Elements: Collaborative Research: Community-driven Environment of AI-powered Noise Reduction Services for Materials Discovery from Electron Microscopy Data

该项目的目标是创建由人工智能（AI）驱动的网络基础设施（CI），以实现材料科学的持续创新。对材料的深刻理解对于能源、通信、建筑、交通和人类健康等相关技术的进步至关重要。深度学习的革命性进步得益于开源人工智能模型和开放获取基准数据库的可用性。然而，现有的与图像处理相关的代码库和数据集主要集中在摄影图像上。为了促进能够对材料科学产生重大影响的人工智能技术的持续发展，提供适合该领域的数据和人工智能模型至关重要。开发的CI将通过提供软件来处理从电子显微镜获得的图像来满足这一需求，电子显微镜是一种使原子可视化的技术，具有在材料科学的各种重要领域实现变革性突破的潜力。CI的明确设计是为了在材料和数据科学社区的交叉点促进人工智能技术用户和开发者的可持续社区的发展，并使材料科学家能够模拟他们自己的数据集，并为科学发现开发他们自己的人工智能模型。因此，开发的人工智能驱动的CI将使对材料的原子级理解取得变革性进展，这将对健康、能源、环境和生物技术产生更广泛的影响。CI环境将有助于材料科学家在人工智能技术方面的培训，将他们与人工智能社区联系起来，并提供软件、数据和支持材料，以启动人工智能驱动的研究。教育和推广计划旨在促进材料科学和人工智能社区之间的互动。专门针对普通公众、高中教师及其学生的外展活动将使他们了解材料科学、电子显微镜和人工智能。该项目致力于为女性和代表性不足的群体提供机会，并将与纽约大学数据科学中心多样性委员会合作，优先考虑多样性。发展对原子水平结构和动力学的基本理解对于材料科学的变革性进步至关重要。像差校正透射电子显微镜是实现这一目标的主要工具。不幸的是，显微镜数据的信息内容可能严重限制了较差的信噪比。对于辐射敏感材料和需要高时间分辨率来研究动态动力学过程的实验尤其如此。人工智能方法可以通过训练深度神经网络并进行大量模拟来利用材料结构的先验信息。这些方法可能显著优于现有的最先进的方法，特别是对于非周期性结构，包括缺陷、界面和表面。开发的CI将提供人工智能降噪服务，这将对沸石、金属有机框架、蛋白质材料界面、液相成核和生长、液固界面以及催化纳米颗粒的流动行为产生直接的进步和影响。此外，该项目将推进面向人工智能的CI设计方法。CI的战略目的是为特定科学领域的人工智能技术的使用和开发创造一个整体环境。它将通过提供对最终用户透明的人工智能技术的软件，吸引缺乏人工智能专业知识的领域科学家。接触这项技术将激励科学界设计和训练他们自己的模型，这将由人工智能存储库中的开源代码库提供便利。开放访问数据库与存储库相结合，将吸引几乎没有领域专业知识的人工智能从业者，让他们能够访问精心整理的数据和相关人工智能任务的明确规范。这些服务将通过多种教育和外联活动迅速启动和支持。该奖项反映了美国国家科学基金会的法定使命，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。

The goal of this project is to create cyberinfrastructure (CI) powered by artificial intelligence (AI) for sustained innovation in materials science. Deep understanding of materials is critical for progress in technologies related to energy, communication, construction, transportation and human health. The revolutionary progress of deep learning has been enabled by the availability of open-source AI models and open-access benchmark databases. However, the existing codebases and datasets relevant to image processing focus mostly on photographic images. In order to promote the sustained development of AI technology that can have significant impact in materials science, it is critical to provide data and AI models that are tailored to this domain. The developed CI will address this need by providing software to process images obtained from electron-microscopes, a technique enabling atoms to be visualized, and has the potential to enable transformative breakthroughs in varied and important areas of materials science. The CI is explicitly designed to foster the growth of a sustainable community of users and developers of AI technology at the intersection of the materials and data science communities, and to empower materials scientists to simulate their own datasets and develop their own AI models for scientific discovery. The developed AI-powered CI will therefore enable transformative progress in atomic-level understanding of materials, which will have broader impacts in health, energy, environment, and biotechnology. The CI environment will contribute to training materials scientists in AI technology, connecting them to the AI community, and providing software, data, and support materials to initiate them in AI-powered research. Educational and outreach plans are designed to facilitate interactions between the materials science and AI communities. Outreach activities specifically targeted to the general public, and to high-school teachers and their students, will expose them to materials science, electron microscopy, and AI. The project is committed to providing opportunities to women and underrepresented groups and will prioritize diversity in collaboration with the NYU Center for Data Science diversity committee.Developing a fundamental understanding of atomic level structure and dynamics is critical for transformative advances in materials science. Aberration-corrected transmission electron microscopy is a primary tool to accomplish this goal. Unfortunately, the information content of microscopy data may be severely limited by poor signal-to-noise ratios. This is particularly true for radiation sensitive materials and experiments where high time resolution is required to investigate dynamic kinetic processes. AI methodology can exploit prior information about material structure by training deep neural nets with extensive simulations. These approaches may significantly outperform existing state-of-the-art methods, especially for non-periodic structures, including defects, interfaces, and surfaces. The developed CI will provide AI noise reduction services which will yield immediate advances and impacts for zeolites, metal organic frameworks, protein-material interfaces, liquid phase nucleation and growth, liquid-solid interfaces, and fluxional behavior in catalytic nanoparticles. In addition, the project will advance methodology for the design of AI-oriented CI. The CI is strategically designed to create a holistic environment for the use and development of AI technology in a specific scientific domain. It will attract domain scientists with little AI expertise, by providing software where the AI technology is transparent to the end user. Exposure to the technology will motivate the scientific community to design and train their own models, which will be facilitated by the open-source codebase in the AI repository. The open-access database combined with the repository will attract AI practitioners with little domain expertise, by giving them access to well-curated data and a clear specification of the relevant AI tasks. These services will be jump-started and supported through multiple educational and outreach activities.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.