A Machine Learning Framework for Acceleration of Materials Prediction

用于加速材料预测的机器学习框架

• 批准号: 1410514
• 负责人: Alexey Kolmogorov
• 金额: $ 37.2万
• 依托单位: SUNY at Binghamton
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2014
• 资助国家: 美国
• 起止时间: 2014-09-01 至 2018-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1410514&HistoricalAwards=false
• 关键词: Machine; Learning; Framework; Acceleration; Materials

NON-TECHNICAL SUMMARYComputational research has been playing an increasingly important role in the development of new materials. The central aim of this project is to create a theoretical framework and computational tools capable of speeding up the prediction of new synthesizable materials by orders of magnitude. The efficiency and the reliability of the method will be achieved by combining two bio-inspired algorithms. A learning "neural network" scheme will be implemented in the "Module for Ab Initio Structure Evolution" package which presently has the capability to perform global structure optimization with an evolutionary algorithm. The introduced computational approach will be applicable to a broad range of material classes and expected to accelerate the exploration of complex systems. A particular focus will be placed on the development of metal oxide catalysts and metal-based battery electrodes for energy-related applications.Artificial neural networks are a powerful tool used in many research areas for dealing with classification, control, and interpolation problems in multi-dimensional spaces. The application of the method to solid state problems requires a formalism specifically designed for building and training neural networks to map the potential energy profiles of given atomic configurations. Namely, an automated algorithm should be able to parse an arbitrary atomic configuration into a suitable set of input parameters, train the neural network on a relevant dataset, and monitor the ability to produce accurate total energy and atomic forces during simulations. The PI's preliminary work and the recent research in the field have shown that, compared to widely used quantum mechanical and classical models, neural networks have the potential to provide a good balance between accuracy and efficiency.The project will include multiple educational activities to foster high-school and undergraduate students' interest in science and mathematics. Students will have the opportunity to learn about computational methods and take part in the research under PI's supervision. A part of the outreach work will be done through Binghamton University's Evolutionary Studies program which brings together students and researchers from biological, physics, and computer sciences.TECHNICAL SUMMARYThis award supports computational and theoretical research and education directed at advancing analytical and computational techniques for materials discovery. Computational input can be particularly valuable at the initial stages of materials development providing libraries of synthesizable candidates and possible synthesis conditions. The success of determining new stable materials depends on the careful global optimization of crystal structures and the accurate evaluation of their thermodynamic stability. Traditionally, the two challenges have been addressed separately. This project will combine a neural network method with an evolutionary algorithm to automate and accelerate compound prediction. The neural network formalism will be generalized to describe a wide range of interatomic interactions with near ab initio accuracy while scaling linearly with the system size. Preliminary tests have shown the ability of such models to capture subtle many-body effects. The developed computational approach is expected to accelerate the exploration of systems known to exhibit particularly complex structures, such as metal oxide catalysts or metal-based battery electrodes. Neural network models will be trained, in particular, on already generated databases containing thousands of entries. Analysis of the constructed neutral-network-based interpolators will advance the fundamental understanding of the bonding mechanisms for future rational materials design. The compound prediction tool offering a combination of two bio-inspired algorithms will be released as an open-source code. One of the objectives is to create a shared online resource with full access to all built interatomic models and all ab initio data used for training the neural networks. Such a platform will ensure reuse of generated density functional theory data and reduce the redundancy of expensive ab initio calculations. Validation of the developed predictive approach will be carried out in close collaborations with Chemistry and Engineering experimental groups at Binghamton University. The project will also include multiple educational activities to foster high-school and undergraduate students' interest in science and mathematics. Students will have the opportunity to learn about computational methods and take part in the research under PI's supervision. A part of the outreach work will be done through Binghamton University's Evolutionary Studies program which brings together students and researchers from biological, physics, and computer sciences.

非技术总结计算研究在新材料的开发中发挥着越来越重要的作用。该项目的中心目标是创建一个理论框架和计算工具，能够加速预测新的可合成材料的数量级。通过结合两种生物启发算法，实现了该方法的高效性和可靠性。将在“Ab Initio结构进化模块”软件包中实施学习“神经网络”方案，该软件包目前具有使用进化算法进行全局结构优化的能力。引入的计算方法将适用于广泛的材料类别，并有望加速复杂系统的探索。一个特别的重点将放在开发金属氧化物催化剂和金属基电池电极的能源相关的应用。人工神经网络是一个强大的工具，用于处理分类，控制和插值问题，在许多研究领域的多维空间。固态问题的方法的应用需要一个专门设计的形式主义，用于建立和训练神经网络映射给定的原子配置的势能分布。也就是说，自动化算法应该能够将任意原子配置解析为一组合适的输入参数，在相关数据集上训练神经网络，并在模拟过程中监控产生准确总能量和原子力的能力。PI的前期工作和最近在该领域的研究表明，与广泛使用的量子力学和经典模型相比，神经网络具有在准确性和效率之间取得良好平衡的潜力。该项目将包括多种教育活动，以培养高中和本科生对科学和数学的兴趣。学生将有机会学习计算方法，并在PI的监督下参与研究。一部分推广工作将通过宾厄姆顿大学的进化研究计划来完成，该计划汇集了来自生物学，物理学和计算机科学的学生和研究人员。技术总结该奖项支持旨在推进材料发现的分析和计算技术的计算和理论研究和教育。计算输入在材料开发的初始阶段可能特别有价值，提供可合成候选物和可能的合成条件的库。确定新的稳定材料的成功取决于晶体结构的仔细的全局优化和对其热力学稳定性的准确评估。传统上，这两个挑战是分开处理的。本计画将联合收割机结合神经网路方法与演化演算法来自动化及加速化合物预测。神经网络的形式主义将被推广到描述广泛的原子间的相互作用与近从头算的精度，同时线性缩放与系统的大小。初步测试表明，这种模型能够捕捉到微妙的多体效应。预计开发的计算方法将加速对已知具有特别复杂结构的系统的探索，例如金属氧化物催化剂或金属基电池电极。神经网络模型将在已经生成的包含数千个条目的数据库中进行训练。 对所构造的基于神经网络的插值器的分析将为未来合理的材料设计提供对键合机理的基本理解。提供两种生物启发算法组合的化合物预测工具将作为开源代码发布。目标之一是创建一个共享的在线资源，可以完全访问所有构建的原子间模型和用于训练神经网络的所有从头算数据。这样的平台将确保生成的密度泛函理论数据的重复使用，并减少昂贵的从头计算的冗余。开发的预测方法的验证将与宾厄姆顿大学的化学和工程实验组密切合作进行。 该项目还将包括多种教育活动，以培养高中和本科生对科学和数学的兴趣。学生将有机会学习计算方法，并在PI的监督下参与研究。一部分推广工作将通过宾厄姆顿大学的进化研究计划完成，该计划汇集了来自生物学、物理学和计算机科学的学生和研究人员。