实验条件下，然后可以预测输出物理性质，这可以适用于复杂的工程解决方案。所提出的方法将采用，增强，并开发具体的数学，统计和信息的方法，发现在材料工程，将联合收割机物理，信息和生物科学相结合。给定一组测量值，该团队将应用ML/AI进行推断并学习真实底层过程的模型，并使用这些推断和不确定性的量化，该团队将设计测试床，以最大限度地利用模型获得的信息。通过收集数据并在迭代循环中进行关联，发现的步伐将加快，比标准方法更快地缩小知识差距。该研究将使用模型选择，鲁棒统计，自适应学习，以及生物纳米界面静态和动态表示的原型验证。该项目的最终目标是通过将生物学与未来的固态器件相结合，通过神经网络的形成，为技术和医学领域的各种关键湿件器件建立基本规则。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Rapidly expanding machine learning (ML) and artificial intelligence (AI) algorithms offer tools for Materials Science and Engineering (MSE), unprecedented just a few of years ago. Even the most advanced traditional tools today (microfluidics, advanced modeling, and supercomputers) cannot keep up with the opportunities offered by data intensive ML/AI tools. The main barrier to achieve the goals laid out in the Materials Genome Initiative (MGI) strategic plan is the difficulty to access clean, curated, comprehensive, meaningful, and most of all, standardized data that can be used in predictive design and modeling of engineered systems, especially true in highly complex and dynamic interfaces between biology and materials science. The ability to deploy these powerful algorithms in domain sciences has remained limited due to the sheer number of dimensions of the parameter space and enormous variability in the data. With the goal of overcoming the current barriers, this project aims to develop a modular software framework, dubbed Materials Intelligence (Mat-I) towards accelerating discovery and innovation in MSE. By exploiting Mat-I technology, the scientific community has the high likelihood of accelerating research, and the collaborating industry (Microsoft, Amazon, Google, NVIDIA, Real Networks, Proctor and Gamble, Allen Institute for Artificial Intelligence, and Intel) will have the crucial tools to develop materials and methods with tailored bio-nano interfaces at the critical intersection of biology, solid-state systems, and informatics in designing devices such as bionanosensors for cancer diagnostics, biomolecular fuel cells for energy harvesting, and neuromorphic networks towards brain-like computers. The project will educate the next generation of innovative scientists, undergraduates, PhDs, and post-doctoral researchers, bolstering the traditional competitive edge of the US at the world stage.The technical aim of the convergence science team, with expertise in genomics, computer science, physics, and materials science and engineering, is to construct a modular Mat-I software framework towards accelerating discovery and innovation. The research will generate and make accessible comprehensive maps among the input space of structures (peptides and single atomic layer solids, the smallest viable entities in biology and physical sciences, respectively) to the output target space of physical properties under a wide range of experimental conditions. The goal is to learn correlations among the three parameters such that, given the sequence/structure representations and experimental conditions, one can then predict the output physical properties, which may be adapted to complex engineered solutions. The proposed approach will employ, enhance, and develop specific mathematical, statistical, and information approaches for discovery in materials engineering that will combine physical, information, and biosciences. Given a set of measurements, the team will apply ML/AI to make inferences and learn a model of the true underlying process and, using these inferences and quantifications of uncertainty, the team will devise test-beds to maximize the information gained with respect to the model. By collecting data and making correlations in an iterative loop, the pace of discovery will be accelerated in closing the knowledge gaps faster than standard methods. The research will use model selection, robust statistics, and adaptive learning, and prototype validation in both static and dynamic representations of bio-nano interfaces. The project will establish foundational rules of a wide range of key wetware devices for technology and medicine through neural network formation by incorporating biology with solid-state devices of the future, the ultimate goal of the project.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.