Elements: DeepPDB: An open-source automated framework to enable high-fidelity atomistic simulations in unexplored material space

解决世界能源挑战的一个关键要求是开发新的节能和智能材料。为了帮助这一过程，可以使用最先进的和革命性的计算工具来模拟新材料，从而全面了解它们的特性和行为。这种材料模拟工作可以将新材料的上市时间从数十年大幅缩短到数月。传统的模拟方法可以在尽可能小的尺度（原子）和宏观水平（即毫米或更大）准确地预测材料的行为特性。许多关键的材料特性是在从几纳米到微米的尺度上定义和需要的，但在这些层面上缺乏模拟。虽然确实存在算法来模拟这些尺度的属性，但我们通常缺乏新材料的基本参数，如下一代太阳能电池，电池和喷气涡轮机合金。使用传统方法来确定这些力场是费力的，需要大量的专业知识，因此受到人类在环的限制。拟议的深潜数据库（DeepPDB）的主要目标是提供一个开源工具包，能够使用先进的基于几何的计算工具自动生成力场参数的估计。 我们还将策划和传播一个经过验证的第一性原理数据集及其相应的无机材料潜力库。DeepPDB将为材料科学和机器学习社区提供服务，为前者提供解决材料挑战的关键参数，为后者提供机器学习开发的基准数据集。由此产生的协同效应将使人工智能和机器学习在计算下一代挑战的关键材料特性方面发挥更大的作用。 DeepPDB还将服务于一个关键的教育目标，允许新一代材料科学家的萌芽，他们了解如何使用深度学习来解决材料科学挑战。 DeepPDB旨在建立一个用于模拟无机材料的深度神经网络势（DNP）数据库。在此过程中，DeepPDB将：（1）开发自动化工作流程，其在给定目标组成的情况下将运行必要的密度泛函理论（DFT）计算，训练DNP，针对由训练数据施加的度量进行验证，识别具有最大不确定性的输入数据空间，并进行验证，直到训练出最佳DNP;（2）开放地传播训练DFT数据沿着预先训练的DNP;（3）开发透明的自动化验证，包括传统的基于DNP的方法以及包括目标指标的完全集成的测试。为了实现这一目标，DeepPDB将基于谨慎的软件工程实践构建一个工具包：基于功能和基于sprint的开发周期的组合;使用单元测试和集成测试作为里程碑的持续集成;以及面向数据库的数据和工作流管理方法。由此产生的开源工具包将作为一个基础工具，以调查迄今为止看不见的材料的性质，在长度和时间-该奖项由美国国家科学基金会高级网络基础设施办公室颁发，并得到美国国家科学基金会数学和物理科学理事会材料研究部的共同支持。该奖项反映了美国国家科学基金会的法定使命，并通过评估被认为值得支持使用基金会的知识价值和更广泛的影响审查标准。

A key requirement to address the world’s energy challenges is the development of new energy-efficient and smart materials. To aid in this process, state-of-the-art and revolutionary computational tools can be used to simulate new materials, thus providing a thorough understanding of their characteristics and behavior. Such materials simulation efforts can drastically reduce the time-to-market of new materials from decades to months. Traditional simulation approaches can accurately predict the behavioral properties of materials both at the smallest possible scale (atomic) and at the macroscopic level, i.e. millimeter or larger. Many critical materials properties are defined and needed at scales ranging from a few nanometers to micrometers, yet simulations are lacking at these levels. While algorithms do exist to simulate properties at these scales, often we lack the fundamental parameters, termed force-fields, for novel materials such as those for next-generation solar cells, batteries and jet turbine alloys. These force-fields are laborious to determine using traditional methods, requiring significant expertise and thus restricted by the human-in-the-loop. The primary goal of the proposed Deep Potential DataBase (DeepPDB) will be to offer an open-source toolkit with the ability to automatically generate estimates of force-fields parameters using advanced empirical-based computational tools. We will also curate and disseminate a validated repository of first-principles datasets and their corresponding potentials for inorganic materials. DeepPDB will serve both the materials science and machine learning communities, by providing the former with critical parameters to solve materials challenges and the latter by benchmark datasets for machine learning development. The resulting synergy will enable artificial intelligence and machine learning to play a greater role in computing critical materials properties for next-generation challenges. DeepPDB will also serve a critical educational objective, allowing the budding of a new generation of materials scientists, who understand how deep learning can be used to solve materials science challenges. DeepPDB aims to build a database of deep neural network potentials (DNP) for the simulation of inorganic materials. In the process DeepPDB will: (1) develop automated workflows that given a target composition, will run the necessary density functional theory (DFT) calculations, train DNPs, validate against metrics imposed by the training data, identify the input-data space with the largest uncertainty and iterate until an optimal DNP is trained; (2) openly disseminate the training DFT data along with the pre-trained DNPs; (3) develop transparent automated validation that encompasses both traditional DNP based methods as well as fully integrated tests that include target metrics. To accomplish this, DeepPDB will build a toolkit based on careful software engineering practices: a combination of feature- and sprint-based development cycles; constant continuous integration using unit-tests and integration tests as milestones; and a database-oriented approach to data and workflow management. The resulting open-source toolkit will serve as a foundational tool to investigate the properties of hitherto-unseen materials at length- and time-scales previously not possible.This award by the NSF Office of Advanced Cyberinfrastructure is jointly supported by the Division of Materials Research within the NSF Directorate of Mathematical and Physical Sciences.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.