EAGER: Type II: Deep Learning and Combinatorial Algorithms for Inorganic Crystal Structure Prediction

非技术总结这个热切的奖项支持在MATDAT18数据马拉松活动中引发的新合作的研究和教育，重点是开发人工智能方法，以发现新材料或识别具有所需应用特性的特定材料。与材料科学家和工程师使用的传统方法相比，涉及计算、材料数据和数据科学工具的方法提供了以更快的速度和更低的成本找到或设计具有所需性能的材料的可能性。在这个项目中，研究小组将开发新的机器学习技术，从各种公开可用的材料及其性能数据库中挖掘知识。所获得的知识可用于选材和设计。该团队将首先专注于使用数据科学的方法和从一个大型社区储存库获得的材料数据，这些数据是通过使用计算机和理论来计算从材料的组成元素形成材料所需的能量。在这个项目中开发的所有技术都将被编码为适用于不同计算机的软件。软件将通过包括GitHub在内的一系列机制以开放源代码的形式向材料科学和数据科学界发布。该项目还将为研究生和本科生提供教育机会，并在材料科学的数据分析方面获得第一手研究经验。该项目的成果将纳入适当的本科生和研究生课程。将作出强有力的努力，将少数民族和妇女包括在内。该项目的结果将通过期刊和国际会议上的出版物广泛传播。技术总结这一热切的奖项支持研究和教育，涉及在MATDAT18数据松活动中引发的新合作，重点是开发形成能量和其他材料性质的深度学习预测器。在材料基因组倡议下开发的材料项目和AFLOWLIB等计算材料属性的大型数据库，存储了数万种材料的属性。它们主要用于筛选各种目标应用的材料，如光催化和电池材料。这类数据库还可以用来开发基于深度学习的材料性能预测值。这些预测预计将比使用传统机器学习(ML)技术进行的预测更准确。与深度学习中用于挖掘海量数据的多层深度人工神经网络相比，即使是最新的传统最大似然方法，如梯度增强或随机树林，也具有有限的能力或学习能力。在这个项目中，研究小组的目标是开发一种用于晶体形成能的深度学习预测器。研究人员还建议开发其他相关的组合算法来解决这个问题。形成能，即晶体和原子形式的组成元素之间的能量差，是这些数据库提供的最可靠的属性之一。该项目的重点是利用深度学习系统和其他算法从大数据中学习的优越能力，对材料的形成能和稳定性进行快速和高精度的预测。该项目将提供一个可公开访问的网络基础设施，实现一个能够预测无机材料形成能的深度学习系统，其精度远远优于用传统ML模型建立的预测器，以及可以重复使用的材料化学表示的新形式，以预测材料的其他性质。深度学习技术应用中的挑战之一是这些算法所需的大量训练时间。研究团队计划通过各种算法创新来应对这一挑战，包括新颖的并行训练算法。调查人员计划采用一些并行架构，包括CPU集群和GU。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

NONTECHNICAL SUMMARYThis EAGER award supports research and education involving a new collaboration kindled at the MATDAT18 Datathon event focused on developing artificial intelligence methods to discover new materials or identify specific materials with desired properties for an application. Methods involving computation, materials data, and the tools of data science offer the potential to find or design a material with desired properties much faster and at lower cost than traditional methods used by materials scientists and engineers. In this project, the research team will develop novel machine learning techniques to mine knowledge from various publicly available databases on materials and their properties. The knowledge thus gained can be utilized for material selection and design. The team will focus first on using the methods of data science and materials data from a large community repository obtained from using computers and theory to calculate the energy needed to form a material from its constitutive elements. All the techniques developed in this project will be coded as software for different computers. Software will be released as open source codes to the materials science and data science communities via a number of mechanisms including the GitHub. This project will also provide educational opportunities to graduate and undergraduate students and a first-hand research experience in data analysis for materials science. Results of this project will be incorporated in appropriate undergraduate and graduate courses. Strong efforts will be made to include minorities and women. Results of this project will be disseminated widely via publications in journals and international conferences.TECHNICAL SUMMARYThis EAGER award supports research and education involving a new collaboration kindled at the MATDAT18 Datathon event focused on developing deep learning predictors for formation energies and other materials properties. Large databases of computed material properties, such as The Materials Project and AFLOWLIB developed under Materials Genome Initiative, host properties of tens of thousands of materials. They are primarily employed to screen materials for various target applications such as photocatalysis and battery materials. Such databases can also be utilized to develop deep learning-based predictors of materials properties. These predictions are expected to be more accurate than predictions made using traditional machine learning (ML) techniques. Even cutting edge conventional ML methods such as Gradient Boosting or Random Forest of Trees have limited capacity, or the ability to learn, when compared to multi-layer deep artificial neural networks employed in deep learning to mine vast data. In this project the research team aims to develop a deep learning predictor for formation energy of crystals. The investigators also propose to develop other relevant combinatorial algorithms for solving this problem. Formation energy, which is the energy difference between the crystal and the constituent elements in their atomic form, is one of the most reliable properties available from these databases. The focus of this project is on fast and highly accurate prediction of formation energies and stability of materials by utilizing the superior capacity of deep learning systems and other algorithms to learn from big data.The project will deliver a publicly accessible cyber infrastructure implementing a deep learning system capable of predicting formation energies for inorganic materials with an accuracy that is vastly superior to that of the predictors built with traditional ML models, and new forms of chemical representations of materials that can be reused to predict other properties of materials. One of the challenges in the employment of deep learning techniques is in the large training times taken by these algorithms. The research team plans to address this challenge with a variety of algorithmic innovations including novel parallel training algorithms. The investigators plan to employ a number of parallel architectures including CPU clusters and GPUs.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.